{"id":24824,"date":"2026-08-11T04:27:25","date_gmt":"2026-08-11T04:27:25","guid":{"rendered":"https:\/\/jurn.link\/dazposer\/?p=24824"},"modified":"2026-08-11T19:51:26","modified_gmt":"2026-08-11T19:51:26","slug":"minimax-h3-on-a-3060-12gb-card","status":"publish","type":"post","link":"https:\/\/jurn.link\/dazposer\/index.php\/2026\/08\/11\/minimax-h3-on-a-3060-12gb-card\/","title":{"rendered":"Minimax H3 on a 3060 12Gb card"},"content":{"rendered":"

How times change. I’d resigned myself to never generating any worth-having generative video on a humble entry-level 3060 12Gb graphics card. Indeed, the card was starting to look ‘entry-level’ no longer. Then about a week ago came the open release of the incredible Minimax H3. Free, fully local, unrestricted, and fully supported on release day by the free ComfyUI. And it will run on my card.<\/p>\n

It was worrying to continually read that one needed 32Gb of fast DDR4 system RAM on the PC. I only have 24Gb of DDR3 RAM.<\/p>\n

But after a little tinkering I can now generate video on my humble 3060 12Gb VRAM card. 9 minutes to generate 8 seconds of 16:9 video (736 x 412px), with excellent prompt adherence. At 24 frames per second. GGUFs and a turbo LoRA mean 22Gb of files are required, all told, beyond the installation of a fresh ComfyUI Portable 0.30.2. <\/p>\n

The largest video I can get with a quick GGUF workflow is a text-to-video 0.3 (736 x 412px) video output, generated using 8 steps with the turbo LoRA and Euler \/ beta for the settings. Generation times are similar when inputting a first-frame for the video to start with, and offering this in combination with a more basic text prompt.<\/p>\n

The drawback of such a ‘smallest possible’ GGUF workflow is that the auto-matching sound (ambient + SFX + music) is not good, but I don’t much care about sound. But I’m surprised that small supercrushed GGUFs can even generate video without producing visual gloop and glitches. Just for the record here’s what the GGUF workflow runs, 22Gb in total…<\/p>\n

Model:<\/strong> H3-FLVA-Pruned-Q3_K-M.gguf (8.2Gb).
\nVAE:<\/strong> Standard safetensors (5.4Mb for both together).
\nClip:<\/strong> qwen3vl-32B-MiniMax-H3-Q2_K.gguf (7.9Gb).
\nTurbo LoRA:<\/strong> minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors with Euler \/ beta, 8 steps (600mb).<\/p>\n

After this initial success of the first basic “can it even work” test, the next steps are to: clear space on the hard-drive; work out which is the best much-larger model and clip file in the mysterious new Int8 ConvRot format (the replacement for the GGUF format); and then keep a look for better turbo LoRAs (new ones are being cooked as we speak). After that, have a look at the more sophisticated references-to-video model (also free and open). It seems the latter can do a “makeover” of video input, suggesting obvious uses for enhancing Poser and DAZ animation videos.<\/p>\n

So, I hope this initial success may encourage those with a similar old PC \/ card, and only 24Gb of older RAM. It can be done.<\/p>\n

\n

Oh, and the model doesn’t know H.P. Lovecraft, other than that (on a good day) he should look something like this…<\/p>\n

\"\"<\/a><\/p>\n

This screenshot gives you a flavour of the sort of “crunched” quality at 0.2 (lowest, 608 x 352px) and with just 6 steps on the Turbo LoRA. A bit granulated, but at 24 frames per second it’s acceptable for drafts or for pushing through Klein 4B for a comic-book panel re-style. Obviously 0.3 and 8 steps produces a larger and better video. I can’t do 0.4, as 0.3 is the limit on this quick lashed-together GGUF workflow.<\/strike> Update: Now I can, with a new VAE video decoder, new turbo LoRA and a few tweaked settings.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

How times change. I’d resigned myself to never generating any worth-having generative video on a humble entry-level 3060 12Gb graphics card. Indeed, the card was starting to look ‘entry-level’ no longer. Then about a week ago came the open release of the incredible Minimax H3. Free, fully local, unrestricted, and fully supported on release day […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13,9,4,7],"tags":[],"class_list":["post-24824","post","type-post","status-publish","format-standard","hentry","category-companion-software","category-freebies","category-spotted-in-the-news","category-the-animation-industry"],"_links":{"self":[{"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/posts\/24824","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/types\/post"}],"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=24824"}],"version-history":[{"count":13,"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/posts\/24824\/revisions"}],"predecessor-version":[{"id":24845,"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/posts\/24824\/revisions\/24845"}],"wp:attachment":[{"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/media?parent=24824"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/categories?post=24824"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/jurn.link\/dazposer\/index.php\/wp-json\/wp\/v2\/tags?post=24824"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}