https://github.com/jcelerier/librediffusion

C++ / CUDA implementation of StreamDiffusion
https://github.com/jcelerier/librediffusion

genai stablediffusion streamdiffusion

Last synced: 10 days ago
JSON representation

C++ / CUDA implementation of StreamDiffusion

• Host: GitHub
• URL: https://github.com/jcelerier/librediffusion
• Owner: jcelerier
• License: agpl-3.0
• Created: 2025-12-24T15:14:17.000Z (6 months ago)
• Default Branch: main
• Last Pushed: 2026-05-22T13:14:51.000Z (30 days ago)
• Last Synced: 2026-06-11T21:35:45.906Z (10 days ago)
• Topics: genai, stablediffusion, streamdiffusion
• Language: C++
• Homepage:
• Size: 1.51 MB
• Stars: 32
• Watchers: 0
• Forks: 3
• Open Issues: 5
• Metadata Files:
  • Readme: README.md
  • Funding: .github/FUNDING.yml
  • License: LICENSE

Awesome Lists containing this project

README

          
# librediffusion

A C++ / CUDA / TensorRT implementation of StreamDiffusion

Implemented in [ossia score](https://ossia.io)

# Benchmarks

On a RTX 5090 at 1 step: 

SDXL Turbo 1024x1024: stable 26 fps

![sdxl](https://github.com/user-attachments/assets/340bf804-6822-46c6-87d8-bf784722f3b5)

SD Turbo 512x512: stable 96 fps

![sdturbo](https://github.com/user-attachments/assets/10ebc9f7-d5b7-487b-a301-b1b7e2370d55)

SDXS: above 600 fps

![sdxs](https://github.com/user-attachments/assets/9f735f86-d162-4c4c-b781-163cadf166a5)

Models need to be converted to TensorRT through the Python script [train-lora.py] beforehand:

```bash

$ uv run train-lora.py --model stabilityai/sd-turbo --min-batch 1 --max-batch 1 --opt-batch 1 --min-resolution 512 --max-resolution 1024 --output ./engines-sd-turbo

```