Four H3 workflows on one Halloween scene: fastest in 3:52, stock 25-step takes 14:32

rileygstaliger · reddit · 2026-10-28

Riley G Staliger, executive producer at OME OMY Productions, ran four HunyuanVideo (H3) recipes on the same Halloween porch scene using a local CUT/ComfyUI setup: each produced a 10-second opening plus a continuation from its own saved latents, generated at 960×544 and upscaled 2× with a learned-latent upscaler to 1080p.

The four recipes: 10Eros beta6 Turbo with VSA at 8 steps; Singularity V2 Turbo at 4 steps; Singularity V2 Turbo at 8 steps with the same adapter pairing; and stock H3's pruned INT8 ConvRot variant at 25 steps with no speed LoRA
Graph-time totals (opening + continuation + two HR passes): Singularity4 3:52; 10Eros8 4:37; Singularity8 5:11; stock25 14:32 — though the author cautions these are this run's times, not a universal speed ranking
Each model generated its own dialogue and scene sound from shared voice references; audio got only a single constant gain per take, no replacement speech or EQ
First completed scored takes kept as-is, including flaws; continuation resets and duplicated speech remain visible as comparison evidence
The free comparison video plays takes A–D back to back, with the recipe reveal at 3:24

The author also asks the community what works best for identity and geography across continuation seams — character-sheet and reference-layout choices in particular.

Original post →

Four H3 workflows on one Halloween scene: fastest in 3:52, stock 25-step takes 14:32

More from Multimodal