Running MiniMax H3 ref-to-video on 12GB VRAM, chasing facial consistency
PerfectionPlayer · reddit · 2026-11-21
A Reddit user shares a local MiniMax H3 reference-to-video workflow on an RTX 4070 Super (12GB VRAM, 32GB RAM), using an int8-quantized video model, a Qwen3-VL-32B CLIP (NVFP4 AWQ), and an int8 VAE. Compared with Qwen Image Edit and Qwen Image 2.1, MiniMax H3 keeps facial identity better but still falls short. They ask the community three questions: how much better hardware/stronger models help identity consistency, whether training a custom LoRA is really worth it and how, and which ref-to-video models best preserve faces today.
More from Multimodal
- Merzmensch launches trilingual AI-art newsletter MERZblown, opens with Pangram witch-hunt — Merzmensch · 2026-11-21
- Redditor shares stunning anime-character photorealistic rain-scene generations — plsdontultme · 2026-11-21
- Midjourney + Nano Banana 2.1: turning anime stills into photoreal photos — gizakdag · 2026-11-21
- isometric-cities: coding agents turn real neighborhoods into pixel-art maps on Mac — zsakib_ · 2026-11-21
- AI turns a simple story into an entertaining video — HeyZoyaKhan · 2026-11-21
- Alchemical Circuit Board prompt fuses medieval mysticism with electronics — LudovicCreator · 2026-11-21