Qwen-Image-Edit-e4m3

FP8 (e4m3) dynamically-quantized Qwen-Image-Edit, saved as a complete QwenImageEditPipeline.

What was changed

All 60 blocks of the QwenImageTransformer2DModel are quantized to e4m3_e4m3_dynamic — float8_e4m3fn weights with dynamically scaled float8_e4m3fn activations. The Qwen2.5-VL text encoder, the processor, the VAE and the transformer's non-block tensors are untouched and stay in bf16. The transformer drops from ~40.9 GB to ~20.5 GB.

Quantization was done with WaveSpeed's xelerate.ao.quantize. Weights are stored as pickled .bin shards, so loading requires use_safetensors=False. FP8 matmul needs Hopper (H100/H200) or newer to actually be faster than bf16.

Usage

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

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

image = load_image("input.png")
out = pipe(image=image, prompt="make it a winter scene").images[0]

wavespeed/Qwen-Image-Edit-l8v1.1-e4m3 — the same quantization with an 8-step Lightning LoRA fused in.

License

Apache-2.0, inherited from Qwen-Image-Edit.

Downloads last month: 17

Model tree for wavespeed/Qwen-Image-Edit-e4m3

Base model

Qwen/Qwen-Image-Edit

Quantized

(16)

this model

Collection including wavespeed/Qwen-Image-Edit-e4m3

Quantized weights we publish

Collection

FP8, INT8, NF4 and precision-cast variants of open models. Each card says exactly which tensors changed and which were left alone. • 16 items • Updated 3 days ago

wavespeed
/

Qwen-Image-Edit-e4m3