Could Flux Kontext Dev handle a backdrop as well as a character, thus bypassing the need to composite later? To find out I threw together a basic garden around Nursoda’s Ronk figure and his snail. Obviously, one would spend a lot more time constructing a garden that was destined to appear in many scenes in a storybook or comic. But this is just for a workflow demo.

Pretty ugly from Poser (Comic Book mode lineart and a bright light preset helps it along, but like all 3D it’s desperate to go ‘dark and grungy’). Yet Kontext handles it nicely. Note the new word at the start of the prompt, ‘Filter …’

The problem is then the garish day-glo nature of the colouring on the new image. But because we have 1:1 registration with the Poser source-image, we can easily lay the colours back in by using it as a colour blending layer in Photoshop. Here that’s been done. Then just a little of the Kontext colour has been brought back in. The layer was then flattened and auto-contrast applied, then desaturated slightly to take account of the colour-boost caused by the auto-contrast. The final result…

And since it’s come from Poser, we can have easy-select masks galore via a clown pass / toonID render, should any further postwork be needed. And if a holding-line around the character, or a blurring or fading of the background, is needed… then Poser can also supply the masks needed.

A quick Poser experiment with the new Flux Kontext Dev. Nursoda’s Ronk and his snail, in Poser. Render to real-time Preview at 2048px, with high texture quality and a little Comic-Book applied. Lay this Poser render on white in Photoshop, reduce to 1024px and use this as the seed image.

The prompt gives a pencil and watercolour effect, but does not cause the layer-registration to shift. It remains an exact 1:1 match, despite the style change. In other words, Kontext can act exactly like a Photoshop filter would. Takes about 70 seconds on a 3060 12Gb graphics-card, at 1024px. This speed is comparable with intensive Photoshop filter plugins such as Reactor or G’Mic. There is a ‘turbo’ version from a third party, said to give a 2x speed up, but it appears to require intense Python wrangling and lots of tracking down dependencies to get it to work.

A 1:1 match means we can restore the Poser colour, by using the original render as a colour-blending layer in Photoshop. Which means we can have consistent colour from panel to panel and page to page, when storytelling in a comic or storybook.

We get a little drop-out of definition. For instance, the spiral of the snail’s shell is lost. If we had a lineart only Firefly render from Poser, we could bring it back by layering in Photoshop.

Update: It appears that if you go back to it then next day, and experiment with style descriptions, then try to go back to the original prompt, the earlier styled generations somewhow adversely affect the later output (more hard and cartoony than it should be). Possibly old latents are being partly re-used? Anyway… start from a fresh launch of Comfy, then go to the workflow and don’t tinker or change anything before starting your output.

Update: It seems a Poser .PNG render with transparency is the best to drop in as the seed image. Rather than needing to first place it onto a white background. Also, “filter” rather than “convert” seems a better choice of words for the prompt.