- Trained on up-to-date data, data cut last month. However, characters such as Ye Shunguang (ZZZ) still require many support tags and it will only depict her head right. - Slightly improved prompt adherence. - Brighter faces, less in shadows. - More stable overall output, though somewhat more restrictive. - Fewer unintended multiple-view compositions. - Characters previously trained new in version 3.2 are now better trained in. Worse: - More sensitive to poorly structured prompts or a low number of prompt tags. - Eye colors may occasionally differ from what is specified in the prompt (can be avoided by lowering cfg to 6). - Style can be more inconsistent on some concepts, especially for characters. Notes: - Style can be inconsistent, especially for characters. Because the style merge is only strong enough to avoid breaking artist and character tags and that additional loras don't break the model to much. - Use "Euler a" as the sampler, this is the sampler the model was primarily tested on. - If only one character is desired, use “1girl” and add “solo” if necessary. - To avoid nsfw art, put "nude" into the negative The model was additional trained longer on a more unbiased dataset. Now trained on bf16, prodigy with simplified ademamix, Factored + stochastic rounding, OrthoGrad to prevent overfitting and active Kourkotas Beta.