reb82/qwen-image-edit-2511-lightning-fp8

Qwen-Image-Edit-2511 · Lightning-8step · fp8 (pre-quantized transformer)

The transformer only of Qwen/Qwen-Image-Edit-2511, with the lightx2v/Qwen-Image-Edit-2511-Lightning 8-step LoRA pre-fused and then fp8-quantized with torchao Float8DynamicActivationFloat8WeightConfig.

Pre-quantizing once removes the ~2–3 min runtime fuse+quantize warmup: this loads in ~1s and is bit-identical to quantizing at runtime (verified MAD 0.000/255). ~20GB vs ~41GB bf16.

Use

import torch
from diffusers import QwenImageEditPlusPipeline, QwenImageTransformer2DModel

# pickle (.bin) format — safetensors can't serialize torchao tensor subclasses,
# so use_safetensors=False is REQUIRED.
transformer = QwenImageTransformer2DModel.from_pretrained(
    "reb82/qwen-image-edit-2511-lightning-fp8",
    torch_dtype=torch.bfloat16, use_safetensors=False)

# the text encoder / VAE / tokenizer / processor / scheduler come from the base repo
pipe = QwenImageEditPlusPipeline.from_pretrained(
    "Qwen/Qwen-Image-Edit-2511", transformer=transformer, torch_dtype=torch.bfloat16)
pipe.to("cuda")
img = pipe(image=src, prompt="…", true_cfg_scale=1.0, num_inference_steps=8).images[0]

Requirements

• torchao installed (the weights are Float8Tensor subclasses).
• fp8-capable GPU (Ada/Blackwell, sm_89+) — fp8 dynamic-activation has no Ampere kernels.
• 8 steps @ true_cfg_scale=1.0 (the LoRA is CFG-distilled).

Format note

Saved with safe_serialization=False (PyTorch pickle) because safetensors cannot hold torchao tensor subclasses. Load only from a source you trust.

License & attribution

Apache-2.0, a derivative of two Apache-2.0 works: Qwen/Qwen-Image-Edit-2511 and lightx2v/Qwen-Image-Edit-2511-Lightning.