Inference Frontier

For one model on one GPU, every optimization recipe is measured against the baseline. The frontier is the recipe that solves

minimizelatency(recipe)subject toloss(recipe, baseline) ≤ ε

for the quality loss ε you choose — LPIPS for image and video, word-error-rate increase for speech.

Image models

seconds per image · LPIPS vs baseline

baseline 13.6s

LPIPS	Latency	Speedup	Recipe
.011	7.1s	1.9×	DPCache K=20
.080	4.5s	3.0×	DPCache K=12
.129	2.0s	6.7×	FP8 W8A8 per-channel + SageAttention2 + fused text-encoder and VAE kernels + DPCache K=12

View benchmark

baseline 17.8s

LPIPS	Latency	Speedup	Recipe
.120	3.1s	5.8×	FP8 W8A8 (double-stream blocks BF16) + SageAttention2 + fused concat + FP8 quantize + DPCache K=16
.193	2.6s	6.9×	FP8 W8A8 + SageAttention2 + fused concat + FP8 quantize + DPCache K=16

View benchmark

Baseline: the engine's native BF16 run at the page's step count. Every recipe is compared to its output.
Quality: a loss against the baseline output for the same prompt and seed, over a fixed prompt set — LPIPS for image, LPIPS for video, ΔWER for speech.
Speed: latency of a single request, median after warmup — seconds per image or clip, milliseconds to first audio for speech.
Verified: re-measured by the maintainers on a private held-out prompt set.

One pull request, one recipe config; maintainers measure and publish it. GitHub ↗