Chroma

[dxqb]

Use and test:
git clone https://github.com/dxqb/OneTrainer -b chroma
./update.sh
Select the #chroma preset

select a LoRA layer filter, not full, otherwise distilled_guidance_layer is trained which is not recommended.

code comments:

• Chroma is different enough from Flux that I made it a separate model, instead of changing the Flux model code
• therefore there is a lot of code duplication - but this duplication also exists between other models, so I've decided against touching that in this PR. More code sharing between models can be done as a separate PR

open points:

• support Chroma timestep distribution? https://huggingface.co/lodestones/Chroma?not-for-all-audiences=true#tldr-a-custom-timestep-distribution-prevents-loss-spikes-during-training. LOGIT_NORMAL seems close enough though for small finetunes - done
• exclude the distilled_guidance_layer from training. For LoRA training the layer filter can be used, but not for full finetuning. Hardcode this exception, or add a layer filter for full finetuning? LoRA: Custom layer filters that support regular expressions #778 - done, hardcoded for now
• Bug: only works if Debug is switched off - fixed
• compare sampling parameters between Comfy/Swarm and OneTrainer - appears to be different on average - they are indeed different, but model creator has confirmed that our parameters are correct
• Chroma seems to scale differently based on sequence length than Flux. Is this expected? Compare with other trainers?

Chroma bs2 512 2s/it
Chroma bs2 1024 10.3 s/it
Flux bs2 512 2.1s/it
Flux bs2 1024 8.7 s/it

• "Force Attention Mask" is not useful with Chroma; remove
• implement Model loading from huggingface using parallel downloads #770 for Chroma
• there are reports that embedding training fails on Chroma - could be an issue with the code, or inherent to the model. Check embedding vectors during training - everything seems ok
• Loading Chroma from single file safetensors fails, probably a diffusers bug

[dxqb]

thank you for your work. is this stable enough to test/try training a lora with on chroma?

several people have tested it successfully. all known open issues you can find above in the task list.