nunchaku-qwen-image-edit-2511

nunchaku-qwen-image-edit-2511 (SVDQuant W4A4 NVFP4)

4-bit SVDQuant of Qwen/Qwen-Image-Edit-2511 for the nunchaku inference engine. Same file layout and usage as the official nunchaku-tech/nunchaku-qwen-image-edit release — load it with stock from_pretrained (one small forward shim is needed for the edit pipeline, see Usage).

• Weights: W4 (NVFP4 e2m1) + low-rank branch (rank 32 or rank 128), activations A4 (NVFP4)
• Modulation layers: W4 AWQ (symmetric, group 64) + super-channel fold — fully baked in, self-contained
• Requires: Blackwell GPU (RTX 50-series sm_120 / GB10 sm_121) for NVFP4
• Measured: 4.82 s/step on DGX Spark GB10 (1.48× vs BF16, see benchmarks); on real-world edits quality is on par with the BF16 base model

Models

Both checkpoints are W4A4 NVFP4 and differ only in the SVDQuant low-rank branch size; on real edits they are visually equivalent. NVFP4 needs a Blackwell GPU (RTX 50-series sm_120 / GB10 sm_121); INT4 builds are not provided yet.

Usage

import torch
from diffusers import QwenImageEditPlusPipeline
from diffusers.utils import load_image
from nunchaku.models.transformers.transformer_qwenimage import NunchakuQwenImageTransformer2DModel
from nunchaku.utils import get_precision

rank = 32  # 32 (smaller/faster) or 128 (higher fidelity)
transformer = NunchakuQwenImageTransformer2DModel.from_pretrained(
    f"stuqiu/nunchaku-qwen-image-edit-2511/svdq-{get_precision()}_r{rank}-qwen-image-edit-2511.safetensors"
)

# nunchaku is currently only lightly maintained and its API lags behind diffusers: its forward still
# expects the (now deprecated) `txt_seq_lens`, which QwenImageEditPlusPipeline in diffusers >= 0.37 no
# longer passes (it switched to `max_txt_seq_len`). We back-fill it with the exact value diffusers
# itself used to compute (`mask.sum(-1)`), so the output is unchanged — no other side effect.
_orig = transformer.forward
def _forward(*a, **k):
    if k.get("txt_seq_lens") is None and k.get("encoder_hidden_states_mask") is not None:
        k["txt_seq_lens"] = k["encoder_hidden_states_mask"].sum(-1).tolist()
    return _orig(*a, **k)
transformer.forward = _forward

pipe = QwenImageEditPlusPipeline.from_pretrained(
    "Qwen/Qwen-Image-Edit-2511", transformer=transformer, torch_dtype=torch.bfloat16
).to("cuda")

image = load_image("input.png").convert("RGB")
out = pipe(
    image=image,
    prompt="把天空换成绚丽的黄昏晚霞",
    num_inference_steps=40, true_cfg_scale=4.0, negative_prompt=" ",
).images[0]
out.save("edit.png")

Tested with diffusers==0.38.0, nunchaku==1.3.0.dev (cu13 / torch2.13), torch==2.13.

Benchmarks

1024×1024 edit, 40 steps, true_cfg 4.0, stock pipeline:

Samples

Real photos, identical seed/params, only the transformer differs (BF16 vs this W4A4 model). Columns: input · BF16 · nunchaku r32 · nunchaku r128.

On typical edits (sky / object / background / color / decoration) the quantized model is visually on par with BF16.

Quantization pipeline

deepcompressor SVDQuant calibration (32 samples, smoothing grid 20, rank-32 / rank-128 low-rank branches) for attention/MLP, official symmetric AWQ recipe (scale = absmax/7, zero ≡ 7) for modulation, plus the temb super-channel fold so the 4-bit modulation stays accurate and the checkpoint loads fully self-contained.

Acknowledgements

• Qwen team for Qwen-Image-Edit-2511
• Nunchaku / MIT HAN Lab for SVDQuant, deepcompressor and the nunchaku inference engine
• Example inputs: COCO val2017