HaozheLiu-ST/T-GATE
A training-free method that accelerates diffusion models for text-to-image generation by decomposing and gating cross-attention mechanisms.
Not currently ranked — collecting fresh signals.
star history
The project provides a technique to accelerate diffusion model inference through temporal attention decomposition. T-GATE (Temporally Gating Attention) identifies that cross-attention becomes redundant in later timesteps and proposes a gating approach to skip expensive computations without retraining. The method works across multiple diffusion frameworks including original DDPM, Diffusers, and ControlNet, offering speedups while maintaining image quality.
Frequently asked
- What is HaozheLiu-ST/T-GATE?
- A training-free method that accelerates diffusion models for text-to-image generation by decomposing and gating cross-attention mechanisms.
- Is T-GATE open source?
- Yes — HaozheLiu-ST/T-GATE is open source, released under the MIT license.
- What language is T-GATE written in?
- HaozheLiu-ST/T-GATE is primarily written in Python.
- How popular is T-GATE?
- HaozheLiu-ST/T-GATE has 418 stars on GitHub.
- Where can I find T-GATE?
- HaozheLiu-ST/T-GATE is on GitHub at https://github.com/HaozheLiu-ST/T-GATE.