MiniMax H3 I2V: how to stop secondary faces' eyes drifting in 9:16 clips
Downtown-Term-5254 · reddit · 2026-10-21
A Reddit user making 9:16 gym promo clips with MiniMax H3 image-to-video in ComfyUI finds the main face holds up, but the secondary character's eyes (90-100 px wide) drift and deform frame to frame, looking obviously AI on close inspection.
Workflow: Qwen Image 2.1 Edit for a clean 1152×2048 start frame → minimaxh3fl2vturbo LoRA, 8 steps, 5s @ 24fps (124 frames) → RTX Video Super Resolution to 1080×1920; 34 min per clip on an RTX 3060.
Tried: 0.6 MP wrecks small faces; 1.5 MP keeps the main face clean but secondary faces still drift; turbo off at 20 steps shows no gain; Realism People LoRA improves skin but destroys the camera move. They want to avoid CodeFormer/GFPGAN per-frame restoration ("restored" look + flicker) and are asking for temporally consistent fixes like masked low-denoise video-to-video on the face region. Workflow JSON offered.
More from Multimodal
- Flux Klein's reference latent node excels at style transfer — can Qwen 2.1 do it? — Inevitable_Pen9043 · 2026-10-21
- AI-generated character IP demo: one mascot across posters, merch and conference scenes — eyal_zuri · 2026-10-21
- Cinematic AI fantasy romance short 'Alpha Meik & Luna Martina' released — Soundangel-Martina · 2026-10-21
- AcademiaSD adds Fibo 2.0 LoRA training with NF4 quantization, cutting VRAM from 16GB to 5.2GB — AcademiaSD · 2026-10-21
- "We went deeper and found more sky": stunning AI-generated megastructure sci-fi art — piotrbinkowski · 2026-10-21
- AI reimagines the 1993 skating game Skitchin with strikingly faithful 3D visuals — demondisc · 2026-10-21