Introduction

Local MiniMax H3 just got its own director's chair. On October 16, NVIDIA's Sol engine team pushed Sol-H3 in ComfyUI to the sol-engine branch of the NVlabs Sana repo: an Apache-2.0 custom node that opens Sol-H3 Studio, a full-screen video workspace over the ComfyUI canvas. You type a prompt, pick a mode and a pipeline, hit Generate, and a 5-second 1344×768 clip with native audio comes back in 35.5 seconds on a single RTX 5090. The original 50-step H3 needs 989 seconds for the same job on the same card.

We've covered the Sol engine racing H3 on datacenter racks before. This is the part makers have been waiting for: the speedups packaged as something you can actually click on your own GPU.

What shipped

Sol-H3 Studio lives in models/minimax_h3/Sol-H3-ComfyUI and treats ComfyUI as the harness. The Studio sits on top in English or Chinese, with live progress, a gallery wall and a library of past videos. The model runs in its own worker process, so your ComfyUI packages stay untouched, and a Nodes toggle drops you back into the same graph whenever you want to tinker.

Two pipelines ship side by side:

  • One stage: the 8-step HyperFlow LoRA merged into an FP8 H3 DiT, plus an exact AdaLN table (instead of 26 GB of projection weights), Sol-Attn with SageAttention 2++, regional torch.compile, and a distilled noncausal VAE. It handles text, a first frame, or reference images, at 768p or 480p.
  • Two stage: an H3 draft refined by LTX, driven by the existing Sol-H3-RTX5090 package (or Sol-H3-Spark on DGX Spark). It is the fastest path, text-to-video at 768p.

Clips are 16:9, 5 seconds, with sound. The DiT streams block by block and the Qwen text encoder waits in host RAM between requests, so the whole thing fits a 32 GB card.

The speed numbers

NVIDIA's README times one 1344×768 clip (121 frames with audio, model already loaded) from request to finished MP4, Qwen encoding included:

  • Original MiniMax H3, 50 steps, BF16, RTX 5090: 989.0 s
  • HyperFlow 8-step LoRA, BF16, RTX 5090: 193.6 s
  • Sol-H3 one stage, RTX 5090: 67.2 s
  • Sol-H3 two stage, RTX 5090: 35.5 s (27.9× faster)
  • Sol-H3 two stage, DGX Spark: 64.3 s
  • Sol-H3 one stage at 480p (864×480), RTX 5090: 22.0 s

A quick reality check, straight from the repo's validation notes: the RTX 5090 rows are single formal requests on a host with a 90 GB RAM limit, not averages, and model loading and warmup sit outside the clock. The team also says plainly that the distilled noncausal decoder may change quality and claims no quality equivalence. First requests load and compile, so expect a slow warmup before the fast clips start.

What you need

This is a power-user install, so check the box before you start:

  • Linux, one RTX 5090 (32 GB) and at least 90 GB of host RAM
  • ComfyUI 0.30 or newer
  • A runtime env with Python 3.12, PyTorch 2.12 (CUDA 13), Triton 3.7, FastVideo, and FlashAttention 4 plus SageAttention 2.2 built for SM120
  • For one stage, an overlay with Diffusers 0.40, PEFT and hyperflow_h3

Weights come from their own homes: MiniMax-H3, the NVFP4 Qwen encoder from Comfy-Org, the HyperFlow 8-step LoRA, the FastH3 dense-datafree adapter, and the optional Sol-H3-VAE distilled decoder that NVIDIA posted to Hugging Face the same evening. Skip the VAE and one stage falls back to the original H3 decoder, about 10 seconds slower per clip.

Install

git clone -b sol-engine https://github.com/NVlabs/Sana.git
ln -s "$PWD/Sana/models/minimax_h3/Sol-H3-ComfyUI" /path/to/ComfyUI/custom_nodes/

Then build the one-stage checkpoints once (the FP8 DiT, its AdaLN table and the reference-image DiT), or prepare Sol-H3-RTX5090 for two stage. Run scripts/configure.py with your paths; it writes config.json and prints the SOL_ENGINE_CONFIG=... line to export. Start ComfyUI and the Studio opens with it. Want a peek without a GPU? Run python3 -m http.server in the node folder and open scripts/studio_preview.html for a stubbed frontend preview.

License notes

The Studio code is Apache-2.0. Model weights and external runtimes are not bundled and keep their own terms, and MiniMax-H3 ships under the MiniMax H3 Community License Agreement. Read that license for your region before you build anything commercial on a local install.

Why it matters

H3 has been the open video model everyone wants to run and few could afford to wait on. A 16-minute render is a coffee break per idea; 35 seconds is a creative loop. Packaging the Sol tricks as a click-to-generate Studio, with the node graph still one toggle away, is exactly the kind of bridge that turns a benchmark blog into a daily tool. The hardware bar is still steep, but the direction is lovely: datacenter speedups landing on a desk.

Original Source

Announcement from Enze Xie:

Conclusion

Sol-H3 Studio is NVIDIA's Sol engine with a friendly front door: MiniMax H3 video with native audio, 5 seconds in 35.5 seconds on one RTX 5090, inside ComfyUI and open source. If you have the card and the RAM, it is the fastest local H3 path we've seen packaged for creators. If you don't, the cloud is still a click away.

No RTX 5090 handy? Try MiniMax H3 on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=sol-h3-studio-comfyui-minimax-h3-rtx-5090

—Aurelia ♡