---
license: apache-2.0
base_model:
- Qwen/Qwen-Image-Edit-2511
- Comfy-Org/Qwen-Image-Edit_ComfyUI
base_model_relation: quantized
tags:
- comfyui
- diffusion-single-file
---

# NVFP4 Qunatization model of Qwen-Image-Edit-2511 for ComfyUI

I did rough calibration, but the quality seems okay.

### Generation speed

- Tested on RTX5090(400W), ComfyUI with torch2.9.1+cu130
- Generates 832x1280, 4steps, cfg1.0
- Second run measured by varying only the seed

```
#nvfp4
100%|███████████████████| 4/4 [00:01<00:00,  2.52it/s]
Prompt executed in 3.45 seconds
#fp8mixed
100%|███████████████████| 4/4 [00:04<00:00,  1.02s/it]
Prompt executed in 6.09 seconds
#bf16
100%|███████████████████| 4/4 [00:06<00:00,  1.62s/it]
Prompt executed in 9.80 seconds
```

### Quantized modules
```json
{
  "block_name": "transformer_blocks",
  "rules": [
    { "policy": "keep", "match": [] },
    { "policy": "float8_e4m3fn",  "match": ["img_mod.1", "txt_mlp.net.0.proj", "txt_mlp.net.2", "txt_mod.1"] },
    { "policy": "nvfp4","match": ["attn.to_k", "attn.to_q", "attn.to_v", "attn.to_out.0", "img_mlp.net.0.proj", "img_mlp.net.2"] }
  ]
}
```

### How to Reproduce

You need nVidia GPU and around 40GB RAM.

1. Download https://github.com/bedovyy/comfy-dit-quantizer
2. activate python environment from ComfyUI.
3. `quantize.py configs/qwen-image-edit-2511.json <SOURCE> uncalibrated.safetensors`
4. `add_input_scale.py calibs/qwen-image-edit-2511.json uncalibrated.safetensors <TARGET>`