Running MiniMax H3 locally on a 12GB GPU: 0.6MP at 4:3 is the sweet spot
vortis23 · reddit · 2026-10-06
A Reddit user tried making a cohesive movie trailer locally on a 12GB GPU using ComfyUI and the 30-second Ref2V MiniMax H3 workflow, with character/environment references generated in Akool.
Key findings:
- At 16:9, a 12GB card only handles 0.3-0.4 megapixels without heavy distortion; quality needs 0.5-0.6MP and 25-30 steps, which crashes or takes over an hour per generation
- The sweet spot is 0.6MP at 4:3 or 3:2 — enough character detail, consistency and prompt adherence, 19-21 minutes per clip; complex 30s clips took 1h05m
- Audio is the weak link: the character flubbed dialogue often, wasting many otherwise-good generations
- With exact prompting, scheduling and steps, MiniMax H3 can rival Seedance 2/2.5 — but demands extreme prompt precision and hardware
More from Infra
- XFreeze: high-bandwidth memory holders will win the superintelligence race in 1-3 years — XFreeze · 2026-10-06
- 691K H100-hours: community estimates compute behind from-scratch checkpoint trained on just 320 H100s — teortaxesTex · 2026-10-06
- Dev reports Clef-Flash runs fast locally even on memory-bandwidth-limited Jetson Orin — gregmushen · 2026-10-06
- Running LLMs Fully Client-Side: llama.cpp Compiled to WASM with WebGPU Proof of Concept — Numerous-Fan8138 · 2026-10-06
- Llama.wasm: llama.cpp Compiled to WASM with WebGPU Runs LLMs Fully in the Browser — Numerous-Fan8138 · 2026-10-06
- NVIDIA Dynamo lets coding agents point at self-hosted endpoints with native tracing — TheZachMueller · 2026-10-06