Efficient-Large-Model

/

Open-H3-Regenerate-2K

Like 15

Follow

Efficient-Large-Model 715

Weights for https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K (project page https://nvlabs.github.io/Sana/Open-H3-Regenerate-2K/): merged refiner LoRA and Sharp curve branches, the latent upscaler, the fast H3 VAE decoder (StreamRender-H3), and an unmodified copy of MiniMax-H3 @ 6818f6c. Byte-identical clone of Sensen02/Open-H3-Regenerate-2K at aaa3d97db80f7d6c2c251ef0cb8702537f2093f1; every file's size and sha256 checked after upload.

Thanks for putting this together. The weights check out on my side:

• All 62 files under base_model/ match MiniMaxAI/MiniMax-H3 @ 6818f6c by sha256.
• latent_upsampler/ matches LBH-123-AI's bf16 file.
• The four release files match the EXPECTED hashes in download.py.

Before we merge, a few licensing fixes, mostly about the second commit:

1. Please restore base_model/LICENSE and add a NOTICE file at the repo root. Section III of the MiniMax H3 Community License requires any redistribution to include a copy of the agreement (III.1) and a NOTICE file with this exact text (III.4):

   MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

   base_model/README.md also still has license_link: LICENSE, which is a dead link without the file.
2. Please restore latent_upsampler/LICENSE. MIT requires the copyright notice to stay with copies. (LBH-123-AI's HF card says apache-2.0 while the GitHub LICENSE is MIT; either way the license text needs to ship with the file.)
3. Please add license metadata to the top-level README, e.g. license: other, license_name: minimax-h3-community-license-agreement, license_link: base_model/LICENSE.
4. Could you confirm that the LightX2V Turbo and LMS Sharp LoRAs allow redistribution? Neither upstream repo states a license for those files, and both are merged into our LoRA.

Minor:

• The project page https://nvlabs.github.io/Sana/Open-H3-Regenerate-2K/ currently returns 404. We'd like to merge after it is live.
• The LMS Sharp link in the README points to main, while lora_config.json pins 849fb3d. Pinning the link to the same revision would keep them consistent.

If the LICENSE files were removed on purpose, let me know the reason and we can sort it out.

Thanks for the README updates. The License section and the license metadata look good. What is still missing are the license files that need to ship with the weights:

1. LICENSE (MiniMax H3 Community License, official text)
   • Restore base_model/LICENSE verbatim from MiniMaxAI/MiniMax-H3 @ 6818f6c (the same file removed in c2a37aa). This also fixes license_link: LICENSE in base_model/README.md.
   • Add the same file as LICENSE at the repo root. The refiner LoRA and the fast H3 VAE decoder are Model Derivatives, so Section III applies to them as well.
2. NOTICE at the repo root, required by Section III.4, containing exactly:

   MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

3. Apache 2.0 license text for two components:
   • base_model/text_encoder/LICENSE: the text encoder is Qwen3-VL-32B, which is Apache 2.0 (noted at the end of the H3 license).
   • latent_upsampler/LICENSE: the weights come from LBH-123-AI/Minimax_h3_latent_Upscaler on HF, which is tagged apache-2.0. (The file removed earlier was the MIT license of the GitHub code repo.)

These are text files only; the weights don't need to change. Thanks!