Skywork / UniPic2-SD3.5M-Kontext-GRPO-2B

🌌 UniPic2-SD3.5M-Kontext-GRPO-2B

📖 Introduction

UniPic2-SD3.5M-Kontext-GRPO-2B is a post-trained grpo version T2I model built on the UniPic2-SD3.5M-Kontext-2B, with enhanced text rendering. It excels at text-to-image generation and image editing, delivering high quality at fast speeds, and runs smoothly on a single 16 GB consumer GPU.

📊 Benchmarks

🧠 Usage

1. Clone the Repository

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

2. Set Up the Environment

conda create -n unipic python=3.10
conda activate unipic
pip install -r requirements.txt

3.Text-to-Image Generation

import torch
from PIL import Image
from unipicv2.pipeline_stable_diffusion_3_kontext import StableDiffusion3KontextPipeline
from unipicv2.transformer_sd3_kontext import SD3Transformer2DKontextModel
from diffusers import FlowMatchEulerDiscreteScheduler, AutoencoderKL, BitsAndBytesConfig
from transformers import CLIPTextModelWithProjection, CLIPTokenizer, T5EncoderModel, T5TokenizerFast

# Load model components
pretrained_model_name_or_path = "/mnt/datasets_vlm/chris/hf_ckpt/Unipic2-t2i"
# int4 is recommended for inference:lower VRAM with no quality loss  {"int4", "fp16"} 
quant = "int4"  

# BitsAndBytes config
bnb4 = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,         
    bnb_4bit_quant_type="nf4",              
    bnb_4bit_compute_dtype=torch.bfloat16, 
)
bnb8 = BitsAndBytesConfig(load_in_8bit=True)

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
    ).cuda()
    text_qconf = bnb8
    vae_dtype = torch.float16
else:  # fp16
    transformer = SD3Transformer2DKontextModel.from_pretrained(
        pretrained_model_name_or_path, subfolder="transformer",
        torch_dtype=torch.float16, device_map="auto", low_cpu_mem_usage=True
    ).cuda()
    text_qconf = None
    vae_dtype = torch.float16

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

# Load text encoders
text_encoder = CLIPTextModelWithProjection.from_pretrained(
    pretrained_model_name_or_path, subfolder="text_encoder",
    quantization_config=text_qconf, torch_dtype=None, device_map="auto", low_cpu_mem_usage=True
)
tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder="tokenizer")

text_encoder_2 = CLIPTextModelWithProjection.from_pretrained(
    pretrained_model_name_or_path, subfolder="text_encoder_2",
    quantization_config=text_qconf, torch_dtype=None, device_map="auto", low_cpu_mem_usage=True
)
tokenizer_2 = CLIPTokenizer.from_pretrained(pretrained_model_name_or_path, subfolder="tokenizer_2")

text_encoder_3 = T5EncoderModel.from_pretrained(
    pretrained_model_name_or_path, subfolder="text_encoder_3",
    quantization_config=text_qconf, torch_dtype=None, device_map="auto", low_cpu_mem_usage=True
)
tokenizer_3 = T5TokenizerFast.from_pretrained(pretrained_model_name_or_path, subfolder="tokenizer_3")

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

# Create pipeline
pipeline = StableDiffusion3KontextPipeline(
    transformer=transformer, vae=vae,
    text_encoder=text_encoder, tokenizer=tokenizer,
    text_encoder_2=text_encoder_2, tokenizer_2=tokenizer_2,
    text_encoder_3=text_encoder_3, tokenizer_3=tokenizer_3,
    scheduler=scheduler)

# Generate image
image = pipeline(
    prompt='a pig with wings and a top hat flying over a happy futuristic scifi city',
    negative_prompt='',
    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

# Load and preprocess image
def fix_longer_edge(x, image_size, factor=32):
    w, h = x.size
    if w >= h:
        target_w = image_size
        target_h = h * (target_w / w)
        target_h = round(target_h / factor) * factor
    else:
        target_h = image_size
        target_w = w * (target_h / h)
        target_w = round(target_w / factor) * factor
    x = x.resize(size=(target_w, target_h))
    return x

image = Image.open("text2image.png")
image = fix_longer_edge(image, image_size=512)

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."

# Edit image
edited_image = pipeline(
    image=image,
    prompt="remove the pig's hat",
    negative_prompt=negative_prompt,
    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_img.png")
print(f"Edited Image saved to edited_img.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}, 
}