toxicdog/MOSS-Image-SFT-INT4-ConvRot-ComfyUI

MOSS-Image-SFT INT4 ConvRot for ComfyUI

Repackaged model files for ComfyUI and others, with INT4 ConvRot converted for testing in Radiant Canvas for fast macOS inference: Radiant Canvas on the App Store.

The original model is OpenMOSS-Team/MOSS-Image-SFT, the supervised fine-tune of MOSS-Image-Base from OpenMOSS.

File

• diffusion_models/moss_image_sft_int4_convrot.safetensors: the MOSS-Image-SFT diffusion transformer as a single file in ComfyUI's native ConvRot W4A4 format.

For ComfyUI, place the file in:

ComfyUI/models/diffusion_models/

ComfyUI support comes from the official MOSS-Image plugin (search for MOSS-Image in ComfyUI Manager, or see the ComfyUI guide). Load the file with the plugin's diffusion model loader. It uses the same tensor names as the plugin's export.py output, and the plugin reads ComfyUI's native quantization metadata. You need a ComfyUI build with INT4 ConvRot support.

The Qwen3.5 text encoder and the VAE are not included here. Get them from the original model, or export them with the plugin's export.py.

The original model recommends 50 steps, CFG 3, shift 5 and negative-branch skipping at layer 22.

Conversion details

• Original model revision: daf706d2e1c7a6369c17b6d1658671aee6db1ebe
• Source: BF16 diffusers transformer, 3 shards, 463 tensors, merged into one file
• Quantized layers: 252 (attention q, k, v, o and the MLP ffn_up_gate and ffn_down in all 42 blocks)
• Format: comfy-kitchen 0.2.37 TensorCoreConvRotW4A4Layout
• ConvRot group size: 256
• Scales: one float32 scale per output row (the layout's INT4 codes are -7 to 7)
• Scale choice: each row's scale is searched and refined to minimize the squared error of the rotated weights, instead of using the row's largest value. This lowers the mean relative weight error from 15.8% to 11.4% and removes diagonal streaks and smearing seen with the first upload.
• Output size: 3,035,537,312 bytes
• SHA-256: bc3862d71000723369f8d8d42adc1680aef8fd401e68d9be44d8d676175c154b

The text embedder, time embedding, shared and head modulation, input embedder, output projection and normalization weights are kept in BF16.

This is a test conversion. Check output quality for your own prompts and workflows before relying on it.

License

The original model is released under the Apache License 2.0. A copy of the license is included in this repository.