Qwen-Image-2.1-Nunchaku
Collection
5 items • Updated
Qwen-Image-2.1-Turbo-Nunchaku
Language: English | ä¸æ–‡
9.20 GiB less after load (46.4%) · text-to-image 2.43× faster · editing 2.02× faster
FP4 on an RTX 5090, compared with the BF16 transformer.
4-bit weight-and-activation quantization of the Qwen-Image-2.1-Turbo 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-Turbo-Nunchaku
- FP4 transformer (Blackwell / 50-series):
svdq-fp4_r32-Qwen-Image-2.1-Turbo-Nunchaku.safetensors - INT4 transformer (30- and 40-series):
svdq-int4_r32-Qwen-Image-2.1-Turbo-Nunchaku.safetensors - Upstream: Qwen/Qwen-Image-2.1-Turbo
Qwen-Image-2.1-Turbo is the 8-step distilled version of Qwen-Image-2.1. The visual generator is still about 7B parameters, a 32-layer single-stream DiT. The same weights do text-to-image, reference-image editing, and transparent (RGBA) output. Official sampling is 8 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, 8 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 | 5.625s | 4.778s | 1.18× |
| Cold inference | 3.429s | 1.449s | 2.37× |
| Steady inference | 3.080s | 1.269s | 2.43× |
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 | 5.500s | 4.796s | 1.15× |
| Cold inference | 4.531s | 2.256s | 2.01× |
| Steady inference | 4.062s | 2.014s | 2.02× |
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 ↑ | 20.14 | 19.88 | 24.84 | 26.99 | 15.21 |
| Text-to-image | SSIM ↑ | 0.843 | 0.853 | 0.937 | 0.961 | 0.604 |
| Text-to-image | LPIPS ↓ | 0.185 | 0.164 | 0.320 | 0.047 | 0.531 |
| Image editing | PSNR ↑ | 19.17 | 19.75 | 23.47 | 24.32 | 13.62 |
| Image editing | SSIM ↑ | 0.719 | 0.717 | 0.879 | 0.928 | 0.397 |
| Image editing | LPIPS ↓ | 0.156 | 0.124 | 0.283 | 0.055 | 0.392 |
Use in ComfyUI
Install the Vitoom Nunchaku plugin. ComfyUI 0.44 or newer is required, started with --enable-manager. Installing the plugin also installs a matching nunchaku wheel.
- Open ComfyUI and click Manage Extensions in the top-right corner.
- Search for
Vitoom Nunchakuand click Install. - Restart ComfyUI.
Place this repository's transformer in ComfyUI/models/diffusion_models: svdq-fp4_r32-Qwen-Image-2.1-Turbo-Nunchaku.safetensors for RTX 50 series, svdq-int4_r32-Qwen-Image-2.1-Turbo-Nunchaku.safetensors for RTX 30 and 40 series. The text encoder is the same as the non-Turbo model: svdq-int4-Qwen3-VL-8B-Instruct-Nunchaku.safetensors, in ~/weights or ComfyUI/models/text_encoders. Place the VAE qwen_image_2.1_vae_bf16.safetensors in ComfyUI/models/vae.
The plugin does not ship a separate Turbo workflow. Open example_workflows/qwen_image21.json, select the Turbo checkpoint above on Vitoom Nunchaku DiT Loader, and set KSampler to 8 steps. Connect Vitoom Nunchaku DiT Loader to KSampler, and Vitoom Nunchaku Text Encoder to TextEncodeQwenImage21. For editing, connect the reference image to the encode node's image_1 and write <image1> in the prompt.
Use in Python
Installation
Nunchaku must provide NunchakuQwenImage21Transformer2DModel, and Diffusers must provide QwenImage21Pipeline (after 2026-09-18, including PR #14804; Turbo's sampling schedule also needs PR #14950). 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-Turbo-Nunchaku"
NAME = "Qwen-Image-2.1-Turbo-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=8,
true_cfg_scale=1.0,
use_kv_cache=True,
generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
image.save("qwen21-turbo-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=8,
true_cfg_scale=1.0,
use_kv_cache=True,
generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
edited.save("qwen21-turbo-edit-w4a4.png")
Acknowledgments
- nunchaku: a high-performance inference engine.
- deepcompressor: a model compression toolbox.
- fluxsd.com: image, video, poster, comic, and picture-book generation.
- createpptx.com: high-quality editable PowerPoint generation.
License
These quantized weights are derived from Qwen/Qwen-Image-2.1-Turbo 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.
- Downloads last month
- 865