Realtime Interactive Video on MiniMax H3: Character LoRA Training Puzzles and a Key ComfyUI Conversion Trap
Big-Set9728 · reddit · 2026-09-07
A developer built a realtime interactive video scene on MiniMax H3 — you type lines and the character answers with generated audio, with 5s clips generating slightly faster than realtime. The bottleneck is character consistency: they want a trained character LoRA instead of paying for reference conditioning each generation.
Key findings and a trap:
- FastVideo's FastH3 (4-step DMD2 distill) does 6s for a 5s clip on 4x B200 and ships a pre-extracted LoRA with lorapath/lorastrength hooks
- diffusion-pipe and ai-toolkit both list H3 support
- VSA weights don't survive stock ComfyUI conversion: 50 togatecompress tensors are silently dropped, producing noise; dense converts cleanly
Open questions: whether a base-H3 character LoRA transfers to the distilled checkpoint, whether it can stack with the distill LoRA, if T2I stills suffice for talking-face identity, and multi-GPU sequence parallelism numbers. Author offers paid consulting.
Related event: Dev Tests Real-Time Interactive Video with MiniMax H3(2 posts)→
More from Multimodal
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- Full animation short made with Google Gemini tools, 10,564 frames — AI_Andrew · 2026-10-26
- Same prompt, two styles: Claude Opus 5.5 vs GPT-6 Astra on a C-section 3D animation — Sourcecode12 · 2026-10-26
- ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26
- Community checkpoint 'Narrative Drift' brings art-directed visuals to Krea 2 Turbo — MoonbearAIArt · 2026-10-26
- Same prompt, split results: Claude Opus 5.5 vs GPT-6 Astra on a 3D C-section explainer — Sourcecode12 · 2026-10-26