Chroma - Nerogar/OneTrainer GitHub Wiki

Chroma is now available on the main branch. Chroma1-HD is an 8.9B parameter text-to-image foundational model based on Flux, but it is fully Apache 2.0 licensed, ensuring that anyone can use, modify, and build upon it.

This page is a work in progress as information as the Chroma fork is tested and Chroma continues to be developed.

OneTrainer Details for Flux:

• Huggingface Link: lodestones/Chroma1-HD
• PR: https://github.com/Nerogar/OneTrainer/pull/945
• Do not use FULL layers, as this will include the distilled_guidance_layer, which is not recommended to train. Needs lora layers update. Hardcoded for finetune not to include.
• Lora's work in Comfy/Swarm (same key structure as Flux)
• Sample Settings for OneTrainer - CFG 3, Steps 40 (Sampler is always Euler, no matter what you choose)
• Token Limit for Chroma: 512
• Training at 1024px is slow, even compared to Flux. Note: Most foundational training for Chroma was 512px, before finishing at 1024px.
• Use of Dynamic Timestep Shift is not recommended (Chroma was not trained using it)

Current Tests:

• Lora works in ComfyUI (Swarm)
• Dora works in ComfyUI (Swarm)
• Lora works in SD.Next
• Lora does not work in Forge (depends on Fork?)

VRAM usage:

• See presets for VRAM settings, with 8GB, 16GB and 24GB presets being created for both Finetune and Lora

Other Notes:

• FP8 Finetune "works", but likely requires more research and using various versions of fp8 for forward pass vs. backward pass. Any use of this is considered strictly research. Recommended to use 8GB or 16GB templates with BF16