>>35278
model: WAI-NSFW-illustrious-SDXL
workflow:
basically this with minor modifications https://comfyanonymous.github.io/ComfyUI_examples/2_pass_txt2img/ (the first one)
Inpainting would be a separate workflow, which I basically never do.
lora:
you can get good chubby or fat (not obese) girls out of the box, as well as many artist styles.
for anything else try https://civitai.com/models/645787/hyper-girls or symix https://bbw-chan.link/bbwai/res/23982.html#34472 or simply go to civitai and explore
the fewer loras the better and also don't necessarily use them at 1.0 strength
prompt engineering is probably the most important part:
you should make yourself familiar with how danbooru uses tags to describe poses, image composition, body types, actions and such. I don't use free-form textual descriptions at all. Keep the prompt simple and structured. Follow the advice given with the checkpoint and don't copy-paste those huge pre-made positive/negative prompts you see elsewhere. I'll attach an example.
other than that just try out different things and see what changes. you can set the seed to a fixed value and play with the prompt to see what effect a certain keyword/lora actually has. this is more fun if you have decent hardware, for example this pic took me 20s to generate.