---
tags:
- text-to-image
- lora
- diffusers
- template:diffusion-lora
base_model: Qwen/Qwen-Image
instance_prompt: Qwen-Image, distillation
license: apache-2.0
pipeline_tag: text-to-image
library_name: diffusers
---

# Glance: Accelerating Diffusion Models with 1 Sample

This model is presented in the paper [Glance: Accelerating Diffusion Models with 1 Sample](https://huggingface.co/papers/2512.02899).

Project Page: https://zhuobaidong.github.io/Glance/

Code: https://github.com/CSU-JPG/Glance

## About
Glance introduces a novel approach to accelerate diffusion models by intelligently speeding up denoising phases. Instead of costly retraining, Glance equips base models with lightweight Slow-LoRA and Fast-LoRA adapters. This method achieves up to 5x acceleration over base models while maintaining comparable visual quality and strong generalization on unseen prompts. Notably, the LoRA experts are trained with only 1 sample within an hour.

---

# 🧪 Usage

## 🎨 Inference

We provide solid 4-GPU inference code for easy multi-card sampling. You can experience our Glance model by running:

```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 python infer_Glance_qwen.py
```

### Glance (Qwen-Image)
```python
import torch
from pipeline.qwen import GlanceQwenSlowPipeline, GlanceQwenFastPipeline
from utils.distribute_free import distribute, free_pipe

repo = "CSU-JPG/Glance"
slow_pipe = GlanceQwenSlowPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.float32)
slow_pipe.load_lora_weights(repo, weight_name="glance_qwen_slow.safetensors")
distribute(slow_pipe)

prompt = "Please create a photograph capturing a young woman showcasing a dynamic presence as she bicycles alongside a river during a hot summer day. Her long hair streams behind her as she pedals, dressed in snug tights and a vibrant yellow tank top, complemented by New Balance running shoes that highlight her lean, athletic build. She sports a small backpack and sunglasses resting confidently atop her head."
latents = slow_pipe(
    prompt=prompt,
    negative_prompt=" ",
    width=1024,
    height=1024,
    num_inference_steps=5,
    true_cfg_scale=5,
    generator=torch.Generator(device="cuda").manual_seed(0), 
    output_type="latent"
).images[0].unsqueeze(0).detach().cpu()
free_pipe(slow_pipe)

fast_pipe = GlanceQwenFastPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.float32)
fast_pipe.load_lora_weights(repo, weight_name="glance_qwen_fast.safetensors")
distribute(fast_pipe)

image = fast_pipe(
    prompt=prompt,
    negative_prompt=" ", 
    width=1024,
    height=1024,
    num_inference_steps=5, 
    true_cfg_scale=5,
    generator=torch.Generator(device="cuda").manual_seed(0), 
    latents=latents.to("cuda", dtype=torch.float32)
).images[0]
image.save("output.png")
```

### 🖼️ Sample Output

![Sample Output](./assets/qwen.png)

## 🚀 Training

### Glance_Qwen Training

To start training with your configuration file, simply run:

```bash
accelerate launch train_Glance_qwen.py --config ./train_configs/Glance_qwen.yaml
```

> Note: All the training code is primarily based on [flymyai-lora-trainer](https://github.com/FlyMyAI/flymyai-lora-trainer).

Ensure that `Glance_qwen.yaml` is properly configured with your dataset paths, model settings, output directory, and other hyperparameters. You can also explicitly specify whether to train the **Slow-LoRA** or **Fast-LoRA** variant directly within the configuration file.

If you want to train on a **single GPU** (requires **less than 24 GB** of VRAM), run:

```bash
python train_Glance_qwen.py --config ./train_configs/Glance_qwen.yaml
```


### Glance_FLUX Training

To launch training for the FLUX variant, run:

```bash
accelerate launch train_Glance_flux.py --config ./train_configs/Glance_flux.yaml
```

## Citation
```
@misc{dong2025glanceacceleratingdiffusionmodels,
      title={Glance: Accelerating Diffusion Models with 1 Sample}, 
      author={Zhuobai Dong and Rui Zhao and Songjie Wu and Junchao Yi and Linjie Li and Zhengyuan Yang and Lijuan Wang and Alex Jinpeng Wang},
      year={2025},
      eprint={2512.02899},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.02899}, 
}
```