Update: I just uploaded the latest x0 version. Because the file has no version number to distinguish it, I appended the SHA256 hash to the filename.

Note: The "expanded" version simply has some components (basically the extra tensors that are specific to Radiance) left uncompressed. Both versions give identical outputs, just that I find the non-expanded version to tank in speed with tiling enabled, which for some setups might not be worth the ~0.4GB size decrease.

For more information (including how to compress models yourself), check out https://huggingface.co/DFloat11 and https://github.com/LeanModels/DFloat11

Feel free to request for other models for compression as well, although compressing models that do not use the Flux architecture might be slightly tricky for me.

How to Use

ComfyUI

Install my own fork of the DF11 ComfyUI custom node: https://github.com/mingyi456/ComfyUI-DFloat11-Extended. After installing the DF11 custom node, use the provided workflow json, or simply replace the "Load Diffusion Model" node of an existing Chroma workflow with the "DFloat11 Model Loader" node. If you run into any issues, feel free to leave a comment. The workflow is also embedded in the below png image.

diffusers

As far as I know, this model architecture is currently not yet supported in diffusers.

Compression Details

This is the pattern_dict for compressing Chroma-based models in ComfyUI:

pattern_dict_comfyui = {
        r"distilled_guidance_layer\.layers\.\d+": (
            "in_layer",
            "out_layer"        
        ),
        r"double_blocks\.\d+": (
            "img_attn.qkv",
            "img_attn.proj",
            "img_mlp.0",
            "img_mlp.2",
            "txt_attn.qkv",
            "txt_attn.proj",
            "txt_mlp.0",
            "txt_mlp.2",
        ),
        r"single_blocks\.\d+": (
            "linear1",
            "linear2",
        ),
        r"nerf_blocks\.\d+": (
            "param_generator", 
        )
    }