Abiray/MOSS-Image-SFT-INT8-convrot

MOSS-Image-SFT (INT8 ConvRot) for ComfyUI

This repository provides INT8 ConvRot quantizations for OpenMOSS-Team/MOSS-Image-SFT

The native ComfyUI workflow file is included in this repository:

moss-image_comfyUI_workflow.json.

---

📂 Repository Structure

The files are organized into standard ComfyUI directory paths:

Abiray/MOSS-Image-SFT-INT8-convrot/
├── diffusion_models/
│   └── moss-image-sft-dit-int8-convrot.safetensors       # 5.41 GB (INT8 DiT Backbone)
├── text_encoders/
│   └── moss-image-sft-qwen35-int8-convrot.safetensors     # 8.45 GB (INT8 Qwen3.5 Text Encoder)
├── vae/
│   └── moss-image-sft-vae.safetensors                    # 336 MB  (FLUX.2 VAE, BF16)
└── moss-image_comfyUI_workflow.json                      # Ready-to-use ComfyUI workflow

---

🔬 Quantization Summary

Standard INT8 round-to-nearest (RTN) or naive absmax clipping often causes severe structural degradation in diffusion models due to high-magnitude activation outliers across specific channels.

ConvRot applies orthogonal Hadamard matrix rotations to the weights and activation spaces prior to quantization. This flattens outlier peaks, allowing fast 1-pass absmax INT8 mapping to maintain near-lossless fidelity.

Component	Quantized Layers	Target Precision	Original BF16	INT8 ConvRot	Mean Relative Error
DiT Transformer (5B)	252 GEMM linear projections across 42 layers	INT8 (.comfy_quant)	~10.0 GB	5.41 GB	0.868% (cos ~1.0005)
Qwen3.5 Text Encoder	308 GEMM projections + rotated per-row embeddings	INT8 (.comfy_quant)	~16.0 GB	8.45 GB	0.868% (cos ~1.0002)
VAE	Preserved in original BF16	BF16	336 MB	336 MB	Lossless (0.00%)

All modulation layers, LayerNorms/RMSNorms, RoPE bands, and attention scale vectors are retained in original BF16/FP32 precision.

---

🖼️ Sample Generations

Prompt

Showcase Image 1

Prompt

Showcase Image 2

Prompt

Showcase Image 3

---

⚙️ Recommended Inference Settings

Parameter	Recommended Value	Notes
Sampler	Euler / Flow-matching	Default ODE solver
Sampling Steps	30 – 50	50 steps provide maximum detail
CFG Scale	3.0	Recommended range: 2.5 – 3.5
Flow Shift	5.0	Standard flow sigma shift
Resolutions	1024x1024, 768x1360, 1360x768	Width and height must be multiples of 16

---

🚀 Setup & Usage in ComfyUI

1. Download Models

Run this Python script inside your ComfyUI environment to place each model directly into its respective folder:

import os
from huggingface_hub import hf_hub_download

COMFY_ROOT = "/path/to/ComfyUI/models"  # e.g., /tmp/workspace/ComfyUI/models
REPO_ID = "Abiray/MOSS-Image-SFT-INT8-convrot"

files = [
    ("diffusion_models", "moss-image-sft-dit-int8-convrot.safetensors"),
    ("text_encoders", "moss-image-sft-qwen35-int8-convrot.safetensors"),
    ("vae", "moss-image-sft-vae.safetensors"),
]

for subfolder, filename in files:
    target_path = f"{subfolder}/{filename}"
    print(f"⬇️ Downloading {filename}...")
    hf_hub_download(
        repo_id=REPO_ID,
        filename=target_path,
        local_dir=COMFY_ROOT,
    )

print("✅ All models downloaded successfully!")

2. Run Inference

1. Download moss-image_comfyUI_workflow.json from this repository.
2. Drag and drop the .json file into your ComfyUI window.
3. Verify node selections:
   • Load Diffusion Model: moss-image-sft-dit-int8-convrot.safetensors
   • CLIPLoader / Text Encoder: moss-image-sft-qwen35-int8-convrot.safetensors
   • VAELoader: moss-image-sft-vae.safetensors
4. Queue Prompt.

---

📄 License & Attribution

• Base model developed by the OpenMOSS Team: OpenMOSS-Team/MOSS-Image-SFT.
• Released under the Apache-2.0 License.