---
base_model:
- Qwen/Qwen-Image-Edit
- lightx2v/Qwen-Image-Lightning
library_name: diffusers
license: apache-2.0
pipeline_tag: image-to-image
tags:
- qwen-image
- image-to-image
- image-editing
- quantized
- fp8
- e4m3
- lora
- lightning
- diffusers
base_model_relation: quantized
---
# Qwen-Image-Edit-l8v1.1-e4m3

[Qwen-Image-Edit](https://huggingface.co/Qwen/Qwen-Image-Edit) with the
[Qwen-Image-Lightning 8-step V1.1](https://huggingface.co/lightx2v/Qwen-Image-Lightning)
LoRA fused into the transformer, then quantized to FP8 (e4m3). Saved as a
complete `QwenImageEditPipeline`.

`l8v1.1` in the repo name is `Lightning, 8 steps, V1.1`.

## What was changed

1. `Qwen-Image-Lightning-8steps-V1.1.safetensors` is loaded as a LoRA, fused
   into the base transformer, and unloaded — so the published weights carry the
   step distillation directly and no LoRA is needed at inference.
2. All 60 transformer blocks are then quantized to `e4m3_e4m3_dynamic`
   (`float8_e4m3fn` weights, dynamically scaled `float8_e4m3fn` activations).

The Qwen2.5-VL text encoder, the processor and the VAE are untouched and stay
in bf16. Weights are pickled `.bin` shards, so loading requires
`use_safetensors=False`.

## Usage

Run it at roughly 8 steps with CFG off — that is what the fused LoRA was
distilled for. Running it at 40 steps like the undistilled model wastes compute
and does not improve the result.

```python
import torch
from diffusers import QwenImageEditPipeline
from diffusers.utils import load_image

pipe = QwenImageEditPipeline.from_pretrained(
    "wavespeed/Qwen-Image-Edit-l8v1.1-e4m3",
    torch_dtype=torch.bfloat16,
    use_safetensors=False,
).to("cuda")

out = pipe(
    image=load_image("input.png"),
    prompt="replace the sky with a clear night sky",
    num_inference_steps=8,
    true_cfg_scale=1.0,
).images[0]
```

## License

Apache-2.0. Both Qwen-Image-Edit and Qwen-Image-Lightning are Apache-2.0.