Original Title: Efficient-Large-Model/Open-H3-Regenerate-2K
NVIDIA (NVlabs Sana / SoL-Engine) provides an open alternative to the closed-source H3-Regenerate-2K API: it upscales MiniMax-H3's 768P draft (1344×768, 121 frames) by 2x in latent space, then uses MiniMax-H3 with a merged LoRA for three-step denoising to output 2560×1440 videos; the code also supports 1080p and two-stage upscaling to 4K. The merged LoRA concatenates three adapters by rank: the Open-H3 refiner (rank 384), LightX2V FL2VA Turbo v1.2 (rank 128), and LMS Sharp v1.0 (rank 64). The authors claim that compared to attaching the three adapters separately, the relative error across all 312 linear layers is only 2.2e-8. The repository also includes the authors' 3D convolutional latent space amplifier, a fast VAE decoder distilled from StreamRender-H3 (1.62B parameters), and an unmodified copy of the MiniMax-H3 base model. The repository was created on 10 26, and was not included in previous issues.
Rising Score 79 · Likes 223 · Downloads 562558
Reasons
- Addresses the need to "generate quickly at low resolution first, then produce 2K once satisfied," without paying for closed-source API calls; inputs can be H3 draft videos, existing H3 latents, or you can add --hyperflow to generate a 768P draft from a single prompt first and then refine it.
- Officially, on a single B200 GPU, processing a 5-second clip from 768P latents to 2560×1440 takes: 87.26 seconds with the original dense attention BF16; 37.39 seconds (2.33x speedup) after switching to Sol-Attn sparse attention, MXFP8, and CUDA Graphs; and a total of 22.45 seconds (3.89x speedup) after further switching to the fast VAE decoder (decoding time drops from 15 seconds to 1.6 seconds). The code README lists the contribution of each optimization individually.
- High download count of 562558 and 223 likes (indicating attention level only).
Usage
- The code is located in the sol-engine branch of GitHub NVlabs/Sana under models/Open-H3-Regenerate-2K: requires Linux, Python 3.11, PyTorch 2.11 + CUDA 13.0. Run pip install -r requirements.txt, then pip install -e ../../techniques/sparse_backends (Sol-Attn, required for forward pass acceleration); run python download.py --weights ./weights --hyperflow to download all weights.
- With an existing draft: python infer.py --weights ./weights --input-video draft.mp4 --prompt "…" --output out.mp4; with only a prompt: add --hyperflow; for fastest speed use --fwd sol_engine_fast, for output closest to the original use the default sol_engine or --fwd standard; --resolution options are 768p / 1080p / 2k / 4k, change --steps 2 to two steps.
- Sampling settings when integrating into your own pipeline: three-step Euler, sigmas [0.909375, 0.857143, 0.666667, 0], CFG 1.
Risks
- Speed figures are measured on B200 / GB200 hardware; MXFP8 and Sol-Attn have specific requirements for GPU and software stack; the authors have not provided data for consumer-grade GPUs, nor are there ComfyUI workflows.
- The first encounter with a new latent shape requires building the engine and recording CUDA Graphs (approx. 39 seconds); sol_engine_fast is faster, but the authors state that details deviate more from the original output.
- Licenses are combined: the refiner LoRA, fast VAE, and base model copy are under the MiniMax H3 Community License (including geographic, usage, and redistribution terms), while the merged LightX2V Turbo, LMS Sharp, and latent space amplifier are under their respective original licenses; there is no unified license.