File size: 9,278 Bytes
e410c29
 
 
 
 
 
 
 
 
 
 
 
 
 
a1cde68
e410c29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
---
pipeline_tag: any-to-any
library_name: transformers
tags:
- text-to-image
- image-editing
- image-understanding
- vision-language
- multimodal
- autoregressive
- unified-model
license: mit
---

## 🌌 UniPic2-Metaquery-GRPO-Flash

<div align="center">   
  <img src="skywork-logo.png" alt="Skywork Logo" width="500"> 
</div>

<p align="center">
  <a href="https://github.com/SkyworkAI/UniPic">
    <img src="https://img.shields.io/badge/GitHub-UniPic-blue?logo=github" alt="GitHub Repo">
  </a>
  <a href="https://github.com/SkyworkAI/UniPic/stargazers">
    <img src="https://img.shields.io/github/stars/SkyworkAI/UniPic?style=social" alt="GitHub Stars">
  </a>
  <a href="https://github.com/SkyworkAI/UniPic/network/members">
    <img src="https://img.shields.io/github/forks/SkyworkAI/UniPic?style=social" alt="GitHub Forks">
  </a>
</p>


## 📖 Introduction

**UniPic2-Metaquery-GRPO-Flash** is a quantized variant of UniPic2-MetaQuery-GRPO, offering end-to-end image understanding, text-to-image (T2I) generation, and image editing. Optimized for efficiency, it runs smoothly on NVIDIA RTX 40-series GPUs with under 16 GB VRAM — without any performance degradation.
<div align="center"> <img src="teaser.png" alt="Model Teaser" width="720"> </div>
<div align="center"> <img src="understanding.png" alt="Model Teaser" width="720"> </div>


## 📊 Benchmarks

<div align="center"> <img src="eval.png" alt="Model Eval" width="720"> </div>


## 🧠 Usage

### 1. Clone the Repository

```bash
git clone https://github.com/SkyworkAI/UniPic
cd UniPic-2
```

### 2. Set Up the Environment
```bash
conda create -n unipic python=3.10
conda activate unipic
pip install -r requirements.txt
```


### 3.Text-to-Image Generation
```bash
import torch
from PIL import Image
from unipicv2.pipeline_stable_diffusion_3_kontext import StableDiffusion3KontextPipeline
from unipicv2.transformer_sd3_kontext import SD3Transformer2DKontextModel
from unipicv2.stable_diffusion_3_conditioner import StableDiffusion3Conditioner
from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLProcessor
from diffusers import FlowMatchEulerDiscreteScheduler, AutoencoderKL,BitsAndBytesConfig

# Load model components
pretrained_model_name_or_path = "/path/to/UniPic2-Metaquery-Flash/UniPic2-Metaquery"
vlm_path   = "/path/to/UniPic2-Metaquery-Flash/Qwen2.5-VL-7B-Instruct-AWQ"

quant = "int4"  # {"int4", "fp16"}

bnb4 = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,  # 与 LMM/Cond 对齐
)

if quant == "int4":
    transformer = SD3Transformer2DKontextModel.from_pretrained(
        pretrained_model_name_or_path, subfolder="transformer",
        quantization_config=bnb4, device_map="auto", low_cpu_mem_usage=True
    )
elif quant == "fp16":
    transformer = SD3Transformer2DKontextModel.from_pretrained(
        pretrained_model_name_or_path, subfolder="transformer",
        torch_dtype=torch.float16, device_map="auto", low_cpu_mem_usage=True
    )
else:
    raise ValueError(f"Unsupported quant: {quant}")


vae = AutoencoderKL.from_pretrained(
    pretrained_model_name_or_path, subfolder="vae",
    torch_dtype=torch.float16, device_map="auto", low_cpu_mem_usage=True).cuda()

# Load Qwen2.5-VL model
lmm = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    vlm_path,
    torch_dtype=torch.bfloat16,device_map="auto",
    attn_implementation="flash_attention_2")

processor = Qwen2_5_VLProcessor.from_pretrained(vlm_path)
processor.chat_template = processor.chat_template.replace(
    "{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}",
    "")

# 加上cuda
conditioner = StableDiffusion3Conditioner.from_pretrained(
    pretrained_model_name_or_path, subfolder="conditioner", torch_dtype=torch.float16).cuda()

scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(pretrained_model_name_or_path, subfolder="scheduler")

# Create pipeline (note: text encoders set to None)
pipeline = StableDiffusion3KontextPipeline(
    transformer=transformer, vae=vae,
    text_encoder=None, tokenizer=None,
    text_encoder_2=None, tokenizer_2=None,
    text_encoder_3=None, tokenizer_3=None,
    scheduler=scheduler)

# Prepare prompts
prompt = 'a pig with wings and a top hat flying over a happy futuristic scifi city'
negative_prompt = ''

messages = [[{"role": "user", "content": [{"type": "text", "text": f'Generate an image: {txt}'}]}]
            for txt in [prompt, negative_prompt]]

texts = [processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True) for msg in messages]
inputs = processor(text=texts, images=None, videos=None, padding=True, return_tensors="pt").to("cuda")

# Process with Qwen2.5-VL
input_ids, attention_mask = inputs.input_ids, inputs.attention_mask
input_ids = torch.cat([input_ids, input_ids.new_zeros(2, conditioner.config.num_queries)], dim=1)
attention_mask = torch.cat([attention_mask, attention_mask.new_ones(2, conditioner.config.num_queries)], dim=1)
inputs_embeds = lmm.get_input_embeddings()(input_ids)
inputs_embeds[:, -conditioner.config.num_queries:] = conditioner.meta_queries[None].expand(2, -1, -1)

outputs = lmm.model(inputs_embeds=inputs_embeds, attention_mask=attention_mask, use_cache=False)
hidden_states = outputs.last_hidden_state[:, -conditioner.config.num_queries:]
prompt_embeds, pooled_prompt_embeds = conditioner(hidden_states)

# Generate image
image = pipeline(
    prompt_embeds=prompt_embeds[:1],
    pooled_prompt_embeds=pooled_prompt_embeds[:1],
    negative_prompt_embeds=prompt_embeds[1:],
    negative_pooled_prompt_embeds=pooled_prompt_embeds[1:],
    height=512, width=384,
    num_inference_steps=50,
    guidance_scale=3.5,
    generator=torch.Generator(device=transformer.device).manual_seed(42)
).images[0]

image.save("text2image.png")
print(f"Image saved to text2image.png (quant={quant})")
```


### 4.  Image Editing
```bash
# Load image for editing
image = Image.open("text2image.png")
image = fix_longer_edge(image, image_size=512)

prompt = "remove the pig's hat"
negative_prompt = "blurry, low quality, low resolution, distorted, deformed, broken content, missing parts, damaged details, artifacts, glitch, noise, pixelated, grainy, compression artifacts, bad composition, wrong proportion, incomplete editing, unfinished, unedited areas."

# Prepare messages with image input
messages = [[{"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": txt}]}]
            for txt in [prompt, negative_prompt]]

texts = [processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True) for msg in messages]

min_pixels = max_pixels = int(image.height * 28 / 32 * image.width * 28 / 32)
inputs = processor(
    text=texts, images=[image]*2,
    min_pixels=min_pixels, max_pixels=max_pixels,
    videos=None, padding=True, return_tensors="pt").cuda()

# Process with vision understanding
input_ids, attention_mask, pixel_values, image_grid_thw = \
    inputs.input_ids, inputs.attention_mask, inputs.pixel_values, inputs.image_grid_thw

input_ids = torch.cat([input_ids, input_ids.new_zeros(2, conditioner.config.num_queries)], dim=1)
attention_mask = torch.cat([attention_mask, attention_mask.new_ones(2, conditioner.config.num_queries)], dim=1)
inputs_embeds = lmm.get_input_embeddings()(input_ids)
inputs_embeds[:, -conditioner.config.num_queries:] = conditioner.meta_queries[None].expand(2, -1, -1)

image_embeds = lmm.visual(pixel_values, grid_thw=image_grid_thw)
image_token_id = processor.tokenizer.convert_tokens_to_ids('<|image_pad|>')
inputs_embeds[input_ids == image_token_id] = image_embeds

lmm.model.rope_deltas = None
outputs = lmm.model(inputs_embeds=inputs_embeds, attention_mask=attention_mask,
                    image_grid_thw=image_grid_thw, use_cache=False)

hidden_states = outputs.last_hidden_state[:, -conditioner.config.num_queries:]
prompt_embeds, pooled_prompt_embeds = conditioner(hidden_states)

# Generate edited image
edited_image = pipeline(
    image=image,
    prompt_embeds=prompt_embeds[:1],
    pooled_prompt_embeds=pooled_prompt_embeds[:1],
    negative_prompt_embeds=prompt_embeds[1:],
    negative_pooled_prompt_embeds=pooled_prompt_embeds[1:],
    height=image.height, width=image.width,
    num_inference_steps=50,
    guidance_scale=3.5,
    generator=torch.Generator(device=transformer.device).manual_seed(42)
).images[0]

edited_image.save("edited_image.png")
print(f"Image saved to edited_image.png (quant={quant})")

```


## 📄 License
This model is released under the MIT License.


## Citation
If you use Skywork-UniPic in your research, please cite:
```
@misc{wang2025skyworkunipicunifiedautoregressive,
      title={Skywork UniPic: Unified Autoregressive Modeling for Visual Understanding and Generation}, 
      author={Peiyu Wang and Yi Peng and Yimeng Gan and Liang Hu and Tianyidan Xie and Xiaokun Wang and Yichen Wei and Chuanxin Tang and Bo Zhu and Changshi Li and Hongyang Wei and Eric Li and Xuchen Song and Yang Liu and Yahui Zhou},
      year={2025},
      eprint={2508.03320},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.03320}, 
}
```