Real-world local video gen on a 4090: MiniMax H3 caps at 2-4 seconds
reason808 · reddit · 2026-10-24
A user with a traditional 3D animation background shared detailed findings from running ComfyUI video generation locally on a 24GB RTX 4090.
- MiniMax H3: at best 4 seconds at 0.2MP (608x342) or 2 seconds at 0.6MP (1024x576); anything beyond, or adding a reference image, causes black-screen crashes — far short of reported "15s at 720p" on the same hardware.
- Wan 2.2 5B: handles 124 frames at 1280x704 for t2v/i2v, but crashes after roughly 7 successful generations.
- Image generation is far more stable: Flux.2 [Klein] 4B Distilled produced 300 1024x1024 images without issues; Flux.1 Dev fp8 and Qwen Image 2.1 t2i also work, though reference-image edit workflows crash easily.
Troubleshooting revealed the frequent "GPU crashes" (black screens while the PC kept running) were actually caused by a faulty monitor — hardware passed days of stress testing. The author mitigates GPU load via MSI Afterburner power limiting and .bat launch tweaks, and asks the community what's realistically achievable on 24GB VRAM versus going cloud.
More from Infra
- PS5 Pro sells out in Toronto as memory shortage hits consumer hardware — Rasmic · 2026-10-24
- Open-source kapitsa engine boosts Mac local agent throughput 23% with tree speculative decoding — DotaMate · 2026-10-24
- Nvidia compute unlock boosts RTX 3080 20GB prefill by up to 29%, decode unchanged — No-Refrigerator-1672 · 2026-10-24
- Conflicting RSI evidence: NVIDIA says AI-assisted CUDA optimization went 5x, then 2x in a month — soumitrashukla9 · 2026-10-24
- Miles v0.1.3 cuts first RL weight update from 99.5s to 3.9s, shares GPUs across train/rollout — hsu_byron · 2026-10-24
- 98%+ Tensor Core utilization shown in LLM training run, drawing admiration — gharik · 2026-10-24