Local MiniMax H3 fight-scene testing on RTX 5090: HyperFlow wins, plus a double-patching trap

glusphere · reddit · 2026-09-29

A detailed local ComfyUI test of MiniMax H3 action scenes on an RTX 5090, with dozens of controlled renders.

Verdict: for action, HyperFlow 8-step with no LoRA plus taomate 3-step LoRA on the upscale pass gave the best weapon consistency and motion (25% slower than 4-step turbo); turbo remains better for dialogue/drama.
Method: same 30s fight (two chained 15s clips), fixed seed and prompts, scored by a local vision model across six axes (/30) plus manual review.
Pitfalls: step-distilled LoRAs must run at their trained step count—more steps made every one worse and all 8-step arms lost the kill; Weapon Combat LoRA at 0.7 beat 1.0.
HyperFlow trap: the original node patches the model's time embedder and ComfyUI caches the UNET, so a second HyperFlow job in the same process runs double-patched and produces garbage; fix via the Adudeguyman fork (v1.4.0) or restarting ComfyUI before each job—freeing VRAM doesn't help.
Notably, the vision judge ranked turbo first, but side-by-side the author found HyperFlow keeps the trident intact with cleaner motion—suggesting the judge is unreliable on this axis.

More from Multimodal