Prompt guide?
I've been looking at the guides in civitai, and there are few tags which should be used, specially the "aesthetic" tag with the score.
They also mention stuff like
digital_media_(artwork)
photography_(artwork)
oil_painting_(artwork)
sketch_(artwork)
retro_snapshot_(artwork)
(I'm not sure if the last two are properly worded). Is there some other instructions, styles, etc?
It is a pity there is no prompt guide availabe. I have discussed with Gemini which claims to have been used partially to label the dataset but I am very uncertain whether this is just pure hallucination or actual fact.
Luckily, a few enthusiasts on /r/stable-diffusion kept posting great images and shared their experience, otherwise I would have completely slept on Chroma. I have now started testing and this model is simply amazing.
The one thing hindering broder adoption, from my point of view, is the missing prompt guide or dataset information. This is why anima gets a lot of traction much faster, as it is excellently documented and the dataset is actually public (danbooru).
@lodestones it would be great if you could at least share a couple of samples from the dataset from a few selected style categories so we get a better understanding of how the model was trained. A description of the labeling pipeline would also be super helpful. People claim we can use some booru tags, but I do not find anything in that regard in official model descriptions. If you could publish the technical report, even in an unfinished / unpolished state, it would be very helpful to maximize what we can get out of your great work!
I currently use Gemma 4 e4b to convert booru tags to natural language prompts and it works okay but I think there is much left on the table we could harness if we had better knowledge about the training dataset and prompt labels.