Custom SPEED Sampler for ComfyUI Speeds Up Image Generation by 46%

A Reddit user (Common-Objective2215) has released a “vibecoded” ComfyUI custom node for SPEED (Spectral Progressive Diffusion), a novel sampling technique that reduces image generation time by roughly 46%. In posted comparisons, the SPEED node completed an image in 14 seconds versus 26 seconds without it, using the same seed and model (Anima). The core idea behind SPEED is that diffusion models don't need full high-resolution computation from the start. Instead, the process begins with lower-resolution denoising and progressively expands resolution during the final sampling steps, cutting unnecessary computation early while preserving detail later.

The implementation is described as a quick, unofficial port – the official code hasn't been released yet, so the node may not perfectly replicate the paper. Users can connect the “Sampler SPEED (Spectral Progressive)” node into any standard ComfyUI workflow via SamplerCustomAdvanced. However, the author warns that outputs can show artifacts and drift, likely due to the upsampling stages. Interestingly, torch.compile was found to make sampling slower and is not recommended. The repository includes before/after images and invites community improvement. For now, the node offers a practical, open-source way to test progressive upsampling speed gains in ComfyUI.