What is Chroma1-HD?

Flux.1 [dev] was a great model, but it had two major issues for widespread community use.

  • Since it is strongly distilled, compatibility with LoRA and full fine-tuning is not very good.
  • Non-commercial license.

Therefore, several attempts appeared to reconstruct "Undistilled Flux" starting from Flux.1 [schnell] which has a more generous license, and one of the most labor-intensive ones among them is Chroma1-HD by LodestoneRock.

Since it has been re-trained including NSFW and anime images, it has a higher degree of freedom in style than the original Flux series, and it fits better to perceive it as a new general-purpose model based on Flux architecture rather than "Undistilled Flux".

Model Download

Unlike Flux.1, Chroma1-HD does not use CLIP, but uses only T5.

📂ComfyUI/
└── 📂models/
    ├── 📂diffusion_models/
    │   └── Chroma1-HD.safetensors
    ├── 📂text_encoders/
    │   └── t5xxl_fp8_e4m3fn_scaled.safetensors
    └── 📂vae/
        └── ae.safetensors

text2image

chroma1_hd.json
  • The basic construction is almost the same as Flux.1.
  • 🟦 T5TokenizerOptions is a node that fills with padding when the number of tokens is less than the set value. There is no big difference whether you put it in or not.
  • CFG is specified as 4.0, but some say 6 to 7 is good. Please try various things.