qwen-image-layered-int8convrot

Qwen Image Layered INT8 ConvRot for ComfyUI

This repository contains a stock-ComfyUI-native INT8 TensorWise + ConvRot conversion of the official qwen_image_layered_bf16.safetensors checkpoint. It loads directly in ComfyUI 0.27.0 or newer; no quantization custom node is required.

File

The checkpoint contains 839 INT8 weights, 839 FP32 per-output-channel scales, 839 ConvRot metadata tensors, and 1,095 retained BF16 tensors. Every selected weight was rotated with regular-Hadamard groups of 256 before quantization.

Requirements

• ComfyUI >= 0.27.0
• comfy-kitchen >= 0.2.15
• NVIDIA GPU with compute capability 7.5 or newer for the optimized TensorWise INT8 kernels

Install this file in ComfyUI/models/diffusion_models/. The standard Qwen Image Layered workflow also needs:

• qwen_2.5_vl_7b_fp8_scaled.safetensors
• qwen_image_layered_vae.safetensors

Select qwen_image_layered_int8convrot.safetensors in the workflow's Load Diffusion Model node.

Conversion provenance

• Base model: Qwen/Qwen-Image-Layered, revision 8f0ca708dfff6ba1dd5f2d85d78f8c108a040bcf
• ComfyUI source checkpoint: Comfy-Org/Qwen-Image-Layered_ComfyUI, revision 5d93d59c0e34488b94dee03ad33337f7ee997f97
• Converter: Comfy-Org/comfy-quants, revision 1e0d481f24847c4914578f5468917902ad53ea46
• Recipe: qwen_image_layered_int8_tensorwise.yaml

comfy-quants export-model-int8-tensorwise \
  --config configs/qwen_image_layered_int8_tensorwise.yaml \
  --source qwen_image_layered_bf16.safetensors \
  --out qwen_image_layered_int8convrot.safetensors \
  --device cuda:0 \
  --convrot \
  --convrot-groupsize 256 \
  --hash-output

Validation

The safetensors contract and checksum were verified after export. The checkpoint was then loaded by stock ComfyUI, which detected all 839 INT8 weights, and the official image_qwen_image_layered workflow completed with a local 640-pixel input on an RTX A5000. The smoke test produced three non-empty 640x608 RGBA layers.

This conversion retains the upstream Apache-2.0 license.