H3 Ref2V workflow notes: 7 references break cohesion at low steps, 10 steps worked
R34vspec · reddit · 2026-09-05
A creator shares hands-on notes from making the short film "The Bat" with H3 Ref2V: they used up to 7 reference clips at once, a personal record, and found that more references make it harder for the model to produce cohesive shots at low step counts — they had to move from 8 steps with the turbo LoRA to 10 to get stable results. Originally intended for the now-ended H3 audio sync contest, the author feels the piece would have benefited from a slightly longer runtime.
More from Multimodal
- Reddit user compares two MiniMax H3 Turbo LoRAs for video generation — No-Bee-231 · 2026-09-05
- MiniMax H3 workflow: REF2VA visuals + FL2VA audio + LightX2V speed, 198s on RTX 5090 — gabxav · 2026-09-05
- AI short film "Ishiro - The Last Blade" shared on Reddit — Nervous-Barnacle-888 · 2026-09-05
- Dev calls for a MiniMax H3 finetune that fills speech gaps with human filler words — cocktailpeanut · 2026-09-05
- Nvidia DLSS 5 frame interpolation discussed in Stable Diffusion community — KonoTheSavage1 · 2026-09-05
- Perceptron's Multilook API Prefills Video Context Once, Cuts Input Cost to 32% at 16 Prompts — AkshatS07 · 2026-09-05