| 
| 
| 
| 
|
| 
| 
| 
| 
| 
|
| 
| 
| 
| 
| 
|
| 
| 
| 
| 
| 
|
## Interactive Demo
## Model variants
| Steps | BF16 | FP8 | INT8 |
|---|---|---|---|
| 2 | [2step-bf16](./2step-bf16/diffusion_pytorch_model.safetensors) | [2step-fp8](./2step-fp8/) | [2step-int8](./2step-int8/) |
| 4 | [4step-bf16](./4step-bf16/diffusion_pytorch_model.safetensors) | [4step-fp8](./4step-fp8/) | [4step-int8](./4step-int8/) |
| 8 | [8step-bf16](./8step-bf16/diffusion_pytorch_model.safetensors) | [8step-fp8](./8step-fp8/) | [8step-int8](./8step-int8/) |
The INT8 variants use optimized W8A8 quantization and the included CUDA loader. Matrix weights use INT8 with FP32 per-channel scales; time embeddings, modulation, normalization and input/output projections retain BF16.
## ComfyUI
Choose an integration for the checkpoints in this repository:
| Integration | Precisions | Steps |
|---|---|---|
| Official ComfyUI workflows | BF16 | 2 / 4 / 8 |
| [Wuli ComfyUI node](https://github.com/wuli-art/Qwen-Image-2.1-Turbo-ComfyUI) | BF16 / FP8 / INT8 | 2 / 4 / 8 |
### Official ComfyUI — BF16 only
Download a single-file BF16 checkpoint from the table above and place it in `ComfyUI/models/diffusion_models`. Select it in `UNETLoader` in the official [text-to-image workflow](https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/templates/image_qwen_image_2_1_turbo_t2i.json) or [image editing workflow](https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/templates/image_qwen_image_2_1_turbo_image_edit.json). Keep the official text encoder and VAE.
Use **CFG 1.0** and set `ManualSigmas` to the selected schedule, including the final zero:
**2 steps**
```text
1.0, 0.8, 0.0
```
**4 steps**
```text
1.0, 0.9177013039588928, 0.787998616695404, 0.5533687472343445, 0.0
```
**8 steps**
```text
1.0, 0.962988555431366, 0.9177013039588928, 0.8610130548477173, 0.787998616695404, 0.6904187798500061, 0.5533687472343445, 0.3468291163444519, 0.0
```
The FP8 and INT8 checkpoints in this repository require the custom node below.
### Wuli ComfyUI node — BF16 / FP8 / INT8
Our [ComfyUI node](https://github.com/wuli-art/Qwen-Image-2.1-Turbo-ComfyUI) supports all nine variants for text-to-image generation and reference-image editing, with automatic model downloads and CPU offload.
**Install** using the Python environment that runs ComfyUI:
```bash
cd ComfyUI/custom_nodes
git clone https://github.com/wuli-art/Qwen-Image-2.1-Turbo-ComfyUI.git
cd Qwen-Image-2.1-Turbo-ComfyUI
python -m pip install -r requirements.txt
```
For Windows Portable, install dependencies from the portable installation's root:
```powershell
.\python_embeded\python.exe -m pip install -r .\ComfyUI\custom_nodes\Qwen-Image-2.1-Turbo-ComfyUI\requirements.txt
```
Restart ComfyUI, then load a workflow:
- [Text-to-image workflow](https://raw.githubusercontent.com/wuli-art/Qwen-Image-2.1-Turbo-ComfyUI/main/example_workflows/text_to_image.json)
- [Reference-image editing workflow](https://raw.githubusercontent.com/wuli-art/Qwen-Image-2.1-Turbo-ComfyUI/main/example_workflows/image_reference.json)
Leave `model_path` and `base_model_path` empty, select `variant` (for example, `8step-fp8`, `4step-int8` or `2step-bf16`), enter a prompt and run. For editing, also select a reference image in **Load Image**. The node applies the selected checkpoint's saved sampling schedule and CFG automatically.
The first run downloads the selected Turbo weights and the text encoder, processor and VAE from `Qwen/Qwen-Image-2.1`. Later runs reuse the cache, and base components are shared across variants. Use the default `cpu_offload` mode to reduce VRAM usage, or `gpu` if you have enough VRAM.
See the [node README](https://github.com/wuli-art/Qwen-Image-2.1-Turbo-ComfyUI#readme) for local paths, cache settings and other options.
## Requirements
Python 3.11+ with CUDA-enabled PyTorch. Download the helpers and install the pinned Diffusers revision:
```bash
python -m pip install huggingface-hub
python -c 'from huggingface_hub import snapshot_download; snapshot_download("Wuli-art/Qwen-Image-2.1-Turbo", allow_patterns=["*.py", "requirements-*.txt"], local_dir="qwen21-turbo")'
cd qwen21-turbo
python -m pip install -r requirements-diffusers.txt
```
The examples download the selected DiT and reuse the text encoder, VAE and processor from `Qwen/Qwen-Image-2.1`.
## Code Example
### For BF16
```python
import sys
from pathlib import Path
import torch
from diffusers import FlowMatchEulerDiscreteScheduler, QwenImage21Pipeline, QwenImage21Transformer2DModel
from huggingface_hub import snapshot_download
variant = "8step-bf16"
snapshot = Path(snapshot_download(
"Wuli-art/Qwen-Image-2.1-Turbo", allow_patterns=[f"{variant}/*", "*.py"],
))
sys.path.insert(0, str(snapshot))
from model_utils import read_config, resolve_base
dit_path = snapshot / variant
sampling = read_config(dit_path)[0]
transformer = QwenImage21Transformer2DModel.from_pretrained(str(dit_path), torch_dtype=torch.bfloat16)
pipe = QwenImage21Pipeline.from_pretrained(
str(resolve_base()), transformer=transformer, torch_dtype=torch.bfloat16,
)
pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(str(dit_path))
pipe.to("cuda")
image = pipe(
prompt="A red ceramic teapot on a wooden table, soft daylight",
width=1024, height=1024, true_cfg_scale=1.0,
num_inference_steps=sampling["num_inference_steps"], sigmas=sampling["sample_sigmas"],
generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
image.save("teapot-bf16.png")
```
### For FP8
```python
import sys
from pathlib import Path
import torch
from diffusers import FlowMatchEulerDiscreteScheduler, QwenImage21Pipeline
from huggingface_hub import snapshot_download
variant = "8step-fp8"
snapshot = Path(snapshot_download(
"Wuli-art/Qwen-Image-2.1-Turbo", allow_patterns=[f"{variant}/*", "*.py"],
))
sys.path.insert(0, str(snapshot))
from load_fp8 import load_fp8_transformer
from model_utils import read_config, resolve_base
dit_path = snapshot / variant
sampling = read_config(dit_path)[0]
transformer = load_fp8_transformer(dit_path, device="cpu")
pipe = QwenImage21Pipeline.from_pretrained(
str(resolve_base()), transformer=transformer, torch_dtype=torch.bfloat16,
)
pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(str(dit_path))
pipe.to("cuda")
image = pipe(
prompt="A red ceramic teapot on a wooden table, soft daylight",
width=1024, height=1024, true_cfg_scale=1.0,
num_inference_steps=sampling["num_inference_steps"], sigmas=sampling["sample_sigmas"],
generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
image.save("teapot-fp8.png")
```
### For INT8
Download the helpers with the command in **Requirements**, then run:
```bash
python infer_int8.py --variant 8step-int8 \
--prompt "A red ceramic teapot on a wooden table, soft daylight" \
--width 1024 --height 1024 --seed 42 --output teapot-int8.png
```
Select `2step-int8` or `4step-int8` for the other schedules. Add `--reference input.png` for editing; repeat the option for multiple reference images. To download from ModelScope, add `--source modelscope`.
For **vLLM-Omni**, use a separate environment and run from the downloaded helper directory:
```bash
python -m pip install -r requirements-omni.txt
python infer_omni.py --variant 8step-fp8 \
--prompt "A red ceramic teapot on a wooden table, soft daylight" \
--edit-prompt "Make the teapot blue. Preserve its shape and the wooden table." \
--seed 42 --output omni-teapot.png
```
## Acknowledgement
Thanks to the Qwen team for Qwen-Image-2.1. Built with Qwen. Distributed under the [Qwen Research License](https://huggingface.co/Wuli-art/Qwen-Image-2.1-Turbo/blob/main/LICENSE).