>>106203690
>>106203986
>>106204003
here's a similar prompt.
chroma really does struggle to put her behind the glass, even when she's behind it she destroys the reflections. lots of bad hand results too from every version I tried.

I like chroma but we're at v50 and hands and text still aren't consistent. its understanding of complex multilayered, multisubject prompts has improved significantly. but I also have started to find I have to rework old prompts when I'm testing them on the new releases. that really shouldn't happen if prompt comprehension is improving across the board, I'm still engineering prompts to avoid triggering chroma's weird biases and meltdowns.

every time we get a new model like a chroma release or qwen I try it out. but they're still very incomplete compared to SDXL anime models lmao. we probably have another year or two before we get a SOTA model that can do these things:
>styles
>artist names
>tags instead of boomerprompting
>hands
>bonus (difficulty: impossible): it's not bloated and slow

hilarious how SAAS models other than MJ and NovelAI have the same problems too.