---
language:
  - en
pipeline_tag: text-to-image
license: apache-2.0
tags:
  - pvd
  - image-generation
  - class-conditional
  - diffusion
  - lora
  - pytorch
---
<div align="center">
<h1>Two Halves are More than One: Phase-wise Velocity Distillation for Fast and High-Quality Image Generation</h1>

<a href="https://arxiv.org/abs/2610.08070"><img src="https://img.shields.io/badge/Paper-arXiv-red" alt="Paper"></a>
<a href="https://github.com/PolyU-VCLab/PVD"><img src="https://img.shields.io/badge/Code-Github-blue" alt="Code"></a>
<a href="https://huggingface.co/VCLab-PolyU/PVD"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Model-HuggingFace-yellow" alt="Model"></a>
<a href="https://huggingface.co/papers/2610.08070"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Paper-HuggingFace-yellow" alt="HF Paper"></a>
</div>

Model weights for PVD class-conditioned image generation (C2I) and text-to-image generation with SD3.5 Medium, FLUX.1-dev and Qwen-Image.

## Model Details

| Variant | Weight directory | Released components | Configuration |
| --- | --- | --- | --- |
| C2I | `weights/pvd_c2i/` | `model.safetensors`; VAE and latent statistics in `cache/` | ImageNet classes, 256×256 |
| SD3.5 Medium | `weights/pvd_sd35m/` | `part1.safetensors`, `part2.safetensors`, `model_config.json` | 12 layers per transformer |
| SD3.5 Medium (Unsplash) | `weights/pvd_sd35m_unsplash/` | `part1.safetensors`, `part2.safetensors`, `model_config.json` | Further trained on the Unsplash dataset |
| FLUX.1-dev | `weights/pvd_flux/` | `backbone.safetensors`, `part1/` and `part2/` LoRA adapters | 9 dual-stream + 19 single-stream layers; rank 64 |
| FLUX.1-dev (Unsplash) | `weights/pvd_flux_unsplash/` | `part1/` and `part2/` LoRA adapters; reuses `weights/pvd_flux/backbone.safetensors` | Further trained on the Unsplash dataset |
| Qwen-Image | `weights/pvd_qwenimage/` | `backbone.safetensors`, `part1/` and `part2/` LoRA adapters | 30 layers; rank 32 |
| Qwen-Image (Unsplash) | `weights/pvd_qwen_unsplash/` | `part1/` and `part2/` LoRA adapters; reuses `weights/pvd_qwenimage/backbone.safetensors` | Further trained on the Unsplash dataset |

Each adapter folder contains `adapter_model.safetensors` and `adapter_config.json`. The C2I decoder and latent statistics are `cache/vavae-imagenet256-f16d32-dinov2.pt` and `cache/latents_stats.pt`.

## Usage

Place this repository's `weights/` and `cache/` folders in the [PVD code repository](https://github.com/PolyU-VCLab/PVD) root. Keep the folder names as listed above. See the code repository's README for base model components and inference commands.