Feature Summary
SPEED - Spectral Progressive Diffusion for Efficient Image and Video Generation. Increases speed of high resolution generation.
Detailed Description
Website: https://howardxiao.ca/speed/
Paper: https://arxiv.org/abs/2605.18736
(If I understood paper correctly, here Z-Image (1.56×) and Z-Image (4.99×) just mean running at reduced step count to match speedup from SPEED)
From abstract:
"We propose Spectral Progressive Diffusion, a general framework that progressively grows resolution along the denoising trajectory of pretrained diffusion models. To this end, we develop a spectral noise expansion mechanism and derive an optimal resolution schedule from the model’s power spectrum. Our framework supports training-free acceleration and a fine-tuning recipe that further improves efficiency and quality, enabling significant speedups on state-of-the-art pretrained image and video generation models while preserving visual quality."
With this we can get between 1.4 and 2x training-free speedup for both image and video generation, and up to 5x when using special finetune/lora, while mostly preserving quality, although image composition changes quite a bit.
There's no official implementation/code yet, all I've found is comfyui node: https://github.com/ruwwww/ComfyUI-SPEED
Alternatives you considered
No response
Additional context
No response
Feature Summary
SPEED - Spectral Progressive Diffusion for Efficient Image and Video Generation. Increases speed of high resolution generation.
Detailed Description
Website: https://howardxiao.ca/speed/
Paper: https://arxiv.org/abs/2605.18736
From abstract:
"We propose Spectral Progressive Diffusion, a general framework that progressively grows resolution along the denoising trajectory of pretrained diffusion models. To this end, we develop a spectral noise expansion mechanism and derive an optimal resolution schedule from the model’s power spectrum. Our framework supports training-free acceleration and a fine-tuning recipe that further improves efficiency and quality, enabling significant speedups on state-of-the-art pretrained image and video generation models while preserving visual quality."
With this we can get between 1.4 and 2x training-free speedup for both image and video generation, and up to 5x when using special finetune/lora, while mostly preserving quality, although image composition changes quite a bit.
There's no official implementation/code yet, all I've found is comfyui node: https://github.com/ruwwww/ComfyUI-SPEED
Alternatives you considered
No response
Additional context
No response