{"id":24907,"date":"2026-10-03T13:10:13","date_gmt":"2026-10-03T13:10:13","guid":{"rendered":"https:\/\/jurn.link\/dazposer\/?page_id=24907"},"modified":"2026-10-03T13:34:51","modified_gmt":"2026-10-03T13:34:51","slug":"h3","status":"publish","type":"page","link":"https:\/\/jurn.link\/dazposer\/index.php\/h3\/","title":{"rendered":"H3"},"content":{"rendered":"

A Minimax H3 directory<\/strong>, aimed especially at less technical users and those on lower-end graphics cards.<\/p>\n

Updated: 3rd October 2026.<\/em><\/p>\n

This is a personal selection and does not<\/strong> aim to be completist. It is an organised version of the news round-up posts I’ve been making on Reddit. For a big technical listing of all known Minimax H3 items, including a long list of the various Turbo LoRAs which have been produced to date, see https:\/\/github.com\/wildminder\/awesome-minimax-H3<\/a>.<\/p>\n

\n
\n

LICENCE<\/p>\n

-> You can now apply for a commercial-use licence via Comfy, rather than via Minimax.<\/p>\n

https:\/\/comfy.org\/minimax\/license\/<\/a><\/p>\n

\n

MODELS<\/p>\n

-> My previous 11Gb W4A8 files held up reasonably well on a 3060 12Gb card, but for those with low VRAM I would suggest the 22Gb minimax_h3_hybrid_fl2va_ref2va_zs05_b25-49_int8.safetensors<\/em> which is a hybrid with a difference. It started with the Z-Image merge with the H3 FL2VA, then that was merged with the H3 Ref2VA to create the hybrid. Surprisingly this 22Gb file runs fine on a 3060 12Gb with the latest Comfy, with the only drawback being it takes about 70 seconds for the Sampler to actually start doing its thing. Then it works with a normal speed (Turbo + Kitchen + Spectrum, 6 to 8 steps). The point of the original Z-image merge was to generally get richer textures, not to run Z-image LoRAs, but I also find the sound to be significantly improved, and visual glitching is gone.<\/p>\n

https:\/\/huggingface.co\/hoidhxd\/MiniMax-H3-x-Z-Image-hybrid\/tree\/main<\/a><\/p>\n

\n

COMFYUI UPDATES<\/p>\n

The Portable is recommended. As always, update your ComfyUI to the latest Nightly version.<\/em><\/p>\n

Updating from an older version will fix a bug in spoken dialogue<\/d>, add controlnets 1.0 and 2.0 support; add keyframing anywhere; and common movie special-effects as small embeddings. Usage usage example for the latter: try prompting embedding:minimaxh3_art_is_explosion at 00:03.500<\/em><\/p>\n

\n

MEMORY<\/p>\n

-> Mark DK Berry has a useful YouTube video, looking in-depth at the various system RAM considerations for Minimax H3 users. Especially relevant to those on less powerful desktop Windows PCs.<\/p>\n

https:\/\/www.youtube.com\/watch?v=AfohnsAZou0<\/a><\/p>\n

\n

REGENERATING CLIPS<\/p>\n

-> MiniMax-H3-Flow-Aligned-Regenerate for ComfyUI. Generates a lower-res video at 0.7mpx and 14 steps, and it will apparently remember the model\u2019s creative decisions during that pass, and then re-use those for a final high-res pass. Theoretically, doing this should help preserve the motion, composition and overall creative direction from the low-res version. But how much actually shifts, when moving from low-res to the high-res, is not stated. It has three workflows available.<\/p>\n

https:\/\/github.com\/xmarre\/MiniMax-H3-Flow-Aligned-Regenerate<\/a><\/p>\n

\n

TURBO<\/p>\n

-> The most important workflow accelerator ‘ComfyUI Spectrum MiniMax H3’ continues to update. Update if you have an old version, and note the workflow placement and preferred samplers. It’s also important to have Spectrum do only a single-pass if using er_sde.<\/p>\n

https:\/\/github.com\/xmarre\/ComfyUI-Spectrum-MiniMax-H3<\/a><\/p>\n

-> Kijai has a fast video VAE for H3 at the official Comfy-Org HuggingFace. This newer version is a bit smaller (2.8Gb in size, instead of 3.1Gb) compared to the older version he had. Requires ComfyUI 0.36.x or higher.<\/p>\n

https:\/\/huggingface.co\/Comfy-Org\/MiniMax-H3\/tree\/main\/vae<\/a> (for the minimax_h3_video_vae_int8_convrot.safetensors<\/em> file).<\/p>\n

-> HardGravy has 6-step turbo LoRA merge + a Photorealism LoRA merged on top. Works well for me at 8-steps. Though for a natural media style it may be the wrong choice, due to the photorealism. <\/p>\n

https:\/\/huggingface.co\/HardGravy2\/Minimax_H3_FLF_HardGravy_6_Step_Turbo_Merge\/tree\/main<\/a><\/p>\n

https:\/\/github.com\/SparknightLLC\/ComfyUI-QuantizationToolkit<\/a> (ComfyUI Quantization Toolkit – nodes required by the HardGravy workflow)<\/p>\n

-> TaoMate-H3 can use only 3 steps for Ref2VA generations. It’s now ComfyUI-friendly from Kijai.<\/p>\n

https:\/\/huggingface.co\/Kijai\/MiniMax-H3_comfy\/tree\/main\/loras<\/a> (182Mb – minimax_h3_taomate_3step_lora_avg_rank_19_bf16.safetensors<\/em> )<\/p>\n

-> How low can you go? A 2-step LoRA for H3, PDMD 2-NFE, now in ComfyUI format.<\/p>\n

https:\/\/huggingface.co\/Iwannapose\/minimax_h3_pdmd_2nfe_comfyui<\/a><\/p>\n

\n

NATURAL STYLES<\/p>\n

-> An archive of Minimax H3 style LoRAs. The most interesting being an comic-book style in the AstroWitch – Cinematic Comic Style – MinimaxH3 – ASTROWITCHV01H3.safetensors file, with ASTROWITCHV01H3 as the trigger-word. The first non-Japanese comic-book style LoRA I’ve seen, and it has a pleasing sort of US\/UK 2010s ‘amateur indie comic’ look.<\/p>\n

https:\/\/huggingface.co\/EllaPriest45\/MinimaxH3_Styles\/tree\/main<\/a><\/p>\n

-> Rough Sketch Nature LoRA for MiniMax H3. A pleasing style LoRA to enhance detailed nature and sketch-line scenes with a compatible comic-book style. Requires a long introductory prompt. Works best with 20 steps and no Turbo LoRA.<\/p>\n

https:\/\/civitai.com\/models\/2956315\/rough-sketch-nature-for-minimax-h3-a-style-lora-to-enhance-nature-and-sketch-line-scenes<\/a><\/p>\n

-> MiniMax-H3-Rough-2D-Cartoon-Illustration. A 2D style LoRA trained on… “free-flow motion illustrations, rough 2D cartoon styles, and dynamic visual movements”. No demo video. Trigger phrase required: rough 2D cartoon illustration<\/p>\n

https:\/\/huggingface.co\/prithivMLmods\/MiniMax-H3-Rough-2D-Cartoon-Illustration<\/a><\/p>\n

-> Strange Reverie for H3. A cartoon\/storybook style LoRA intended for… “strange little stories, fantasy creatures, moonlit rooms […] exaggerated forms, elongated silhouettes”. Trigger: StrangeDaal and suitable prompting.<\/p>\n

https:\/\/civitai.com\/models\/2930675\/strange-reverie<\/a><\/p>\n

-> Ghosts! A new LoRA attempts… “to make ghosts [appear] out of thin air”. And, as with all ghost-hunting, it’s a bit uncertain and according to the maker… “there will be unexpected results”.<\/p>\n

https:\/\/civitai.com\/models\/2938527\/minimax-ghost<\/a><\/p>\n

\n

COMIC STYLES<\/p>\n

-> Anime Video Line Art Colorization node, which requires DiffSynth-Studio. Input a line-art video, have it beautifully auto-coloured and with temporal consistency. Convincing examples. No idea how well it will work with non-anime animation line-art.<\/p>\n

https:\/\/huggingface.co\/DiffSynth-Studio\/MiniMax-H3-LoRA-LineartAnime<\/a><\/p>\n

-> A Manga Tone Rendering LoRA for H3. Emulates the shaded line-art style common in Japanese b&w manga comics. Teaches H3… “the manga screentone rendering style with motion: monochrome ink rendering, proper tone shading, and tone that holds up under dynamic camera movement and action”. Trigger words: manga-tone rendering plus a suitable prompt. I’m guessing it might be transferable to western greyscale art styles, since I can’t discern actual printed screen dots in the shading? I’m assuming here that the LoRA doesn’t manga-fy the references, adding huge eyes and tiny noses.<\/p>\n

https:\/\/civitai.com\/models\/2930172\/manga-tone-rendering-lora-for-minimax-h3<\/a><\/p>\n

-> A gothicy style LoRA, somewhere between woodcut\/linocut and old-school b&w manga comics.<\/p>\n

https:\/\/civitai.com\/models\/2961693\/dark-life-style-minimax-h3-unusual-visual<\/a><\/p>\n

-> An ‘Anime to Real’ slider for H3, which also goes the other way “Positive values lean towards a 2D anime style”.<\/p>\n

https:\/\/huggingface.co\/adf99\/H3_Ref2V_Anime_Slider_v1<\/a><\/p>\n

-> An interesting experiment in feeding ChatGPT a comic-book page, and asking it to write it as a Minimax H3 prompt. The resulting prompt then generates an animated version of the page, as if H3 had been fed a storyboard and script.<\/p>\n

https:\/\/www.reddit.com\/r\/StableDiffusion\/comments\/1wih5rg\/i_fed_minimax_h3_3_pages_of_a_comic_and_it_spit\/<\/a><\/p>\n

\n

MOVIE STYLES<\/p>\n

-> A video aging slider LoRA for Minimax H3. Described as… “basically a little time machine for your video. The strength controls how ‘old’ the footage looks.”<\/p>\n

https:\/\/civitai.com\/models\/2941229\/mmh3-video-age-slider<\/a><\/p>\n

-> A 1920s Horror Movies LoRA. Trained on American clips rather than German, by the look of the samples, and it also appears to know wide views of the mid 1920s New York cityscape. No trigger needed.<\/p>\n

https:\/\/huggingface.co\/neph1\/1920s_horror_movies_minimax_h3<\/a><\/p>\n

-> A ‘Studio 1939’ LoRA duo, helping you to generate a… “hand-painted, golden-age animation style” from the late 1930s\/40s. Two varieties, ‘painterly’ and ‘full cel’. The LoRAs were trained on clips from public-domain material. The maker says it blends nicely with your own style prompts when set at a lower 0.4 – 0.8 strength. A trigger word is required: gulliv3r – which you may want to add to the filenames.<\/p>\n

https:\/\/huggingface.co\/lovis93\/studio-1939-old-animation-lora-minimax-h3<\/a><\/p>\n

-> A new Film Noir LoRA for Minimax H3. No trigger word, but it was trained to also add small highlight colours in blood, red, and orange + glowing jewellery (seemingly glowing in an eerie ‘Innsmouth green’) and thus the maker suggests prompting for e.g. “Selective red glowing color on lips”. You can also do a normal b&w noir style, so long as no colours are specified in the prompt.<\/p>\n

https:\/\/civitai.com\/models\/2934763\/film-noir-with-selective-colors<\/a><\/p>\n

-> Yee-haw! A 1950s westerns LoRA for H3 FL2VA. As in the wild west and cowboys. No trigger, but the maker suggests adding… “The lighting is natural, with [time-of-day] light. The [photograph] has a vintage, cinematic quality.” He also suggests breaking from 16:9 and using a super-widescreen screen format which these movies often used, such as 1.85:1 or 2.35:1 (Cinemascope). See also a gunfight LoRA.<\/p>\n

https:\/\/huggingface.co\/neph1\/1950s_western_movies_h3<\/a><\/p>\n

https:\/\/huggingface.co\/JOKER141\/MiniMaxH3-GunFu-Gunfight-LORA<\/a><\/p>\n

-> A 1950s Scifi Minimax H3 LoRA. The trigger is 50s_scifi and if you want the Technicolor look in the samples you have to prompt for it.<\/p>\n

https:\/\/huggingface.co\/neph1\/1950sScifiMinimaxH3<\/a><\/p>\n

-> Have Minimax seamlessly and ‘magically’ transform the style of a room from modern to late 1950s, in a few seconds. With an amusing video demo, workflow and prompt.<\/p>\n

https:\/\/old.reddit.com\/r\/StableDiffusion\/comments\/1w0ei8t\/time_period_shift_special_effect_in_minimax_h3\/<\/a><\/p>\n

-> A new ‘Retro Toon 70s’ LoRA for Minimax H3. Helping you to emulate a… “retro 1970s feature-film cel animation \u2014 hand-painted characters with bold ink lines over textured, painterly matte backgrounds, limited-animation motion, and the grainy, earthy palette of the era.”<\/p>\n

https:\/\/huggingface.co\/KennethFal\/retro-toon-70s-lora-minimax-h3<\/a><\/p>\n

-> A ‘1980s VHS videotape style’ LoRA for Minimax H3. Has samples.<\/p>\n

https:\/\/huggingface.co\/KennethFal\/vh5tape-vhs-lora-minimax-h3<\/a><\/p>\n

-> Rewind to the 1980s! MiniMax-H3-Tape-FX for ComfyUI gives Minimax H3 videos the look of… “VHS, BetaMax and LaserDisc \u2014 including the worn-out tape look, tracking errors, dropout, creases, ghosting, head-switch noise, vertical roll and a period-correct VCR on-screen display.”<\/p>\n

https:\/\/huggingface.co\/Smite79\/MiniMax-H3-Tape-FX<\/a><\/p>\n

-> A new 1980s Horror Movies LoRA, which also appears to offer useful ‘night in the graveyard’ lighting. Suggested trigger is 80s_horror If you need some emaciated ghouls for your graveyard, I see there’s also a new Weight Slider LoRA.<\/p>\n

https:\/\/huggingface.co\/neph1\/1980s_horror_movies_minimax_h3<\/a><\/p>\n

https:\/\/huggingface.co\/Playtime-AI\/MM-H3-Weight_Slider<\/a><\/p>\n

-> An 1980s_scifi_movies LoRA for H3, trained on the cyberpunk end of the genre. No trigger word required. <\/p>\n

https:\/\/huggingface.co\/neph1\/1980s_scifi_movies_minimax_h3<\/a><\/p>\n

-> A 3D-Animation-Style-MiniMax-H3 style LoRA, for the popular modern Pixar \/ Dreamworks animation style. Note that Minimax can also more or less emulate this style on its own – see the Style Atlas (link below).<\/p>\n

https:\/\/huggingface.co\/SOLRICKS\/3D-Animation-Style-MiniMax-H3<\/a><\/p>\n

https:\/\/github.com\/hoodtronik\/minimax-h3-style-atlas<\/a><\/p>\n

-> Do you know your tradwives from your truckvibes? Then you may be interested in a new H3 LoRA which claims to style your clips to appeal to the Instagram crowd.<\/p>\n

https:\/\/huggingface.co\/vpakarinen\/insta-tiktok-aesthetics-h3-lora<\/a><\/p>\n

\n

STYLE TRANSFER<\/p>\n

-> Experimental LoRAs for style transfer (“restyles an entire video while attempting to preserve its motion, framing and timing”), and also for adding VFX (“return the same take with only the requested change applied”).<\/p>\n

https:\/\/huggingface.co\/Alissonerdx\/Minimax-H3-ComfyUI<\/a><\/p>\n

https:\/\/huggingface.co\/Alissonerdx\/Minimax-H3-ComfyUI\/blob\/main\/docs\/style-transfer.md<\/a> (Style readme)<\/p>\n

https:\/\/huggingface.co\/Alissonerdx\/Minimax-H3-ComfyUI\/blob\/main\/docs\/vfx-edit.md<\/a> (VFX readme)<\/p>\n

\n

LIGHTING AND RE-LIGHTING<\/p>\n

-> A useful simple day-night slider LoRA for Minimax H3.<\/p>\n

https:\/\/comfy.icu\/models\/2949040\/-MMH3-DayNight-Slider-v1-0<\/a><\/p>\n

-> New H3_slider_experiments. Just the one at present, a simple slider to either add sun (2.0) to the scene, or gloom and depth-fog (-3.0). Has a dinosaur demo video.<\/p>\n

https:\/\/huggingface.co\/siraxe\/H3_slider_experiments<\/a><\/p>\n

-> H3-Relight for ComfyUI. Add up to three scene lights in a visual-preview widget, and the widget creates a reference and matching prompt. The interface is in Portuguese, but many of the label words are similar to those in English. Update: Now has an English version.<\/p>\n

https:\/\/github.com\/NyckM\/H3-Relight-Minimax-ComfyUI\/tree\/main<\/a><\/p>\n

https:\/\/github-com.translate.goog\/NyckM\/H3-Relight-Minimax-ComfyUI\/tree\/main?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a> (English translation)<\/p>\n

-> ‘Skynet Dark – MiniMax H3 Style LoRA’. A style LoRA to enhance a… “dark, rain-soaked cinematic look … warm tungsten practicals, cold rain-window rim light, deep shadows, wet reflective streets”. Trigger: skynetstyle placed first in the prompt. Works best with… “a pencil-sketch storyboard plus character references”, which gives the LoRA some room to enhance the scene.<\/p>\n

https:\/\/civitai.com\/models\/2918509\/skynet-dark-minimax-h3-style-lora<\/a><\/p>\n

-> Siraxe has sliders to either add sun (2.0) to the scene, or gloom and depth-fog (-3.0). <\/p>\n

https:\/\/huggingface.co\/siraxe\/H3_slider_experiments<\/a><\/p>\n

-> Something’s Off. A general LoRA for a ‘make the scene darker and more eerie’ Halloween style.<\/p>\n

https:\/\/civitai.com\/models\/2955672\/somethings-off-h3-minimax-style-lora-by-fos<\/a><\/p>\n

\n

CAMERA MOTION<\/p>\n

-> CrossView-Warp LoRA for H3, with matching nodes. “Give it a video and a camera offset – azimuth, elevation – and it generates the same scene from that new viewpoint”. The process… “reads two videos: a depth-warp of your clip [made by the Comfy node], which carries the geometry, and the [reference] clip itself, which carries identity and appearance.” Has a workflow. Trigger word needed: crosswarp<\/p>\n

https:\/\/huggingface.co\/Cseti\/MiniMax-H3_Ref2VA-LoRA-CrossView-Warp_v1<\/a> (LoRA)<\/p>\n

https:\/\/github.com\/cseti007\/ComfyUI-CrossViewWarp<\/a> (ComfyUI nodes)<\/p>\n

https:\/\/huggingface.co\/datasets\/Cseti\/ComfyUI-Workflows\/blob\/main\/minimax-h3\/crossview-warp\/README.md<\/a> (ComfyUI workflow)<\/p>\n

-> MiniMax-H3-360-Orbit-LoRA for the H3 FL2VA model. Give the model the same image as the first and the last keyframe, and the LoRA orbits the camera all the way around the subject – while the scene\/character stays frozen. The resulting video clip is a loop. Requires a special prompt.<\/p>\n

https:\/\/huggingface.co\/pablodawson\/MiniMax-H3-360-Orbit-LoRA<\/a><\/p>\n

-> A new Camera Motion helper LoRA, including ‘aerial drone shot’ and ‘macro \/ extreme close-up’ (small diorama). Just a 1000 step version, with better versions yet to come.<\/p>\n

https:\/\/huggingface.co\/Jojocodex\/minimax-h3-Camera-Motion-lora<\/a><\/p>\n

https:\/\/huggingface-co.translate.goog\/Jojocodex\/minimax-h3-Camera-Motion-lora?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a> (English translation)<\/p>\n

-> Siraxe has sliders for Camera Arc, Detail (add), Speed, and Smear. No trigger words required.<\/p>\n

https:\/\/huggingface.co\/siraxe\/H3_slider_experiments<\/a><\/p>\n

-> A Camera Movement LoRA for Minimax. 12 camera moves added including ‘Orbit’, but the maker says it works best with ‘Handheld’ and ‘Slow pull out’ \/ ‘Slow push out’. Works in tandem with turbo LoRAs. Several drawbacks: the ComfyUI version seems to require careful choosing; the LoRA gives an ignorable error when loading; and it requires yunjing as the trigger word.<\/p>\n

https:\/\/huggingface.co\/Jojocodex\/minimax-h3-yunjing-lora<\/a><\/p>\n

https:\/\/huggingface-co.translate.goog\/Jojocodex\/minimax-h3-yunjing-lora?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a> (translation)<\/p>\n

-> Fellow British bloke ‘Nerdy Rodent’ has a useful YouTube video. His latest Minimax-focused tutorial covers, among others: continuing an existing clip using the Ref2VA model, and placing two video-reference characters in a new setting with their voices intact; having the camera ‘dive’ into an image and imagine going around a corner or into a building; and he also tests some new acceleration boosters. Free workflows and prompts, though only on the screen.<\/p>\n

https:\/\/www.youtube.com\/watch?v=rZnZcwobcR8<\/a><\/p>\n

-> Also on Youtube, a useful survey from LoopForge of what works in translating the twelve official Minimax camera motions into fourteen industry-recognised shot types (e.g. crash zoom, rack focus, whip pan, crane rise etc). Also has a matching page with the full prompts and workflows.<\/p>\n

https:\/\/www.youtube.com\/watch?v=GCdKh_KaQVs<\/a><\/p>\n

https:\/\/loopforge.cc\/minimaxh3-shots-skills\/<\/a><\/p>\n

-> A Claude Code ‘Skill’ to expand short-code camera prompts into full blown Minimax H3 subject-aware camera prompts… e.g. you start with “\/MS \/LOWANGLE \/DOLLYIN:MS>MCU:SLOW” and then it’s expanded.<\/p>\n

https:\/\/github.com\/SK-SL-source\/cinematic-director-camera-dsl<\/a><\/p>\n

\n

CAMERA PATHING<\/p>\n

-> Traffic Motion Studio is a user-controlled traffic sim built for Minimax H3 output. “Directly draw vehicle trajectories, arrange events such as cutting in, collisions, forced lane changes, and overtaking, export accurate 2D simulation reference videos”. At a guess, it might possibly be useful for blocking and placing cameras in your dramatic car-chase scenes? Or could Minimax be prompted to use the output for reference?<\/p>\n

https:\/\/github.com\/caiusy\/traffic-motion-studio<\/a><\/p>\n

https:\/\/github-com.translate.goog\/caiusy\/traffic-motion-studio?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a> (English translation)<\/p>\n

-> The new ComfyUI-Camera-Path node. “Give it an image (or a batch of frames) plus the matching geometry from ComfyUI’s built-in Run MoGe Inference, draw a camera move in the node’s 3D editor, and it renders the scene from those virtual camera positions. The result is a plain IMAGE batch, so it drops into any workflow that takes a control or reference video.” Has a video demo. This advanced node has been enabled by ComfyUI v0.37.0 adding MoGe 3 geometry estimation, the result of Microsoft Research’s decades-long interest in 3D point-clouds. Running the MoGe nodes in ComfyUI requires the file moge_3_vitg_fp16.safetensors (2.5Gb) or moge_3_vitl_fp16.safetensors (741Mb) in your ..\/models\/geometry_estimation folder.<\/p>\n

https:\/\/github.com\/phazei\/ComfyUI-Camera-Path<\/a><\/p>\n

https:\/\/docs.comfy.org\/built-in-nodes\/MoGeInference<\/a><\/p>\n

https:\/\/huggingface.co\/Comfy-Org\/MoGe\/tree\/main\/geometry_estimation<\/a><\/p>\n

\n

CAMERAS FOR VIDEOGAME EMULATION<\/p>\n

-> WarmBloodAban has launched his H3 LoRA concept series with his ‘Minimax-h3_Third_person_view’ LoRA. This helps with emulating the… “game CGs, HUD UIs, and dynamic 3rd\/1st-person camera controls” commonly found in videogames.<\/p>\n

https:\/\/huggingface.co\/WarmBloodAban\/Minimax_H3_LoRAs<\/a><\/p>\n

-> A camera LoRA for a first-person videogame style ‘head-mounted camera’.<\/p>\n

https:\/\/huggingface.co\/neph1\/minimax_h3_headmounted_fpv_cam<\/a><\/p>\n

-> From the maker of the ‘Studio 1939′ LoRA, a new ’16Bit Pixel \u2014 SNES-era pixel’ LoRA, helping H3 to animate… “like a SNES-era cutscene in motion”. No trigger needed. The same maker has a new hand-drawn LoRA for Minimax H3, but there’s no video demo for that one. My fixed-seed tests ‘with’ (hand-drawn in prompt + LoRA at 0.8) and ‘without’ (just hand-drawn in prompt) suggest no hand-drawn effect either way. But you may discover something different.<\/p>\n

https:\/\/huggingface.co\/KennethFal\/16bit-pixel-lora-minimax-h3<\/a><\/p>\n

https:\/\/huggingface.co\/KennethFal\/hand-drawn-lora-minimax-h3<\/a><\/p>\n

\n

CINEMATOGRAPHY<\/p>\n

-> A Minimax LoRA designed to add the “authentic cinematic texture” as found in traditional film cinematography. Reduce the strength to 0.5 for fast action scenes and for use in Ref2VA workflows. Trigger: DY and perhaps add a prompt addition such as: “High-end commercial product cinematography, 1960s filmic aesthetics, [specific] atmosphere, [specific] lighting, a theatrical interplay of light with the prevailing atmospheric conditions of the scene”.<\/p>\n

https:\/\/huggingface.co\/kirk86413\/cinema-h3\/tree\/main<\/a><\/p>\n

-> ComfyUI-Cinematic-Prompt. “A visual, user-friendly prompt builder for ComfyUI that allows you to construct complex cinematic prompts using a visual interface with previews”. Includes presets for professional movie-camera types, and cinematic film stocks.<\/p>\n

https:\/\/github.com\/yedp123\/ComfyUI-Cinematic-Prompt<\/a><\/p>\n

-> ComfyUI-EasyColorCorrector. Presets for… “professional film stock emulation with highlight roll-off”, and a suite of sophisticated colour correction nodes.<\/p>\n

https:\/\/github.com\/regiellis\/ComfyUI-EasyColorCorrector<\/a><\/p>\n

-> A Cinematic Video Prompt \u2014 H3 Adapter. This is for LLM clients that support Agent Plugins v1.0.0, the new Microsoft and Google-backed standard for universal agent plugins. This H3 plugin… “adapts a 700+ term cinematography vocabulary for MiniMax H3 video generation. The vocabulary is camera-craft language (angles, movement, lighting, composition, lens\/format, style, color, mood, sound design, editing, FACS, storytelling).” The output is in the H3 prompt style.<\/p>\n

https:\/\/github.com\/ssm82\/cinematic-video-prompt-h3-plugin<\/a><\/p>\n

https:\/\/agent-plugins.org\/compatible-clients<\/a> (official list of LLM clients which support Agent Plugins v1)<\/p>\n

\n

MOTION PHYSICS<\/p>\n

-> A Spatial & Physics LoRA for Minimax H3. Intended to help with physics-based prompts that include scene actions such as… “the blocks are slowly stacked on top of each other, then the stack collapses onto the floor”. No trigger word needed.<\/p>\n

https:\/\/huggingface.co\/Jojocodex\/minimax-h3-spatial-physics-lora<\/a><\/p>\n

https:\/\/huggingface-co.translate.goog\/Jojocodex\/minimax-h3-spatial-physics-lora?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a> (translation)<\/p>\n

-> MiniMax H3 Semantic Bridge, for prompt-to-video FL2VA generations only. A RTX 3090 Ti card and the SenseNova model were first used to ‘teach’ Minimax H3 to more correctly understand some core relationships, such as those present in spatial layouts, lighting, and exact human hand movement. For instance, which… “surfaces are transparent or reflective [and] how light passes through one material but reflects from another”. The result has been encapsulated inside this small 11Mb ComfyUI conditioning adapter node. The node’s example workflow shows a glass object displaying convincing see-through\/reflections from its environment. The node also helps H3 understand mirrors, or which of two hands should hold an object. I thus guess it might be useful for your film noir gun-battle scene, set in a ‘Hall of Mirrors’ carnival sideshow – but don’t blame me if your PC explodes while trying to generate that!<\/p>\n

https:\/\/huggingface.co\/speach1sdef178\/MiniMax-H3-Semantic-Bridge<\/a><\/p>\n

https:\/\/www.reddit.com\/r\/StableDiffusion\/comments\/1w7wtco\/i_tried_transferring_semantic_representations\/<\/a><\/p>\n

-> ‘MiniMax-H3 motion adapter (pilot version)’. Another potential helper for fast-motion video scenes. This one… “teaches the base model to spend the extra clock [time] on smoothness”.<\/p>\n

https:\/\/huggingface.co\/MATLOWAI\/MiniMax-H3-Motion-Adapter<\/a><\/p>\n

\n

MOTIONS FOR THE BODY<\/p>\n

-> BUNNY General Motion Continuity Repair LoRA. This is said to be the best H3 motion-fixer LoRA, by those have tested several such helpers. “A general-purpose motion support LoRA, not a combat-only LoRA. It can be used for running \/ sports \/ dance \/ acrobatics \/ character interaction \/ combat \/ weapon motion and other dynamic scenes. … designed around repairing those [‘awkward movement’] moments rather than simply making every motion stronger.” The trigger is bunny_crisp_motion but it will work without a trigger in the prompt. No workflow.<\/p>\n

https:\/\/huggingface.co\/JOKER141\/MiniMax-H3-General-Motion-Continuity-Repair<\/a><\/p>\n

-> MiniMax-H3-Combat-Base-V2 LoRA. “Combat Base V2 is no longer just a combat LoRA \u2014 it now works for both action and dialogue scenes, adding richer body movement, finer sound details, and stronger shot-to-shot continuity to character performance.” No trigger needed, unless you have really high intensity (e.g. Marvel superheroes battle) where you use prfight2 or “heavy stunt impact” (e.g. Indiana Jones style) where prfin1 is the trigger. Has two demo workflows, for FL2VA and Ref2VA.<\/p>\n

https:\/\/huggingface.co\/JOKER141\/MiniMax-H3-Combat-Base-V2<\/a><\/p>\n

-> ‘Fight Prompt Director’ is a Codex \/ Claude Code ‘skill’ which crafts prompts for fight and heavy physical-action scenes in Minimax H3. “It turns text, images, keyframes, or storyboard references into executable video prompts with readable action, spatial continuity, camera intent, and character-specific movement.”<\/p>\n

https:\/\/github.com\/irenerachel\/fight-prompt-director<\/a><\/p>\n

-> MiniMaxH3-GunFu-Gunfight-LORA, a LoRA for a range of gunfighting styles.<\/p>\n

https:\/\/huggingface.co\/JOKER141\/MiniMaxH3-GunFu-Gunfight-LORA<\/a><\/p>\n

-> ‘Better Human Motion’ for Minimax H3 is another general motion improver LoRA.<\/p>\n

https:\/\/huggingface.co\/vpakarinen\/better-human-motion-h3-lora\/tree\/main<\/a><\/p>\n

-> A MiniMax H3 Single-Dancer Prompt Skill. Appears to be a ‘skill’ for Codex. Input a reference image and it then… “displays six options: dance style, duration, style, camera movement, audio, and ending”. Make your choices and out pops a complete H3-ready prompt, ready to copy-paste. Dance styles include Modern Dance, Ballet, Latin, among others.<\/p>\n

https:\/\/github.com\/lj350201364\/MiniMaxH3_Single_Dancer_Prompt_Skill<\/a><\/p>\n

https:\/\/github-com.translate.goog\/lj350201364\/MiniMaxH3_Single_Dancer_Prompt_Skill?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a> (English translation)<\/p>\n

\n

REALISM AND SKIN <\/p>\n

-> MiniMax-H3-Realism-People-LoRA for… “faces that hold up in close-up, natural skin texture, believable expressions and gestures, film-style lighting and documentary camera movement.” Convincing examples.<\/p>\n

https:\/\/huggingface.co\/fal\/MiniMax-H3-Realism-People-LoRA<\/a><\/p>\n

-> A MiniMax-H3-Facial-Realism-CloseUp LoRA, aiming to enhance subtle details and micro-expressions. Has samples.<\/p>\n

https:\/\/huggingface.co\/prithivMLmods\/MiniMax-H3-Facial-Realism-CloseUp<\/a><\/p>\n

-> Vpakarinen’s Natural-face-speech-h3-lora is now newly at v2. This offers… “Face muscle dynamics with clear English speech” and V2 apparently adds… “male\/female voice with flawless speech sync and crystal clear quality”. Also note that this LoRA, and his Better-human-motion-h3-lora, have new demo videos.<\/p>\n

https:\/\/huggingface.co\/vpakarinen\/better-human-motion-h3-lora\/tree\/main<\/a><\/p>\n

https:\/\/huggingface.co\/vpakarinen\/natural-face-speech-h3-lora\/tree\/main<\/a><\/p>\n

-> A Minimax H3 LoRA that aims to give naturally blemished and somewhat pimply skin, judging by the convincing video demos. It even tries to improve hands and eyes. Requires trigger words: perfe8ct; perfect hands; perfect skin or perfect eyes. A bit of a counterintuitive trigger for skin, though, since it seems the LoRA make the skin more realistically grunky, not more perfect.<\/p>\n

https:\/\/civitai.com\/models\/2899220\/minimax-h3-t2va-t2v-polyhedron-perfect-eyes-perfect-skin-perfect-hands<\/a><\/p>\n

-> A note on the above. Specifying realism for the face and eyes, paired with suitably emotive dialogue and tone, might help Minimax to simply infer expressions (thus, they won’t need to be specified in the prompt). For example…<\/p>\n

“… looking toward the camera with a relaxed and genuine expression. We see a detailed natural skin texture with subtle skin tone variations, cheeks slightly rosy to indicate health, and overall a realistic subsurface scattering of light on the face, ears and hair. Small natural catchlights occur on the moist surface of the eyes, as the person’s head and eyes move, and these catchlights reflect the surrounding environment. As the person speaks the camera continually catches the organic micro-expressions made by a human face, and a naturally variable eye blinking pattern.”<\/p>\n

\n

FACEFIX AND FACE ADJUSTMENT<\/p>\n

-> A FaceSwap LoRA for MiniMax H3 Ref2VA (or the hybrid model, presumably). It “changes the Reference video identity face with the Reference” image or images. The trigger keyword is faceswap which should be relatively easy to remember. Has a range of demo videos and a workflow. Note that the workflow requires CRT-Nodes (a nodes pack which I see has also recently updated to support the H3 Fun Controlnet). Works very well, and easily.<\/p>\n

https:\/\/huggingface.co\/UntMods\/FaceSwap_MiniMaxH3_REF2VA<\/a><\/p>\n

https:\/\/github.com\/PGCRT\/CRT-Nodes<\/a><\/p>\n

-> On YouTube, a tested fix for improving ‘faces at distance’. By applying a low denoise video-to-video pass on a 832 x 480px video file reference + one reference image (to ensure character consistency). A ComfyUI workflow for this is freely available, and it’s designed to work on an RTX 3060 12Gb card. The drawback is the video-to-video output is likely to remove any lip-sync from the source video.<\/p>\n

https:\/\/www.youtube.com\/watch?v=d1h5-E7NpuY<\/a> (20 minute tutorial, including good advice on shutting down other GPU-hogging software.)<\/p>\n

<\/a> (for the workflow MBEDIT – MH3_r2v_SingleSampler_Detailer_v17.json )<\/p>\n

-> Face Fix for poorly rendered ‘distant faces’. Requires face_yolov8m.pt for face-detection \/ regeneration.<\/p>\n

https:\/\/github.com\/Carasibana\/ComfyUI-H3-FaceRefine<\/a><\/p>\n

-> The BSAI-ComfyUI-FaceRefine for ComfyUI, another facefixer for small or distant faces.<\/p>\n

https:\/\/github.com\/xm6018924\/BSAI-ComfyUI-FaceRefine<\/a><\/p>\n

-> A simple age-slider LoRA for Minimax H3, with demo video showing advanced old-age to toddler.<\/p>\n

https:\/\/huggingface.co\/Playtime-AI\/Minimax_H3-Age_Slider\/tree\/main<\/a><\/p>\n

-> ComfyUI_MiniMaxH3_Director now also has a Face Refine node for ComfyUI, which also supports Semantic Bridge semantic enhancement.<\/p>\n

https:\/\/github.com\/AIMixer\/ComfyUI_MiniMaxH3_Director<\/a><\/p>\n

https:\/\/github.com\/AIMixer\/ComfyUI_MiniMaxH3_Director\/tree\/main\/director\/face_refine<\/a><\/p>\n

https:\/\/github.com\/AIMixer\/ComfyUI_MiniMaxH3_Director\/blob\/main\/README_EN.md<\/a> (English readme)<\/p>\n

-> ComfyUI-H3-FaceAutoBypass is now newly at v2. Does your generated clip really need a facefix, or not? This node decides for you, potentially saving you time.<\/p>\n

https:\/\/github.com\/fukkun2705-commits\/ComfyUI-H3-FaceAutoBypass<\/a><\/p>\n

-> NKD’s new Face Rig is a slick modern ComfyUI update for the old Live Portrait. “Poses a portrait’s expression by dragging handles that sit on the face itself: brows, eyelids, gaze, mouth corners, jaw, head. The result re-renders live while you drag.” Which would seem to have obvious uses for Minimax, such as quickly adjusting an existing portrait image to use as a starting frame for a Minimax image-to-video clip. No workflow, but judging by the video it looks like a relatively simple node setup.<\/p>\n

https:\/\/github.com\/Nekodificador\/ComfyUI-NKD-Basic-Tools\/blob\/master\/docs\/face-rig.md<\/a><\/p>\n

-> Minimax-H3-ref2v_Anime_2_Realism, which aims to assist with your attempts at anime-to-photoreal conversion.<\/p>\n

https:\/\/huggingface.co\/LiseTY\/Minimax-H3-ref2v_Anime_2_Realism<\/a><\/p>\n

-> 3d_to_real_detail_slider_H3, a simple slider LoRA for use with Minimax H3 text-to-video. Shifts ‘toony into realistic’, and has convincing demo examples.<\/p>\n

https:\/\/huggingface.co\/siraxe\/3d_to_real_detail_slider_H3<\/a><\/p>\n

\n

CHARACTER SHEETS<\/p>\n

-> ComfyUI-OrbitSheets for Minimax H3. Claims to build flawless ‘turnaround’ character and location reference-sheets.<\/p>\n

https:\/\/github.com\/lumos675\/ComfyUI-OrbitSheets<\/a><\/p>\n

-> H3 Character Sheet Generator. “Throw in some rough reference images, get back a character sheet you can reuse forever. […] Six frames from one video generation can’t [make bjorked character views] because they come out of the same pass. That’s the whole trick. The camera does a slow orbit with no hard cuts, the character stands still like a statue, and then this workflow grabs six frames and stitches them together.” A good readme, don’t skip the second half. It also does objects, and there are prompts\/demos for anime-to-real style conversions.<\/p>\n

https:\/\/huggingface.co\/SwagMessiah100\/H3_Character_Sheet_Generator<\/a><\/p>\n

-> The local open-source Inline Studio now lets Minimax H3 train (slowly) a character LoRA as a .char file. The input dataset is just two good images. Has workflows and a specific H3 guide, but note that H3 training requires a 24Gb VRAM graphics card and lots of system RAM. That said, note that a .char is uniquely portable between models – and training a .char on Flux.2 Klein 4B Base appears to need only 10Gb VRAM. So… train on Klein 4B, run on H3 FL2VA? See also: Refmods.<\/p>\n

https:\/\/github.com\/inlineresearch\/Inline-Studio<\/a><\/p>\n

\n

CROWDS AND SPACE<\/p>\n

-> Does your story require an ensemble team of superheroes, or perhaps you just need a crowd scene for your space-station arrivals lounge? The ‘Minimax H3 15+ reference image workflow’ patches ComfyUI to remove the nine reference limit for Ref2VA generation. Use caution: it passed the mods on CivitAI, but this uses a .BAT file for the patch, and it calls a small .PY Python file. The code in there looks safe to me – I can read Python scripts by eye – and there’s also an undo .PY script. Note also that there may be other nodes that claim to offer more than nine, and without patching.<\/p>\n

https:\/\/civitai.com\/models\/2929051\/minimax-h3-15-reference-image-workflow<\/a><\/p>\n

-> H3-Visual-Marker-Control. Experiments which bring Klein’s ‘draw a rectangular red box’ trick to H3, but enhanced. The boxes you draw on the reference image don’t simply specify a target area to focus the change on, as with Klein. Here you can add a blue frame around a character, and blue frame around a target area. The character then moves to that framed area with natural motion. With a two character scene, you use blue for the main fast-moving character and then red for the calmer character. Yellow can also be added to indicate a “distant action target” that the main moving character should naturally move back to. It’s experimental at present, but it obviously works and there are convincing demos and workflows.<\/p>\n

https:\/\/github.com\/RK-BoilingPoint\/H3-Visual-Marker-Control<\/a><\/p>\n

https:\/\/github.com\/RK-BoilingPoint\/H3-Visual-Marker-Control#1-introduction<\/a> (English readme)<\/p>\n

https:\/\/github.com\/RK-BoilingPoint\/H3-Visual-Marker-Control\/tree\/main\/Videos<\/a> (video demos)<\/p>\n

\n

SPRITE PRODUCTION<\/p>\n

-> ‘Blubs-pixel-nodepack’ for ComfyUI. Nodes for… “turning MiniMax H3 output into pixel-art sprite animations”, as commonly used in retro videogames. There are also workflows, and a bridge to the popular $20 Aseprite software.<\/p>\n

https:\/\/github.com\/japaneserunic\/blubs-pixel-nodepack<\/a><\/p>\n

-> A workflow designed for Minimax as a seamlessly looping animated .GIF generator. Requires ComfyUI-LoopGif nodes, which appears to do sophisticated glitch correction at the loop seams.<\/p>\n

https:\/\/civitai.com\/models\/2889463\/using-minimax-h3-as-an-ai-gif-generator<\/a><\/p>\n

https:\/\/github.com\/HM1579\/ComfyUI-LoopGif<\/a><\/p>\n

-> Sprite Sheep. Load one character drawing, prompt Minimax H3 and get a sprite sheet for use in arcade-style games, plus a preview animated .GIF file.<\/p>\n

https:\/\/github.com\/MANFRI-DEV\/sprite-sheep<\/a><\/p>\n

-> The new MiniMax H3 Sprite Motion Tool, which has canned motions to apply to your arcade-style game sprites (“Idle, Walk, Run, Attack, Cast, Hit, Death, Jump, and custom motions”).<\/p>\n

https:\/\/github.com\/SkylightStudio07\/minimax-h3-sprite-tool<\/a><\/p>\n

-> MyPet is new. An… “AI Desktop Pet Project: Using two fully automated pipelines in which a character design is transformed into a set of playable desktop pet action sprites, without the need for manual image-editing.” Uses Z-Image-Turbo and MiniMax H3.<\/p>\n

https:\/\/github.com\/zhongpei\/mypet<\/a><\/p>\n

https:\/\/github-com.translate.goog\/zhongpei\/mypet?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a> (English translation)<\/p>\n

\n

GAME WORLDS & 360 VR<\/p>\n

-> ‘H3-World’ and a matching LoRA for it. Generate walkable video-worlds, controlled by videogame-like WASD keyboard controls. “Given an initial frame and keyboard controls, it generates action-controlled video with coordinated character and camera motion.” Possibly we’re just recreating Myst here, but maybe there’s a whole new approach developing.<\/p>\n

https:\/\/github.com\/Danzer1xxxxChan\/H3-World<\/a><\/p>\n

https:\/\/huggingface.co\/DANNY621\/H3-World<\/a><\/p>\n

-> An Equirectangular 360-degree LoRA for MiniMax H3. With this, Minimax outputs a 360\u00b0 environment that wraps around the viewer. The maker says you simply change the output’s native video ratio to 2:1, then… “tag it with spherical metadata (spatialmedia -i –v2 –stereo=none -p equirectangular), and it plays as an immersive mono-360 clip in a Quest \/ DeoVR \/ Skybox.”<\/p>\n

https:\/\/huggingface.co\/shamanic\/minimax-h3-equi360-lora<\/a><\/p>\n

\n

OUTPAINTING<\/p>\n

-> H3 OutpaintPrep. This ComfyUI node… “prepares a source video for a latent-mask outpaint with MiniMax H3. It pads the frames to the target canvas, and hands back the video\/audio noise masks” in a form that H3 understands.<\/p>\n

https:\/\/github.com\/lovemachine100\/ComfyUI-H3-OutpaintPrep<\/a><\/p>\n

\n

BATCH<\/p>\n

-> Batch Minimax for ComfyUI. Appears to run one reference image against one video reference, along with one prompt file (e.g. video_1.mp4 + refimage_1.png + prompt_1.txt), and it then runs the next set. Has workflows.<\/p>\n

https:\/\/github.com\/TagirovAlex\/BatchMiniMax<\/a><\/p>\n

-> The Tenitsky Prompt Cycler node for ComfyUI.<\/p>\n

https:\/\/github.com\/tenitsky\/tenitsky-prompt-cycler-simple<\/a><\/p>\n

\n

SEED-HUNTING<\/p>\n

-> New on YouTube, a “Minimax H3 Seed Hunter v2.0 Workflow Release + Usage Guide”. With a link to the workflow.<\/p>\n

https:\/\/www.youtube.com\/watch?v=cJYL6W2GtIM<\/a><\/p>\n

<\/a> (ComfyUI workflow bundle is free to all, even non-Patreon members, on his Patreon)<\/p>\n

A 2-step LoRA for H3, PDMD 2-NFE, now in ComfyUI format.<\/p>\n

https:\/\/huggingface.co\/Iwannapose\/minimax_h3_pdmd_2nfe_comfyui<\/a><\/p>\n

\n

STILL IMAGES & TYPOGRAPHY<\/p>\n

-> The maker of Fizgig has released ComfyUI-Fizgig-H3-Still, as nodes plus workflows. He says that with this… “a true single frame renders far faster than the old 5-frame workaround”, even when ramping up the megapixels. Tested by me at 8Mpx \/ 30-steps + Kitchen Attention to see how it handles smaller typography. Even at 8mpx it’s still no good for really miniscule lettering, but you can at least now add film star names and a production company credit in smaller lettering below the main title (see demo link below). Note that it does remove any colour cast (e.g. it gives a pure stark b&w film-noir, removing the very slight sepia warming tone present in the normal output). But as compensation it’s pleasingly smooth with none of blocky pixel seams \/ ‘scan lines’ to be seen on the regular output. Recommended.<\/p>\n

https:\/\/github.com\/shootthesound\/ComfyUI-Fizgig-H3-Still<\/a><\/p>\n

https:\/\/jurn.link\/dazposer\/wp-content\/uploads\/2026\/09\/Fizgig_H3_Still_00011_-copy.jpg<\/a> (typography demo, 8Mpx, 30 steps)<\/p>\n

-> minimax_h3_single_frame_decoder_500k.safetensors (9.69Gb). This new and big VAE looks like a weighty attempt at a single-frame generator. Trained from the original VAE… “on 500,000 unique image-reconstruction examples”, with the apparent aim of generating a no-glitches single-frame output suitable for product concept-shots.<\/p>\n

https:\/\/huggingface.co\/iamkaikai\/MiniMax-H3-Single-Frame-VAE-500K<\/a><\/p>\n

-> ComfyUI-MiniMaxH3-SingleFrame. Two still-image generation nodes, with the second being especially interesting. Given the usual first\/last frames, it attempts to interpolate\/generate a plausible single ‘middle frame’. Has workflows, and requires no ComfyUI Core patching or special VAE.<\/p>\n

https:\/\/github.com\/tori29umai0123\/ComfyUI-MiniMaxH3-SingleFrame#english<\/a> (English ReadMe section)<\/p>\n

-> The ComfyUI_Minimax_H3_Frame_Selector. New to me, though it appears to have been around for a while. Generate the video, manually select one frame, and then save just that frame as an image. Possibly useful if you’re going to generate the video anyway, and want a somewhere-in-the-middle frame extracted as well as first\/last frames.<\/p>\n

https:\/\/github.com\/TripleHeadedMonkey\/ComfyUI_Minimax_H3_Frame_Selector<\/a><\/p>\n

\n

WORKFLOW EXTRAS<\/p>\n

-> How to connect a video clip as a reference-file input in H3’s Ref2VA workflow. Easy, when you know how.<\/p>\n

https:\/\/www.reddit.com\/r\/comfyui\/comments\/1vpiwob\/wheres_the_place_to_attach_video_in_minimax_h3\/<\/a><\/p>\n

-> ComfyUI-MiniMaxH3Mod custom nodes for ComfyUI. Claims faster Ref2VA generation, by turning your references into little .safetensors files. Said to be especially useful for those using a larger video clip as a reference-file?<\/p>\n

https:\/\/github.com\/Luisacaotica\/ComfyUI-MiniMaxH3Mod<\/a><\/p>\n

-> Deno Custom Nodes for ComfyUI added a A\/B ‘Video Compare’ node, a few weeks ago. Works with a simple slider, in the way that image-compare nodes do.<\/p>\n

https:\/\/github.com\/Deno2026\/comfyui-deno-custom-nodes<\/a><\/p>\n

-> ComfyUI-H3-PowerLoraStack. “A stacked multi-LoRA loader for MiniMax H3, in the spirit of rgthree’s Power Lora Loader, but built around the three things that actually break H3 LoRAs.”<\/p>\n

https:\/\/github.com\/cicalooo\/ComfyUI-H3-PowerLoraStack<\/a><\/p>\n

-> ComfyUI-FirstframeLastframeExtractor will help in more or less reproducing a draft H3 FL2VA video at a larger size after locking the seed\/prompt. Works as advertised, very simple. Plug it into the VAE decode node, it auto saves the two frames you need. Then you optionally load these images back to your first-frame\/last-frame inputs, via the native “Load Image (from outputs)” nodes.<\/p>\n

https:\/\/github.com\/RmaNMetaverse\/ComfyUI-FirstframeLastframeExtractor<\/a><\/p>\n

-> The new H3-Shadow-Negative-Nodes for ComfyUI. These aim to add negative prompting to your workflows, if you feel you need it. Note that Minimax can already tolerate some negative prompting within the main prompt.<\/p>\n

https:\/\/huggingface.co\/Leon1000\/H3-Shadow-Negative-Nodes<\/a><\/p>\n

-> ComfyUI-YCNodes-MiniMax-H3. A H3 Sigma Refiner node for ComfyUI, that claims to… “solve the problem of pixel particles and flickering at the edges of high-speed moving objects in H3 videos. Local ‘micro-sculpting’ steps are added to the noise scheduling in the low Sigma (low noise) range”. The new ‘H3 Sigma Refiner’ node is placed between the BasicScheduler and SamplerCustomAdvanced nodes. I use it in all my workflows.<\/p>\n

https:\/\/github.com\/yichengup\/ComfyUI-YCNodes-MiniMax-H3<\/a><\/p>\n

https:\/\/github-com.translate.goog\/yichengup\/ComfyUI-YCNodes-MiniMax-H3?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a><\/p>\n

-> ComfyUI-MiniMax-H3-Keyframe-Offset. Custom node that lets you… “place first\/last keyframes at arbitrary frame[s], instead of only at the start\/end of the clip”. Has added an interesting audio node that can have audio partly conform to a ref_image. e.g. prompt for “echo reverberating through the valley”, add a reference image of stony mountains, and it generates a suitable ‘wide distant echo’ in a big stony valley.<\/p>\n

https:\/\/github.com\/asirusasr-maker\/ComfyUI-MiniMax-H3-Keyframe-Offset<\/a><\/p>\n

-> The new H3 GuideMaster for ComfyUI… “makes it easy to place images and audio on a MiniMax H3 timeline”. Pleasingly designed. Requires the latest ComfyUI, for the ability to add guide images (aka keyframes) at any point in the video.<\/p>\n

https:\/\/github.com\/MajoorWaldi\/ComfyUI-Majoor-H3-GuideMaster<\/a><\/p>\n

-> A workflow for the H3-LongVideos seamless clip chainer. The maker had not supplied one. Though the node has since been reworked so much that it’s effectively a new node – he says that any old workflows are defunct and will need node-changing. Thus this bundle usefully includes the old git clone node folder. Pop it into your Custom Nodes folder, to match the demo workflow, and don’t auto-update the node.<\/p>\n

https:\/\/jurn.link\/dazposer\/wp-content\/uploads\/2026\/08\/H3-Minimax-LongVideos-OLD_VERSIONinc-demo-workflow-for-ComfyUI.zip<\/a><\/p>\n

-> On YouTube, Mark DK Berry has a new short ‘quickstart’ video on using the TAE Preview (taeh3.safetensors) in ComfyUI. This preview node gives your workflow clear… “high-quality real-time latent previews when working with Minimax H3”.<\/p>\n

https:\/\/www.youtube.com\/watch?v=pMRdQvmJbTQ<\/a><\/p>\n

https:\/\/huggingface.co\/Kijai\/MiniMax-H3-TAE\/tree\/main\/vae_approx<\/a><\/p>\n

-> Genkai-ComfyUI-Nodes for ComfyUI – PromptSync node. “Watch a H3 generated video alongside its original timed prompt. Check whether actions, camera cuts, and dialogue happen at the intended moments.”<\/p>\n

https:\/\/github.com\/GENKAIx\/Genkai-ComfyUI-Nodes<\/a><\/p>\n

\n

AUDIO<\/p>\n

-> ‘ComfyUI-H3-AudioRefine’ for users of 4-step turbo LoRAs. An experimental node set that can freeze the video stream while the workflow… “runs additional denoising steps on the audio stream only”. The aim is to improve audio quality, when using 4-step turbo LoRAs. There are a downsides and trade-offs here, so it’s very important to read the readme. Without its video freezing node, my tests on a RTX 3060 12Gb card have it working well. At 6 steps (on its own sampler), it can add maybe 35-45 seconds to a turbo 6-step 0.3 six-second clip generation.<\/p>\n

https:\/\/github.com\/Adudeguyman\/ComfyUI-H3-AudioRefine<\/a><\/p>\n

-> A high-quality ‘ASMR audio and soft whisper acoustics’ LoRA for Minimax H3. With two demo videos.<\/p>\n

https:\/\/huggingface.co\/vpakarinen\/asmr-trigger-audio-h3-lora<\/a><\/p>\n

-> ComfyUI-MiniMaxH3DualClockSampler. A sampler node that runs audio and video generation on different ‘clocks’ during generation, potentially solving the problem of… “severe popping, clipping, and noisy audio issues during the 4-step [turbo LoRA-enhanced] generation process”.<\/p>\n

https:\/\/github.com\/shuaixn\/<\/a><\/p>\n

-> ComfyUI-H3-ExactAudioLock. This… “provides a human Audio Review \/ Accept Gate for choosing one approved take, before it continues through the workflow. [It is designed for H3 productions] where the picture needs to respond to exact user-supplied audio, instead of regenerated or approximate audio”.<\/p>\n

https:\/\/github.com\/badgids\/ComfyUI-H3-ExactAudioLock<\/a><\/p>\n

-> A workflow to… “set custom soundtracks in Minimax without R2VA, by using latent noise masks” in FL2VA. Apparently also works with a lip-sync track. See comments for workflow and enhancements.<\/p>\n

https:\/\/www.reddit.com\/r\/StableDiffusion\/comments\/1vtv0qs\/psa_in_h3_you_can_set_custom_soundtracks_without\/<\/a><\/p>\n

-> A user has discovered that Minimax has a basic understanding of the International Phonetic Alphabet (IPA, aka ‘phonemes’), which are used for precise word transcription and pronunciation. I guess it would make sense for Minimax to know phonemes, re: generating lip-sync. Here he used a large online AI to adjust the IPA translation towards a national accent. My thought is that it may be more trustworthy if you first convert in Balabolka, and then have it read back to you using TTS in Balabolka. Then you can at least be sure the IPA translation is correct in English, before passing it over to the AI for accent adjustment.<\/p>\n

https:\/\/www.reddit.com\/r\/StableDiffusion\/comments\/1w0tb8q\/minimax_h3_accents_mmh3_understands_the_ipa\/<\/a><\/p>\n

https:\/\/www.cross-plus-a.com\/balabolka.htm<\/a> (Balabolka, the only freeware offline English-to-IPA translator for Windows. Use: Select text | Edit | Pronunciation | Read This | select drop-down: ‘Phonemes (IPA)’ | click button ‘Phonemes’ | then click ‘Convert text to IPA’).<\/p>\n

-> ComfyUI-H3-ExactAudioLock is for those who need.. “multi-speaker scenes, overlapping speech, singing”, and who thus want the video… “to respond to exact user-supplied audio instead of regenerated or approximate audio.”<\/p>\n

https:\/\/github.com\/badgids\/ComfyUI-H3-ExactAudioLock<\/a><\/p>\n

\n

CHARACTER REPLACEMENT<\/p>\n

-> A concise prompt for character replacement, with a special command for when your video has a hand or other body-part passing in front of a face.<\/p>\n

https:\/\/old.reddit.com\/r\/MiniMax_AI\/comments\/1w5wdjs\/how_to_made_minimax_h3_video_edits_way_easier_to\/<\/a><\/p>\n

-> The new ComfyUI-Viggle-Animate-H3 fine-tune model is designed for full character replacement. “ComfyUI nodes for Viggle-Animate, a 33.1 B full finetune [21Gb] of MiniMax-H3’s ref2va transformer for character replacement in video. It takes a driving video and a reference still, and re-renders the performer(s) in the clip as the character in the still. Motion, camera, timing, background and lighting come from the video; identity comes from the image.” 22Gb appears to be the smallest H3 Viggle combination (Viggle model + matching DMD2 LoRA).<\/p>\n

https:\/\/github.com\/Saganaki22\/ComfyUI-Viggle-Animate-H3<\/a><\/p>\n

-> MiniMax-H3-Character-Swap-LoRA, a new LoRA which attempts to make character-swop a little better… “In our local comparisons it often preserved the background more closely than the base model, but motion timing, facial expressions, and hard cuts remain unreliable.”<\/p>\n

https:\/\/huggingface.co\/akatz-ai\/MiniMax-H3-Character-Swap-LoRA<\/a><\/p>\n

-> The final 1000-step version of the MiniMax-H3-Character-Swap-LoRA is now available, along with a REF2VA workflow. Assists with swopping source and picture, while keeping all other elements the same. The rendering medium for the character sheet is also transferred (e.g. comic-book style). The drawback is that even id you have close-up facial expressions they are often not followed – though that may not matter if you have a monster, space alien, robot\/cyborg, or masked superhero. No trigger word is required. Note than REF2VA can also natively do character swopping, with relevant prompting.<\/p>\n

https:\/\/huggingface.co\/akatz-ai\/MiniMax-H3-Character-Swap-LoRA\/tree\/main<\/a><\/p>\n

\n

CONTROLNETS<\/p>\n

-> The original MiniMax-H3-Fun-Controlnet-Union file. This Controlnet accepts… “Canny, Depth, HED, MLSD or Pose control [Openpose] videos, and also runs video inpainting.” 6.8Gb in size. Has video examples.<\/p>\n

https:\/\/huggingface.co\/alibaba-pai\/MiniMax-H3-Fun-Controlnet-Union<\/a><\/p>\n

-> The MiniMax-H3-Fun-Controlnet-Union-2.0. Version 2.0 is now available and this can accept additional inputs – Scribble, Layout, and Gray. The number of control blocks have also been doubled. Kijai has uploaded a 4.5Gb INT8 version of v2.0, working with the latest ComfyUI. <\/p>\n

https:\/\/huggingface.co\/alibaba-pai\/MiniMax-H3-Fun-Controlnet-Union-2.0<\/a><\/p>\n

https:\/\/huggingface.co\/Kijai\/MiniMax-H3-experimental\/tree\/main\/model_patches<\/a><\/p>\n

-> A workflow for using the Minimax H3 Fun Controlnet, when using a suitable input video in ComfyUI. In fact, it’s a three-in-one workflow (one live, two bypassed). <\/p>\n

https:\/\/civitai.com\/models\/2920872\/minimax-h3-fun-control-in-comfyui-or-depth-and-pose-video<\/a><\/p>\n

-> ComfyUI-ReShot. This updated a few days ago, and now supports Openpose and Canny. ReShot transforms a regular video input into a Depth Map, Openpose skeleton, or Canny edge lines video. Which can then be used by the Minimax H3 Fun Controlnet.<\/p>\n

https:\/\/github.com\/maosika-ai\/ComfyUI-ReShot<\/a><\/p>\n

-> ComfyUI-H3-FunControl. Another set of nodes that aim to run the new Controlnet in ComfyUI, specifically Kijai’s Comfy-friendly conversion.<\/p>\n

https:\/\/github.com\/wyzborrero\/ComfyUI-H3-FunControl<\/a><\/p>\n

https:\/\/huggingface.co\/Kijai\/MiniMax-H3-experimental\/tree\/main\/controlnet<\/a><\/p>\n

\n

LORA TRAINING<\/p>\n

-> The Fizgig LoRA trainer can train Minimax H3 LoRAs on AMD Radeon graphics-cards. If the card has 16Gb VRAM or higher. Also Refmods. “Minimax H3 now trains on video clips, on their sound, and on voice recordings alone: photos, clips and voice files in one folder to train one LoRA in one run”. Fizgig can do so locally in 16Gb VRAM, without slowdowns. Adds a fix so the LoRAs will work correctly on both the 4-step Turbo or the official workflow, and has also benefitted from a major security audit. Fellow-Brit and industry professional Shoot The Sound has a good 20-minute tutorial today on YouTube, using the latest Fizgig and its dataset prep tool, to train a character+voice LoRA.<\/p>\n

https:\/\/www.reddit.com\/r\/StableDiffusion\/comments\/1vvqrgp\/fizgig_now_trains_loras_on_amd_radeon_flux_2\/<\/a><\/p>\n

https:\/\/github.com\/shootthesound\/Fizgig<\/a><\/p>\n

-> ComfyUI-Fizgig-H3-Tweaks. A optional official bolt-on node for the LoRA trainer and dataset prep tool Fizgig. Includes… “a prompt-strength dial for H3’s CFG-free Turbo renders”.<\/p>\n

https:\/\/github.com\/shootthesound\/ComfyUI-Fizgig-H3-Tweaks<\/a><\/p>\n

-> For LoRA trainers, there’s also DiffSynth Studio’s new ‘MiniMax-H3 DeCFG Training Adapter’ LoRA. They say that… “the base MiniMax-H3 model is CFG-distilled, which can make direct LoRA fine-tuning unstable or degrade the distilled CFG-free behavior. This adapter temporarily pulls the distilled DiT back toward its pre-distillation behavior during training, providing a better optimization landscape for new LoRAs.”<\/p>\n

https:\/\/modelscope.ai\/models\/DiffSynth-Studio\/MiniMax-H3-TrainingAdapter<\/a><\/p>\n

-> Ostris’s H3 Training Adapter is now at version 3.<\/p>\n

https:\/\/huggingface.co\/ostris\/minimax_h3_training_adapter\/tree\/main<\/a><\/p>\n

\n

REFMODS<\/p>\n

-> A YouTube tutorial for beginners, steadily taking you through the creation of a RefMod for Minimax H3, in ComfyUI.<\/p>\n

https:\/\/www.youtube.com\/watch?v=2K6-OtV_Vbc<\/a><\/p>\n

-> The Fizgig LoRA trainer and dataset wrangler. New in 6.x is the ability to make RefMods, as well as LoRAs. Presumably this change now makes Fizgig a tool for those with 12Gb VRAM.<\/p>\n

https:\/\/github.com\/shootthesound\/Fizgig<\/a><\/p>\n

-> A RefMods fork which… “aims to improve the UI” for picking your RefMods.<\/p>\n

https:\/\/github.com\/FranckyB\/ComfyUI-H3RefModPicker<\/a><\/p>\n

\n

VIDEO TO PROMPT<\/p>\n

-> A ComfyUI-VideoAnalyzer. “ComfyUI nodes that turn a video into a MiniMax H3 video-generation prompt […] by analyzing it with a vision-language model.” The user can choose from several Qwen VL models, to align with their available VRAM.<\/p>\n

https:\/\/github.com\/aroslanov\/ComfyUI-VideoAnalyzer<\/a><\/p>\n

-> ‘Video to H3 Prompt’ as an “end-to-end pipeline that turns a video file into a ready-to-paste MiniMax H3 generation prompt”. Appears to be a ‘skill’ for use with a local installation of OpenAI’s Codex, which is a lightweight coding agent.<\/p>\n

https:\/\/github.com\/LoveRain1997\/video-to-h3-prompt<\/a><\/p>\n

https:\/\/github.com\/openai\/codex<\/a><\/p>\n

-> A local video-to-h3-prompt tool. Get a… “complete, ready-to-paste MiniMax H3 prompt [by feeding a video clip to] a five-channel evidence pipeline that reconstructs the causal event chain [in the video], separates live action \/ camera \/ edit-effects \/ sound layers, and compiles the exact H3 fields for every input mode.” Command line, no GUI as yet. Requires ffmpeg \/ ffprobe on PATH in Windows.<\/p>\n

https:\/\/github.com\/LoveRain1997\/video-to-h3-prompt<\/a><\/p>\n

\n

PROMPTS & LLM PROMPT HELPERS<\/p>\n

-> I went looking for straightforward lightweight H3-tuned .GGUF, for local use as a prompt idea rewriter, to be hosted in Jan.ai (Windows desktop freeware, a llama.cpp GGUF runner). This is the one I picked.<\/p>\n

https:\/\/huggingface.co\/geocine\/minimax-video-prompt-enhancer-2.6b-gguf<\/a><\/p>\n

https:\/\/github.com\/janhq\/jan<\/a><\/p>\n

-> A combined package to… “use local Qwen3.8-27B and official MiniMax-H3 Skills in ComfyUI to generate H3 prompts.” This spins up its own standalone local llama.cpp server, and then serves you a Minimax-skilled Qwen3.8-27B model inside a ComfyUI node. Qwen is removed from your VRAM after use. Nice if you can afford the VRAM.<\/p>\n

https:\/\/github.com\/chflame163\/ComfyUI_Qwen_H3_Prompt<\/a><\/p>\n

https:\/\/github-com.translate.goog\/chflame163\/ComfyUI_Qwen_H3_Prompt?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp<\/a> (English translation)<\/p>\n

-> A working prompt to maintain a straight-up character view from head-to-feet, throughout a continuous pull-out shot.<\/p>\n

https:\/\/old.reddit.com\/r\/StableDiffusion\/comments\/1vt2fea\/minimax_h3_prompting_so_that_it_will_keep_the\/<\/a><\/p>\n

-> A new H3 Prompt Journal. Some scenes require complex physical logic from the camera. The Journal’s first three entries demonstrate how to write prompts for such scenes: for a ‘Three-Person Occlusion-Linked Orbital Long Take’ (e.g. elegantly redirect the camera between three moving people, in a single take); ‘Dual-subject-speed-contrast’ (e.g. a dancing master leads in a waltz, while his hesitant student follows his moves); and ‘Single-subject-three-pose’ (e.g. input three poses for one character, then have a gnat-sized camera… “sweep past feet, legs, torso, shoulders, hair – constantly redirecting around the moving [giant] body without ever slowing down”).<\/p>\n

https:\/\/github.com\/LoveRain1997\/h3-prompt-journal<\/a><\/p>\n

-> The standalone offline H3 Prompt Composer continues to develop and refine, and is now at version 5.43.4. With new support for “multi-subject camera prompting” and “better control for more complex camera moves”.<\/p>\n

https:\/\/github.com\/BMB12d3\/minimax-h3-prompt-composer<\/a><\/p>\n

https:\/\/www.youtube.com\/watch?v=Aywx3Sf5Yk0<\/a> (27 minute YouTube tutorial).<\/p>\n

-> ComfyUI-Fantastic-MiniMaxH3-PromptBuilder custom nodes. Helps you build prompts locally inside ComfyUI, while adhering to the built-in official prompt guide… “H3 doesn’t want a casual sentence \u2014 it wants a structured prompt with named sections, shot timings, speaker IDs, and tags pointing at your reference media.” No LLM required. Convoluted workflows, but for simplest use: just plug it into your existing prompt node, then click on the blue box to open the Builder. Then build the prompt, and pass it back to the prompt box.<\/p>\n

https:\/\/github.com\/Adudeguyman\/ComfyUI-Fantastic-MiniMaxH3-PromptBuilder<\/a><\/p>\n

https:\/\/huggingface.co\/MiniMaxAI\/MiniMax-H3\/blob\/main\/docs\/VIDEO_PROMPT_WRITING_GUIDE_base_en.md<\/a><\/p>\n

-> A prompt to neatly mix styles in one video. e.g. SpongeBob SquarePants as a cartoon, appearing in a live-action sitcom.<\/p>\n

https:\/\/www.reddit.com\/r\/StableDiffusion\/comments\/1vp8dpc\/mix_style_inside_the_same_scene_minimax_h3\/<\/a><\/p>\n

-> MiniMax-H3-Prompt-Rewriter-LoRA-GGUF… “A GGUF conversion of lightx2v\/MiniMax-H3-Prompt-Rewriter-LoRA, so the rewriter runs under llama.cpp” and thus in Jan.ai etc, or inside ComfyUI with llama-cpp-python installed to its Python.<\/p>\n

https:\/\/huggingface.co\/ivanfromm\/MiniMax-H3-Prompt-Rewriter-LoRA-GGUF\/tree\/main<\/a><\/p>\n

-> A ComfyUI workflow to… “turn one scene photograph into eight target-centered cinematic camera views, in one MiniMax H3 generation.” Doing it in one generation gives some stability to the scene geometry.<\/p>\n

https:\/\/huggingface.co\/ethanfel\/H3_Cinematic_Multishot_Coverage<\/a><\/p>\n

\n

SCRIPTWRITERS<\/p>\n

-> From Germany, a new Single-Line Video Prompt Engine & Screenplay Generator, aka ‘Drehbuch–Referenzanker-Generator’. This is a… “Professional Dramaturgical Screenplay Studio [which] turns raw bullet points, concept ideas, and reference images into dramaturgically structured video screenplays broken down into contiguous timecoded windows (e.g. 14.000 seconds per window) […] Optimized for MiniMax H3”. Bilingual, in English or German.<\/p>\n

https:\/\/github.com\/wobushannes\/Drehbuch–Referenzanker-Generator<\/a><\/p>\n

\n

SUBTITLES<\/p>\n

-> Multi-language Subtitle generator. The speech is transcribed from the video by local models and translated by a local LLM, and written to local .SRT subtitles files.<\/p>\n

https:\/\/github.com\/dodgypast\/mlsubgen<\/a><\/p>\n

\n

DISCOVERIES & TRICKS<\/p>\n

-> Input an image of a painting, generate a time-lapse style video showing how that painting could have been made. It’s a nice creative idea. Real artists might splutter and say “it’s not done like that”. But, as the poster suggests, using a painter’s technical terms might improve the results.<\/p>\n

https:\/\/old.reddit.com\/r\/StableDiffusion\/comments\/1vzhwj3\/using_minimax_h3_to_reverseengineer_a_paintings\/<\/a><\/p>\n

-> A user shows that Minimax can emulate a ‘time-lapse’ video showing a drawing\/painting being created really fast. Includes the prompt he used.<\/p>\n

https:\/\/www.reddit.com\/r\/StableDiffusion\/comments\/1w0yt06\/h3_speedpainting_v2_prompt_included\/<\/a><\/p>\n

-> A video of Minimax H3 ingesting a six-image storyboard, reading each image in sequence from top-bottom and left-right, and generating the shot sequence as a single video. There’s a workflow, and a full structured prompt. One interesting finding shared in the video is that combining a detailed pencil-sketch storyboard with a photoreal\/cinematic prompt gives Minimax more room for creative interpretation.<\/p>\n

https:\/\/www.youtube.com\/watch?v=LByGCGzu67o<\/a><\/p>\n

https:\/\/github.com\/amao2001\/ganloss-latent-space\/blob\/main\/workflow\/2026-08-26%20minimax_h3_r2v_story_board.json<\/a><\/p>\n

-> A Minimax H3 Storyboard ‘skill’ for Claude. Also note some of the maker’s interesting findings, such as… “H3 drops facial [expression] instructions silently when a shot holds too many of them” for the available processing to handle, and “ref_audio_0 drives the mouth, not the voice” for lip-sync.<\/p>\n

https:\/\/github.com\/phileiny\/h3-storyboard-skill<\/a><\/p>\n

-> How do you keep your scene environment stable, when generating longer clips? This useful Reddit thread discusses various methods and tweaks.<\/p>\n

https:\/\/old.reddit.com\/r\/StableDiffusion\/comments\/1w1z2fu\/environment_consistency_minimax_h3\/<\/a><\/p>\n

-> Reddit advice on the quality and type of the reference images required to get adequate character face consistency in Minimax. Another new thread has some thoughts on style transfer.<\/p>\n

https:\/\/old.reddit.com\/r\/StableDiffusion\/comments\/1w30rsj\/has_anyone_tried_fine_tuning_minimax_h3_for_a\/p6wtd9t\/<\/a><\/p>\n

https:\/\/old.reddit.com\/r\/comfyui\/comments\/1w2sgxf\/currently_what_is_the_best_style_transfer_method\/<\/a><\/p>\n

-> On YouTube, from the maker of Fizgig, a quick… “demonstration that with the right training, training a LoRA with images alone on Minimax H3 does not have to damage movement and composition.”<\/p>\n

https:\/\/www.youtube.com\/watch?v=zhFwI7VkT6s<\/a><\/p>\n

-> ComfyUI Remote MiniMax H3. With this you can apparently… “run MiniMax H3’s Clip \/ Text Encoder processing on another PC to reduce VRAM usage on your main ComfyUI machine.”<\/p>\n

https:\/\/github.com\/aaruyou\/ComfyUI-RemoteMiniMaxH3<\/a><\/p>\n

\n

GUIDES<\/p>\n

-> ComfyUI has a page linking to the official prompting guides (there are two, for the different FL2VA and REF2VA models).<\/p>\n

https:\/\/docs.comfy.org\/tutorials\/video\/minimax\/minimax-h3-prompt-guide<\/p>\n

-> The official Minimax H3 repository has a Skills folder. Eight of these Skills are style\/narrative specific. e.g. “Explain science, education, or general knowledge through tactile handmade papercraft visuals.” The Skills should work with anything that… “can read a SKILL.md local file”.<\/p>\n

https:\/\/github.com\/MiniMax-AI\/MiniMax-H3\/tree\/main\/skills<\/a><\/p>\n

https:\/\/www.jan.ai\/docs\/agent\/skills (readme for Skills in the local desktop llama.cpp GGUF-runner Jan.ai)<\/p>\n

-> The official Awesome MiniMax H3 Integrations page, with links.<\/p>\n

https:\/\/github.com\/MiniMax-AI\/awesome-minimax-h3-integration<\/a><\/p>\n

-> A Minimax H3 Browsable Offline Style Atlas (1.2Gb packed as a .ZIP file). Being… “a browsable index of all 941 distinct visual styles across the 1,000 video clips”. “Styles are grouped into eight media categories (live-action cinematic, film stock & era looks, documentary & broadcast, amateur\/found footage, 2D animation, stop-motion & puppetry, 3D\/CG & game renders, and specialty imaging), with a live text filter for browsing.” Search results are shown initially as quick-loading stills, with each still hyperlinked to its local video clip.<\/p>\n

https:\/\/github.com\/hoodtronik\/minimax-h3-style-atlas<\/a><\/p>\n

-> A fresh Minimax ‘known characters and celebrities’ list, version 2, with matching video clips. My own tests show it has only very a vague idea of H.P. Lovecraft and J.R.R. Tolkien, when run without references.<\/p>\n

https:\/\/huggingface.co\/datasets\/malcolmrey\/various\/blob\/main\/h3-center\/known-characters\/INDEX.md<\/a><\/p>\n

-> A collection of 685 source-attributed MiniMax H3 video examples, with the public prompts that made them. Many just point to Twitter \/ X. Not sure how many of these clips would also get you a ComfyUI workflow, but apparently the video metadata has not been stripped unless they’re on Twitter. So… drop them in ComfyUI and see.<\/p>\n

https:\/\/github.com\/SkyNotSilent\/awesome-minimax-h3<\/a><\/p>\n

-> A big curated and very technical listing of all known Minimax H3 items. Includes a long list of the various Turbo LoRAs which have been produced to date.<\/p>\n

https:\/\/github.com\/wildminder\/awesome-minimax-H3<\/a><\/p>\n

-> The producer of an acclaimed AI short film kindly… “published his prompt collection documents and walked through his entire method on a public livestream.” His release has now been distilled into “a structured, reusable system”, packaged as a Claude Code Skill. There’s also a human-readable one-page quickstart ‘cheatsheet’.<\/p>\n

https:\/\/github.com\/jnMetaCode\/ai-shortfilm-prompts<\/a><\/p>\n

https:\/\/github.com\/jnMetaCode\/ai-shortfilm-prompts\/blob\/main\/cheatsheet.md<\/a><\/p>\n

\n

HARDWARE<\/p>\n

-> A replacement stack to run Minimax H3 faster on a Intel Arc Pro B70 graphics card. Made with C++, Rust, Typescript.<\/p>\n

https:\/\/github.com\/maxfridbe\/sycl_minimax_h3<\/a><\/p>\n

\n

SOFTWARE UTILITIES<\/p>\n

-> Tagscanner. Windows freeware that lets you show your Minimax videos on your blog etc, without also inadvertently sharing your ComfyUI workflow (and any previous prompts stuck in it, e.g. yesterday I shared a H3 workflow for prompt “A western-style fantasy dragon …”, but on inspection in Notepad++ I inadvertently found the .JSON also has the workflow’s starting-point prompt “H.P. Lovecraft walking down a dark alley…” hidden in it. Who knew old prompts could stick around like that?). Tagscanner is Windows freeware. Usage: Drag-drop the .MP4 video file into Tagscanner, right-click it, ‘Cut’, then ‘Save’. That’s it. The software is meant for .MP3 music taggers, but works fine for quick metadata removal.<\/p>\n

https:\/\/www.majorgeeks.com\/files\/details\/tagscanner.html<\/a><\/p>\n

-> Do your Minimax and RTX upscaled .MP4 videos lack thumbnails in Windows? The old-school freeware Icaros… “can provide Windows Explorer thumbnails, for essentially any video media format supported by FFmpeg”. Install as Admin, and I find it works instantly without a PC reboot. Last updated June 2026, supports Windows 11.<\/p>\n

https:\/\/github.com\/Xanashi\/Icaros<\/a><\/p>\n

-> The excellent freeware ‘Screen to GIF’ easily converts your .MP4 videos to re-sized old-school animated .GIFs (or better, animated .PNG files). A modern GUI for the Windows desktop PC, very easy to use, just drag in your .MP4 files. Updated last month. No sound, of course, but the format may be useful to some.<\/p>\n

https:\/\/github.com\/nickemanarin\/screentogif<\/a><\/p>\n

-> The excellent open-source video editor OpenShot has released version 4.0. Among the new features are improved export presets, plus… “creative presets for motion, camera moves, colour, film looks, lighting, and audio”, and… “local AI model downloads and workflows for Object Mask and Object Detection”. Vital tip for first-time users: To zoom into the timeline, hold down the Crtl key and then roll the mouse-wheel forwards. OpenShot also officially provide some experimental ComfyUI integration nodes and workflows. I can offer a useful tip for the free OpenShot video editor. Right-click your imported clip in the Project Files, and “Add profile”. Name the profile “Minimax 0.5” or somesuch. Now when you go to export the edited video, there’s a new export profile preset called “Minimax 0.5” – this lets you export your edited video exactly the same way it went in – pixel-size, frame-rate, and so on. Then for later quality file-size shrinkage, for posting on a blog etc, the dedicated freeware Handbrake is perhaps the better option.<\/p>\n

https:\/\/github.com\/OpenShot\/openshot-qt\/releases\/tag\/v4.0.0<\/a> (desktop versions and changelog)<\/p>\n

https:\/\/github.com\/OpenShot\/OpenShot-ComfyUI<\/a><\/p>\n

https:\/\/handbrake.fr\/features.php<\/a><\/p>\n

-> The open-source sound editor Audacity 4.0.0 has just been released. It’s a huge makeover. The UI no longer looks like it’s from the late 1990s, and it’s now a lot easier to use. Should be of use to those who want maximum video visuals, and who prefer to craft their audio track later.<\/p>\n

https:\/\/github.com\/audacity\/audacity\/releases<\/a><\/p>\n

https:\/\/www.youtube.com\/watch?v=BTQymidLYIM<\/a> (five minute demo)<\/p>\n

-> For character inputs with motion, note that a full Poser 12 Pro licence is just $20 via Graphixly. Poser 12 is mature ‘pose 3D human figures and render’ software from 2022, and its animation tools are fully bundled. Potentially useful for quickly producing clay-shaded 3D figure\/expression animations, for use as H3 reference inputs (Poser’s 3D figures can take any .BVH motion-capture file). Poser also has a real-time comic-book inking mode built in, giving you quick and pleasing line-art from 3D figures. Clay-shaded or line-art renders will prevent reference character-traits from leaking into your Minimax characters, and also give H3 more room to be a little creative. May also be useful for making references for closeups of hands and feet.<\/p>\n

https:\/\/graphixly.com\/products\/poser-pro<\/a><\/p>\n

https:\/\/www.youtube.com\/watch?v=dkN8dN-FtGI<\/a> (short tutorial on Poser’s real-time comic-book inking)<\/p>\n

\n

That’s it!<\/p>\n","protected":false},"excerpt":{"rendered":"

A Minimax H3 directory, aimed especially at less technical users and those on lower-end graphics cards. Updated: 3rd October 2026. This is a personal selection and does not aim to be completist. It is an organised version of the news round-up posts I’ve been making on Reddit. For a big technical listing of all known […]<\/p>\n","protected":false},"author":1,"featured_media":0,"parent":0,"menu_order":11,"comment_status":"closed","ping_status":"closed","template":"","meta":{"footnotes":""},"class_list":["post-24907","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/pages\/24907","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/comments?post=24907"}],"version-history":[{"count":8,"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/pages\/24907\/revisions"}],"predecessor-version":[{"id":24920,"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/pages\/24907\/revisions\/24920"}],"wp:attachment":[{"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/media?parent=24907"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}