Native 3K MiniMax H3 video on a 4090: one ComfyUI node kills the VRAM crash
friedlc · reddit · 2026-10-30
After weeks of MiniMax H3 in ComfyUI crashing above 1080p with CUDA error: illegal memory access on an RTX 4090, the author found Zironic's H3-Optimizations node: dropped between the model chain and guider/sampler, it chunks attention and MLP work to eliminate memory spikes, keeping the math equivalent (no quality loss).
Benchmark results (4090 24GB, ComfyUI 0.43.0, SageAttention, 25 steps, 192 frames / 8s with audio):
- 1920×1088: 65 s/step, 28 min total
- 2560×1440: 1 h 13 min
- 3072×1728 (3K): 2 h 21 min, VRAM steady at 23.6 GB with zero OOMs
- 3840×2176 probe ran; full render estimated 5.6 h
Key takeaways: 3K output lands all five prompted fight beats with no doubled limbs and clearly better detail than 1080p; but changing resolution re-stages the shot, so you can't upscale a favorite take — sweep seeds at target size; native 1080p is steadier frame-to-frame than 720p + SeedVR2 upscale. Time per step grows faster than pixel count (4× pixels ≈ 11× time). The author says it saved $10k on a 5090 and published the full 3K workflow.
More from Infra
- Google's Ed Chi asks: will future phones be thin clients or run more on-device intelligence? — edchi · 2026-10-30
- UkisAI cuts thinking tokens 61.4% on GLM-5.3-Flash with under 0.5% accuracy loss — Secure_Recording_472 · 2026-10-30
- tiny-llm: an open-source course where you build a vLLM-style inference engine for Qwen3 on a Mac — techNmak · 2026-10-30
- WanGP one-click launcher hits Pinokio: AI video generation on 6GB+ VRAM GPUs — cocktailpeanut · 2026-10-30
- DeepSpeed fixes ZeRO-2 fp32 precision bug, users urged to upgrade to 0.19.8+ — StasBekman · 2026-10-30
- AMD ships ROCm HRX Loom v0.1 targeting its CPUs, GPUs and NPUs — Fcking_Chuck · 2026-10-30