Benchmark for SageAttention kernels using real attention shapes logged from ComfyUI models (image / video / audio)

What this is — and what it is not This is not a benchmark of how fast a model generates an image or video. No model weights, no inference pipeline. The benchmark runs on randomly generated tensors that reproduce the exact attention shapes — (batch, heads, seq_len, head_dim, dtype) — that real models use during sampling inside ComfyUI. More precisely: it measures only the attention operation itself, one step inside the denoising loop. Everything else — VAE, CLIP, scheduler, ComfyUI overhead — is

More from StableDiffusion

Nodes With Live Preview inside ComfyUI ?

StableDiffusion·6

Batch Image Captioning Generator

StableDiffusion·5

Anime Documentary Series

StableDiffusion·4

LTX 2.3 Lora Loader Audio / Visual splitter

StableDiffusion·4

Made an AI animated Mortal Kombat fight film to hype the 2026 movie — pure action, no dialogue

StableDiffusion·2

Fast & clean face swap workflow for ComfyUI (FLUX + InsightFace) — ready to use

StableDiffusion·7