[tayo42]

With the luggage example it seems to only generate backgrounds where the lighting makes sense? That's kind of interesting. I was wondering how it would handle the highlight on the right.

[minimaxir]

Giving Stable Diffusion constraints forces it to get creative.

It’s the best argument against “AI generated images are just collages”.

[Der_Einzige]

This is a general result. For example, ChatGPT struggles hard with following lexical, syntactic, or phonetic constraints in prompts due to the tokenization scheme - see https://paperswithcode.com/paper/most-language-models-can-be...

LLMs + Diffusors are super charged when using techniques like constraints, controlnet, regional prompting, and related techniques.

[zamalek]

In ComfyUI you could run the image through a style-to-style (sdxl refinement might even pull it off) model to change the lighting without changing the content. Or use another ControlNet. Your workflow can get arbitrarily complex.