MiniMax H3 Runs Locally on an RTX 4090 as Users Share ComfyUI Workflows

`nvfp4_awq` `pruned_int8_convrot` `fp8_scaled` First run: start small. Template library → Video → MiniMax H3 → Text to Video (the three `api_`-prefixed templates call the cloud API and need a key — ignore them for now). Check that the UNETLoader points at the FL2VA weight, the encoder is the NVFP4 build, and both VAEs are loaded. A sensible 12 GB starting point: 16:9, 864×480 (~0.4 MP), 5 seconds, Turbo on, Turbo steps 8 — this validates the environment rather than chasing quality. Keep the prompt to one scene, one main action, one shot, and sound. The first generation is always slow (about 34 GB of weights load first; reference: 4090 48G first run ≈ 644 s vs ≈ 425 s warm). `api_` [...] Prompt format. H3 uses an official structured prompt format (full Prompting Guidance in the official README). A practical trick: hand your Chinese/plain brief plus the official prompt guide to any LLM and let it rewrite into the structured format, then save as a template — an effective community stand

How I Started Making AI Videos with Audio Using 'MiniMax H3' you need a VRAM 24GB class GPU. That's the world of the RTX 4090 or 5090. This

MiniMax H3 for ComfyUI is here! I am getting insane results out of it on just a RTX 4090 laptop. Leaps and bounds if we can get a replacement

A rough tiering, based on what’s been demonstrated: 6GB VRAM (RTX 2060 class): workable but degraded, expect mushy detail and rougher audio, generation times around 10-15 minutes for short clips. 12-16GB VRAM (RTX 4060 class): noticeably cleaner, faster generation, still below full quality. 24GB VRAM (RTX 4090 class): enough headroom to also do LoRA training on top of generation. 32GB+ VRAM (RTX 5090, RTX 6000): the intended target, supports longer 15-second clips at higher resolution with less compromise. The pattern lines up with most modern diffusion-based video models: more VRAM buys you resolution, generation length, and speed, but low-VRAM setups aren’t dead ends anymore thanks to aggressive quantization and optimization work from the open-source community. [...] MiniMax H3 is fully open source, meaning anyone can download the weights and run generation locally instead of relying on a paid API. VRAM needs scale from roughly 6GB to 24GB+, with lower-VRAM setups producing so

The storybook film took an afternoon once the pipeline worked: six illustrations, three techniques, about 25 minutes of GPU time for the shots plus one reference-mode generation. A year ago this was a render-farm conversation. ## Workflows for download Everything in this post runs on core ComfyUI nodes — v0.30 or later, no custom node packs to install. The weights are on Hugging Face at Comfy-Org/MiniMax-H3, and MiniMax’s structured prompt guides ship in the model repo’s `docs/` folder (`VIDEO_PROMPT_WRITING_GUIDE_base_en.md` and the `_ref_` variant for reference mode). [...] Meanwhile, rendering the same freeform prompt at 5× the pixels changed nothing: 608×352 scored 29.2, 864×480 scored 30.3, 1344×768 scored 26.0. If your generations look soft, the fix is not more resolution. To be fair to the model: the atmospheric prompt asked for shallow depth of field, film grain, sea mist, and volumetric fog — and H3 rendered exactly that. It was doing its job. But the gap between hand-tuned

Very useful resources on making videos with the Minimax H3 model on an RTX 4090 (24GB) 😊 The showcases look amazing, and detailed prompts plus