More fun with ‘Poser to Stable Diffusion’, now that I’ve moved to Windows 11 Superlite and have the AI stuff mostly set up.
This time I can use SDXL rather than SD 1.5. I think regular readers of this blog will recall the previous attempts with the same Poser source, and see quite a difference in the result. I’m using the same test render.
To get this I made a ComfyUI workflow featuring an SDXL turbo model powering Img2Img, plus three LoRAs, and a lineart Controlnet. Not sure the latter is really needed (a relic of the old workflow), provided the colour stays steady from image to image and thus from panel-to-panel and page-to-page in a comic. Or I guess I could go all-in and try four different Controlnets working at once, and see how stable the results are compared to the Poser render.
But this is just a first experiment, and it’s encouraging to get this far immediately.
On the other hand, it’s inventing things like the suit pockets and a waistcoat. Which is annoying since consistency is needed. The reason to use Poser is to have the results be consistent, not full of little differences that either take a lot of postwork to fix, or which are lazily left in and annoy the heck out of the reader. (Update: prompt to “dark 2-piece suit” to get rid of waistocoats)
The result comes in at a healthy 1432px (in about 12 seconds), from a 768px starter Poser render. Meaning that cutout and de-fringing is easier in Photoshop. Here the result is cutout, defringed, and given a Stroke to firm the holding-line. The shadows have also been lifted a little, to give it a more graphic look.
Next step will be to get some more SDXL Controlnets, and output a variety of different Poser renders and then see what combination works the best with this workflow.