Fast FLF Loop Workflow — MiniMax H3 — v1.0

Where this should run: any RTX 40-series or 50-series (Ada/Blackwell) card with 16GB+ VRAM should comfortably handle this at 1.0 megapixel. 30-series and older cards will still work, but Sage Attention falls back to older kernels without the same speedup — expect noticeably longer render times. Under ~12GB VRAM, stick closer to the 0.4 megapixel preview resolution.

A note on that 337s vs 234s gap: your first generation after starting ComfyUI (or after loading a different model) will usually take longer, since models are still being staged into VRAM. Once everything's loaded, a second render back-to-back on the same models will be noticeably faster — that's where the 234s number comes from. This only holds if you queue another generation right after, without switching models or doing anything else that forces a reload in between. Keep that in mind when comparing your own numbers to the ones above.

On EasyCache quality: at the default threshold, EasyCache skips roughly half the sampling steps by reusing near-identical intermediate results. That's a big speedup but not free — abstract/geometric loops hide it well, detail-heavy or character content less so. Recommended: EasyCache on for previews and iteration, bypassed (or threshold lowered to ~0.1) for final renders.

Example generations

Two sample outputs included, both using the cobra reference image bundled with the workflow (matching prompt is preloaded in the first/last frame conditioning):

1.0 megapixel render

0.4 megapixel render

Load the cobra image into both the first-frame and last-frame LoadImage nodes to reproduce the examples, or swap in your own matching first/last frame pair.