re-rendering some Lily scenes using stable diffusion, Comfyui, (credits to original creator), comparing different models
The second model does better than the first, the third is also good, I'd say.
the original for the 2nd image is shown in
#23
For this I recognize pose, create a depth map, and a canny; comfyui custom nodes, all just from the source image.I had previously captioned these images, which I now use as part of the prompt. The pose, canny and depth map are used fro controlnet conditioning. I use the source image as latent input, denoise around 0.6, more info about params in the image metadata (exif data).