Qwen-Image-2.1-Nunchaku

Language: English | 中文

9.20 GiB less after load (46.4%) · text-to-image 2.69× faster · editing 2.19× faster

FP4 on an RTX 5090, compared with the BF16 transformer.

4-bit weight-and-activation quantization of the Qwen-Image-2.1 transformer, rank 32, for inference with Nunchaku. The repo has two checkpoints: FP4 for Blackwell (50-series) and INT4 for 30- and 40-series GPUs. The VAE stays in BF16. Diffusers components are included, so the repo loads as a pipeline.

Model

• This repo: tonera/Qwen-Image-2.1-Nunchaku
• FP4 transformer (Blackwell / 50-series): svdq-fp4_r32-Qwen-Image-2.1-Nunchaku.safetensors
• INT4 transformer (30- and 40-series): svdq-int4_r32-Qwen-Image-2.1-Nunchaku.safetensors
• Upstream: Qwen/Qwen-Image-2.1

Qwen-Image-2.1 is a unified text-to-image and image-editing model from Qwen. The visual generator is about 7B parameters, a 32-layer single-stream DiT. The same weights do text-to-image, reference-image editing, and transparent (RGBA) output. The official example uses 40 steps, true_cfg_scale=1, and prefix KV cache.

Samples

A blonde girl holds a wooden sign in her hand, which reads:：CreatePPTX.com

BF16	FP4

Performance

1024×1024, true_cfg_scale=1, 30 steps, prefix KV cache on, CPU offload off, dtype bfloat16. BF16 and FP4 share the same INT4 text encoder, svdq-int4-Qwen3-VL-8B-Instruct-Nunchaku.safetensors. Measured on an RTX 5090.

Text-to-image

	BF16	FP4
Load	9.549s	5.523s	1.73×
Cold inference	11.143s	4.273s	2.61×
Steady inference	10.973s	4.085s	2.69×

Image editing

The condition image is the BF16 sample above. Prompt: Turn the girl's hair black. Other settings match text-to-image.

	BF16	FP4
Load	5.366s	4.579s	1.17×
Cold inference	13.349s	6.206s	2.15×
Steady inference	13.043s	5.964s	2.19×

Memory

	BF16	FP4	Saved
Allocated after load	19.83 GiB	10.63 GiB	9.20 GiB (46.4%)
Reserved after load	19.89 GiB	10.68 GiB	9.22 GiB (46.3%)
Text-to-image peak allocated	26.44 GiB	17.23 GiB	9.22 GiB (34.8%)
Text-to-image peak reserved	27.92 GiB	18.46 GiB	9.46 GiB (33.9%)
Edit peak allocated	28.45 GiB	17.24 GiB	11.21 GiB (39.4%)
Edit peak reserved	30.47 GiB	18.46 GiB	12.01 GiB (39.4%)

Quality

Same FP4 run on an RTX 5090, 25 images each, compared with the BF16 transformer under the same settings. Higher PSNR and SSIM are closer. Lower LPIPS is closer. p90 is the 90th percentile of that metric.

Task	Metric	Mean	p50	p90	Best	Worst
Text-to-image	PSNR ↑	22.54	22.03	27.32	29.34	13.99
Text-to-image	SSIM ↑	0.877	0.896	0.964	0.974	0.599
Text-to-image	LPIPS ↓	0.164	0.162	0.299	0.041	0.435
Image editing	PSNR ↑	19.94	20.15	24.05	27.64	14.19
Image editing	SSIM ↑	0.752	0.764	0.890	0.918	0.527
Image editing	LPIPS ↓	0.147	0.113	0.281	0.040	0.348

Installation

Nunchaku must provide NunchakuQwenImage21Transformer2DModel, and Diffusers must provide QwenImage21Pipeline (after 2026-09-18, including PR #14804). Prebuilt wheels are in tonera/vitoom-nunchaku.

pip install "git+https://github.com/huggingface/diffusers.git"

python -m pip uninstall -y torch torchvision torchaudio nunchaku
python -m pip install "torch==2.14.*" "torchvision==0.29.*" --index-url https://download.pytorch.org/whl/cu130

# Pick the wheel that matches your Python, CUDA, and PyTorch from tonera/vitoom-nunchaku
hf download tonera/vitoom-nunchaku \
  nunchaku-1.4.0+cu13.0torch2.14-cp311-cp311-linux_x86_64.whl \
  --local-dir ./wheels

python -m pip install ./wheels/nunchaku-1.4.0+cu13.0torch2.14-cp311-cp311-linux_x86_64.whl

Usage

import torch
from diffusers import QwenImage21Pipeline
from PIL import Image
from nunchaku import NunchakuQwenImage21Transformer2DModel
from nunchaku.utils import get_precision

REPO = "tonera/Qwen-Image-2.1-Nunchaku"
NAME = "Qwen-Image-2.1-Nunchaku"

transformer = NunchakuQwenImage21Transformer2DModel.from_pretrained(
    f"{REPO}/svdq-{get_precision()}_r32-{NAME}.safetensors",
    torch_dtype=torch.bfloat16,
    device="cuda",
)
pipe = QwenImage21Pipeline.from_pretrained(
    REPO,
    transformer=transformer,
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

image = pipe(
    prompt="A blonde girl holds a wooden sign in her hand, which reads:：CreatePPTX.com",
    width=1024,
    height=1024,
    num_inference_steps=30,
    true_cfg_scale=1.0,
    use_kv_cache=True,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
image.save("qwen21-w4a4.png")

ref = Image.open("input.png").convert("RGB")
edited = pipe(
    prompt="Turn the girl's hair black",
    image=ref,
    width=1024,
    height=1024,
    num_inference_steps=30,
    true_cfg_scale=1.0,
    use_kv_cache=True,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
edited.save("qwen21-edit-w4a4.png")

License

These quantized weights are derived from Qwen/Qwen-Image-2.1 and follow the Qwen Research License. license: other in the model-card metadata is only a placeholder for the Hugging Face license field. The license linked above is what applies.