README.md · Pruna-Qwen-Image-2.1

1

---

2

license: other

3

license_name: qwen-research

4

base_model: Qwen/Qwen-Image-2.1

5

library_name: diffusers

6

pipeline_tag: text-to-image

7

tags:

8

- qwen

9

- image-generation

10

- image-editing

11

- rgba

12

- lora

13

- distilled

14

- few-step

15

---

16

17

18

<p align="center">

19

<img src="assets/Pruna-Qwen-Image-2.1.png" width="800">

20

</p>

21

22

23

<p align="center">

24

<a href="https://github.com/PrunaAI/pruna">

25

    <img src="https://img.shields.io/badge/GitHub-PrunaAI-9334E9?style=plastic&logo=github&logoColor=white" alt="GitHub">

26

</a>

27

 

28

<a href="https://twitter.com/PrunaAI">

29

    <img src="https://img.shields.io/badge/Twitter%2FX-@PrunaAI-9334E9?style=plastic&logo=x&logoColor=white" alt="Twitter/X">

30

</a>

31

 

32

<a href="https://www.linkedin.com/company/pruna-ai">

33

    <img src="https://img.shields.io/badge/LinkedIn-PrunaAI-9334E9?style=plastic&logo=linkedin&logoColor=white" alt="LinkedIn">

34

</a>

35

 

36

<a href="https://discord.com/invite/JFQmtFKCjd">

37

    <img src="https://img.shields.io/badge/Discord-Join%20us-9334E9?style=plastic&logo=discord&logoColor=white" alt="Discord">

38

</a>

39

 

40

<a href="https://dashboard.pruna.ai/login?utm_source=huggingface&utm_medium=org_card&utm_campaign=hf_traffic">

41

    <img src="https://img.shields.io/badge/Performance%20Models-Try%20them%20now-9334E9?style=plastic&logo=data:image/svg%2bxml;base64,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" alt="Performance Models">

42

</a>

43

</p>

44

<div align="center">

45

46

<h1 style="color: #9334E9;">⚡ Pruna-Qwen-Image-2.1</h1>

47

48

<h2>Few-step LoRA adapters for Qwen-Image-2.1</h2>

49

50

<h3>

51

<span style="color: #9334E9;">5 or 8 steps</span>

52

 · 

53

<span style="color: #9334E9;">Up to 6.3x faster</span>

54

 · 

55

<span style="color: #9334E9;">No CFG</span>

56

 · 

57

Text-to-image and image editing

58

</h3>

59

60

</div>

61

62

**Pruna-Qwen-Image-2.1 is a set of LoRA adapters that let

63

[`Qwen/Qwen-Image-2.1`](https://huggingface.co/Qwen/Qwen-Image-2.1) generate

64

and edit images in only 5 or 8 steps. The adapters load on top of the base

65

model, so the pipeline, text encoder, and VAE stay unchanged. Training is based

66

on DMD. Improved using Qwen.**

67

68

> **v0.1: first release, work in progress.** Pruna-Qwen-Image-2.1 does not yet match the

69

> visual quality of the base model. We are still improving the distillation and

70

> will update this repository as new versions become available.

71

72

## Variants

73

74

Both adapters are v0.1. Pick one based on whether you need **quality** or **speed**.

75

76

| File | Steps | Trade-off |

77

|---|:---:|---|

78

| `p_qwen_image_2.1_8step_v0.1.safetensors` | 8 | **Higher quality.** Recommended as the default. |

79

| `p_qwen_image_2.1_5step_v0.1.safetensors` | 5 | **Higher speed**, with noticeably lower visual quality. |

80

81

Each adapter is trained for its own sigma schedule. Load only one at a time.

82

83

## Prompts, editing, and resolution

84

85

The adapters were trained at **1K resolution only**, with a mix of simple and

86

upsampled prompts, text-to-image generation, and single- and multi-image editing

87

with **up to 3 reference images**. Prompt upsampling is optional; detailed prompts

88

usually give better results.

89

90

Start at **1024 × 1024** and use at most **3 reference images** for editing.

91

Higher resolutions (including 2K) and more reference images may work, but are

92

outside the training coverage, and quality may vary. The 2K timings below show

93

inference speed, not a guarantee of quality at that resolution.

94

95

## Benchmarking

96

97

Use a table; Hugging Face can otherwise render the `<img>` elements as separate blocks.

98

99

<table>

100

<tr>

101

<td width="50%">

102

<img src="assets/Pruna-Qwen-Image-2.1-t2i-1024.png" width="100%">

103

</td>

104

<td width="50%">

105

<img src="assets/Pruna-Qwen-Image-2.1-t2i-2048.png" width="100%">

106

</td>

107

</tr>

108

<tr>

109

<td width="50%">

110

<img src="assets/Pruna-Qwen-Image-2.1-i2i-1024.png" width="100%">

111

</td>

112

<td width="50%">

113

<img src="assets/Pruna-Qwen-Image-2.1-i2i-2048.png" width="100%">

114

</td>

115

</tr>

116

</table>

117

118

Text-to-image with the official Qwen model-card example prompt, BF16, batch size 1,

119

on one NVIDIA H100 80GB. Median of 3 requests after one warmup per configuration.

120

Includes prompt encoding, denoising, and decoding; excludes PNG saving, model

121

loading, and warmup. Base: 40 steps with KV cache on; Pruna-Qwen-Image-2.1: 5 or 8 steps with

122

KV cache off, as configured for this benchmark. The examples below enable KV caching for

123

Pruna-Qwen-Image-2.1; the chart has not been remeasured with that setting. LoRAs were unmerged;

124

no CFG, compilation, or CPU offload. These timings do not imply equal image quality.

125

126

## Quickstart

127

128

Runs text-to-image and image editing with the same pipeline. Needs a **CUDA GPU**.

129

130

```bash

131

pip install 'torch>=2.4.0' 'transformers>=5.17' accelerate peft pillow

132

pip install git+https://github.com/huggingface/diffusers@6256aa7666cedd47443adc8f82da9a10e110b09c

133

```

134

135

```python

136

import torch

137

from PIL import Image

138

from diffusers import FlowMatchEulerDiscreteScheduler, QwenImage21Pipeline

139

140

STEPS = 8 # 8 for higher quality, 5 for higher speed

141

142

# The terminal sigma 0 is appended by the scheduler.

143

SIGMAS = {

144

5: [1.0, 0.94, 6 / 7, 2 / 3, 0.4],

145

8: [1.0, 14 / 15, 6 / 7, 10 / 13, 2 / 3, 6 / 11, 0.4, 2 / 9],

146

}[STEPS]

147

148

pipe = QwenImage21Pipeline.from_pretrained(

149

"Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16

150

).to("cuda")

151

pipe.load_lora_weights(

152

"PrunaAI/Pruna-Qwen-Image-2.1",

153

weight_name=f"p_qwen_image_2.1_{STEPS}step_v0.1.safetensors",

154

)

155

156

# Use the sigmas exactly as given: no extra shifting.

157

pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(

158

pipe.scheduler.config,

159

use_dynamic_shifting=False,

160

shift=1.0,

161

shift_terminal=None,

162

)

163

```

164

165

## For Text-to-Image

166

167

Longer, more descriptive prompts give better results.

168

169

```python

170

image = pipe(

171

prompt=(

172

'A glowing neon shop sign that reads "QWEN IMAGE 2.1", mounted on a brick wall '

173

"in a narrow city alley at night. Heavy rain, wet pavement reflecting pink and "

174

"blue light, shallow depth of field, cinematic photograph."

175

),

176

width=1024,

177

height=1024,

178

generator=torch.Generator("cuda").manual_seed(42),

179

num_inference_steps=STEPS,

180

sigmas=SIGMAS,

181

true_cfg_scale=1.0,

182

use_kv_cache=True,

183

).images[0]

184

```

185

186

### For Image Editing

187

188

```python

189

image = pipe(

190

prompt="Change the background to a sunset beach",

191

image=Image.open("input.png").convert("RGB"),

192

generator=torch.Generator("cuda").manual_seed(42),

193

num_inference_steps=STEPS,

194

sigmas=SIGMAS,

195

true_cfg_scale=1.0,

196

use_kv_cache=True,

197

).images[0]

198

```

199

200

### Recommended settings

201

202

- **Use the sigma schedule that matches your adapter**, and keep the scheduler

203

at `shift=1.0` with dynamic shifting off so the sigmas are not shifted twice.

204

- 8-step: `1 → 14/15 → 6/7 → 10/13 → 2/3 → 6/11 → 0.4 → 2/9 → 0`

205

(shift 2, computed as `σ = 2t / (1 + t)` on evenly spaced `t`)

206

- 5-step: `1 → 0.94 → 6/7 → 2/3 → 0.4 → 0`

207

- **No CFG.** Keep `true_cfg_scale=1.0` and do not pass a negative prompt.

208

- **Keep the LoRA strength at 1.0.**

209

- **Write detailed prompts for text-to-image.** Longer prompts that describe

210

the subject, setting, lighting, and style work noticeably better than short

211

ones.

212

213

## Limitations

214

215

- This is a first version. Quality is below that of the base model.

216

- The 5-step adapter is faster than the 8-step one, but its images are

217

visibly worse.

218

- Short or vague text-to-image prompts give weaker results.

219

- The adapters are only intended for the settings above. Other step counts,

220

schedules, or CFG values are not supported.

221

222

## What this is not

223

224

- Not a standalone model. It needs the `Qwen/Qwen-Image-2.1` base weights.

225

- Not a replacement for the base model when you need its full quality.

226

- Not a finished release. Expect the weights to change in future versions.

227

228

## License

229

230

Pruna-Qwen-Image-2.1 is a derivative of Qwen-Image-2.1 and is distributed

231

under the **Qwen RESEARCH LICENSE AGREEMENT**. Review its use restrictions

232

before using or redistributing the adapters.

233

234

235

236

237

## What's next?

238

239

- **Use Pruna-Qwen-Image-2.1 to generate and edit images with Qwen-Image-2.1 in a few steps.**

240

- Compress your own models with [Pruna](https://github.com/PrunaAI/pruna) and give us a ⭐️ for more efficiency!

241

- Want to use our optimized image models right away?  Check [P-Image-Ideogram](https://www.pruna.ai/p-image-ideogram) and [P-Image-Edit](https://www.pruna.ai/p-image-edit).

242

243

<style>

244

.model-button {

245

display: inline-flex;

246

flex-direction: row;

247

justify-content: center;

248

align-items: center;

249

gap: 8px;

250

251

padding: 8px 20px;

252

border: none;

253

border-radius: 8px;

254

255

background: #9334e9;

256

color: #ffffff;

257

258

font-size: 14px;

259

font-weight: 400;

260

line-height: 1;

261

text-decoration: none;

262

white-space: nowrap;

263

cursor: pointer;

264

265

box-sizing: border-box;

266

overflow: visible;

267

opacity: 1;

268

}

269

</style>

270

271

<a href="https://dashboard.pruna.ai/login?utm_source=huggingface&utm_medium=model_card&utm_campaign=hf_traffic" class="model-button">Try our models</a>