Tried to generate a realistic-looking Jessica Rabbit. Any ideas on how I could improve this?
Prompt info:
- Positive Prompt:
Jessica Rabbit in a red sequin dress and red high heels sitting facing the camera and slightly leaning back, slim waist, (big boobs:1.1), (deep cleavage:1.5), smiling, god rays, volumetric lighting, photorealistic, 8k, hdr - Negative prompt:
bad anatomy, bad hands, three hands, three legs, bad arms, missing legs, missing arms, poorly drawn face, bad face, fused face, cloned face, worst face, three crus, extra crus, fused crus, worst feet, three feet, fused feet, fused thigh, three thigh, fused thigh, extra thigh, worst thigh, missing fingers, extra fingers, ugly fingers, long fingers, horn, realistic photo, extra eyes, huge eyes, 2girl, amputation, disconnected limbs - Steps: 35
- Sampler: DPM++ 2S a Karras
- CFG scale: 7
- Seed: 2155243983
- Size: 768x1024,
- Model hash: 255f68ed9a
- Model: dynavisionXLAllInOneStylized_beta0411Bakedvae
- Clip skip: 2
The face is a little flat and plastic feeling, so it would be nice to get a more natural skin texture. I would also like to bring out a bit nicer of a smile, and zoom out a bit more to show off some more of her legs without the legs fusing - I was fighting with that a fair amount. There were a couple I generated with other models that had her wearing stilleto heels that made who whole figure gorgeous, but there were so other parts that broke the realism.
As far as tooling goes, I've not yet had a chance to use ComfyUI yet, and just used AUTOMATIC1111 on my M2 MacBook Pro.
It's worth considering switching models - dynavision tends to go that way with faces. You might also have some luck with prompt engineering e.g. removing 'photorealistic'. It sounds counterintuitive but I've definitely seen that term make things look more "CGI" - which makes some sense, you dont normally call a real photo "photorealistic". I sometimes try things like "portrait photography" instead. For some better skin texture, sometimes adding a refiner will help as well.
I also noticed you listed your resolution as 1024x768. That's not one of the "recommended" SDXL resolutions ( https://stablediffusionxl.com/sdxl-resolutions-and-aspect-ratios/ ), you might have better luck with 1152x896. Not likely to make a substantial difference with the results youve gotten so far, though.
The real win with what you have so far will probably be changing techniques: A common one is to generate your image as close as possible with txt2img, and then to get a better face, send the image to img2img inpainting. Paint out the face and you can get much more specific with a facial generation on top of the previous image, changing models/lora/etc if needed.