Confused about the different model versions? Check the 'About this version' on the right sidebar for more information.
This is just a quick release that we did since Flux + SD3.5 is taking such a long time (but we are working on it!)
Hopefully this satisfies those furry desires.
Take care!
Model Info
Chroma XL is an experimental model that is a partial merge + train of the brilliant NoobAI-XL (NAI-XL) model by L_A_X and their team, NoobAI is based off of Illustrious. All credit to the model goes to him.
The model only needs low CFG and low step count. Making it highly versatile over a wide range of consumer hardware.
We thank L_A_X and their team for his fantastic efforts in creating this model!
Recommended Settings
We recommend using the following settings with the model for best results.
Resolution: 1024x1024
Steps: 8-12
CFG Guidance: 1.7-3.0
Sampler: Euler A
Scheduler: beta
The model utilises the low guidance and steps. It's advised to keep your step count as well as your guidance low, whilst you may increase the step count if you choose to do so, there is a clear point of diminishing returns, therefore we recommend you to follow our recommended values here.
It is strongly recommended to use an ancestral sampler (e.g. Euler a) with beta schedule, the noise injection in ancestral sampler is absolutely needed to prevent over saturation.
Usage Instructions
All the sample images have their workflow included within them. It’s recommended to download those images and use that workflow in ComfyUI.
For all other clients, just run it as you would with any other SDXL model.
Strengths & Weaknesses
During our evaluations of the model, we find the following strengths and weaknesses in the model.
Strengths
High-level understanding of danbooru and e6 tags.
Relatively good anatomy compared to previous models.
Good understanding of many booru concepts, especially ones related to anime and furry.
Ability to do duo quite coherently, especially compared to prior models.
Ability to do high quality backgrounds.
Weaknesses
Non-existent natural language capabilities (don’t even try to use it).
Slightly incompatibility with SDXL model extensions, such as ControlNet but give it a go.
Characters that are in the distance tend to become extremely blurry.
The whole model has sort of a burned in ‘haze’ which can cause upscaling to look extremely pixelated.
Can look overcooked sometimes.
Prompting & Tips
The model has next to no knowledge of natural language so you should exclusively use tags with this one.
It’s recommended to go over the guidelines provided within the base models CivitAI page here.
https://civarchive.com/models/833294?modelVersionId=998979
The model was trained with artist tags.
It’s recommended to always include solo, duo, trio, etc. Since not adding any of those tags results in a very unpleasant amalgamation of them.
Credits
This model was merged + trained by the Horizon Team.
All credit for the base model goes to L_A_X and their team.
https://civarchive.com/models/833294?modelVersionId=998979
Contact Us
You can reach us at our Discord server!
Description
Mango: Sweet and delicious! A little bit further away from NoobAI 1.0 compared to Spud. It's recommended to try both of them and see which one you personally prefer, there isn't a right answer for which one is better, since both have their own associated pros and cons.
This uses NoobAI's "Epsilon-pred 1.0-Version" rather than "Epsilon-pred 0.75-Version". It's one of the two options provided.
There seems to be a much greater influence from anime overall which does reduce the effectiveness of some furry prompts. It's recommended to try both models out to see which you prefer.
There are improvements in overall fidelity and a reduction of the 'haze' that caused problems with some upscalers in v3.31.
FAQ
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.