Running MiniMax H3 Fully Local on a 16GB GPU: 8-Step Video Generation & QA Lessons
Short_Regular_7191 · reddit · 2026-08-08
A developer shared a hands-on guide to running the 33B omni-modal MiniMax H3 video+audio model entirely locally on a 16GB RTX 5060 Ti.
Stack & Optimization
- Using ComfyUI's native support, NVFP4 pruning, and system RAM offload, the workflow bypasses VRAM limits smoothly.
- By utilizing a Turbo LoRA and a dual video/audio clock sampler, sampling steps were reduced from 20 to 8 steps. Generating a 5.2s native 1344x768 video takes just 11.4 minutes.
Automated Metrics vs. Human Eyes
- An automated QA script picks the best generation based on Whisper transcription accuracy, face similarity, and sharpness (variance of Laplacian).
- The script favored a clip with higher micro-contrast, but human review showed the lower-scoring clip was significantly more lifelike with natural motion. The new rule: whenever face similarity and sharpness metrics conflict, trigger manual human review.
Upscaling: Less is More
- Compared plain Lanczos upscaling against the SeedVR2 3B AI upscaler. For already high-res native generations, AI upscaling added artificial micro-contrast, making the result look "etched." A simple 1.4x Lanczos + crop won visually.
More from Infra
- TokenSpeed MLA 2 hits Vera Rubin: up to 3.44× faster MLA decode for Kimi K3 — zhyncs42 · 2026-10-16
- US vs Chinese hyperscalers: Americans fund and schmooze, Chinese demand your BoM and cap you at 7% margin — jwt0625 · 2026-10-16
- How LiteLLM routes one OpenAI-shaped API across every model provider — Mahmoud_Zalt · 2026-10-16
- OpenAI-Broadcom custom AI chip Jalapeño named a TIME Best Invention of 2026 — itsclivetime · 2026-10-16
- Developer live-runs continual pretraining of a Luganda LLM on Colab, 41% done with loss dropping fast — weskambale · 2026-10-16
- Jensen Huang: Nvidia charges ~$thousands per GPU yearly for software, a major revenue line — rohanpaul_ai · 2026-10-16