minimax-h3-latent-core

Partial rewriting of MiniMax H3 latents. Pure PyTorch - nothing here imports ComfyUI.

Read this before your company adopts it. This is not an open source licence. It is PolyForm Small Business 1.0.0: free for individuals, and free for a company with fewer than 100 people (employees and contractors combined) and under USD 1,000,000 revenue in the prior tax year. Below that line commercial work is fine - the threshold is about your size, not about what you do with it. Above it, use needs a paid licence. Nothing phones home and nothing checks; you are expected to read the line and act on it. COMMERCIAL.md says how to get one. What you generate is yours either way - the licence makes no claim on the output.

An H3 latent is two tensors:

video   (B, 24, T, H, W)     H, W = pixels / 16
audio   (B, 32, 2, L)        L does not depend on resolution

Because they are separate, one can be pinned while the other is resampled - which is what makes an audio-only re-roll cost a fraction of a full generation, and what lets a rectangle of the picture be rewritten without disturbing the lettering around it.

Install

pip install minimax-h3-latent-core

Distribution and import name are the same. Both carry the minimax prefix because h3 on PyPI is Uber's geospatial indexing library, which this has nothing to do with.

Use

import torch
from minimax_h3_latent_core import build_masks, resize_video_only, save_tensors, load_tensors

video = torch.randn(1, 24, 8, 12, 20)
audio = torch.randn(1, 32, 2, 300)

# free the right half of the frame, only in the middle of the clip; pin the soundtrack
masks = build_masks([video, audio], track='video only', strength=0.9,
                    t_start_pct=25, t_end_pct=75, t_feather_pct=0,
                    x_pct=50, y_pct=0, w_pct=50, h_pct=100, feather_pct=0)

small = resize_video_only([video, audio], 6, 10)   # audio is left alone
save_tensors('clip.safetensors', [video, audio])
tensors, nested, prompt, meta = load_tensors('clip.safetensors')

Traps this library exists to avoid

• The audio tensor is 4-D, like an image latent. Resizing "anything 4-D or 5-D" stretches the soundtrack. resize_video_only touches only the 5-D tensor.
• A masked pass needs real noise. Disabling noise collapses the inpaint term and the region comes back as blocks.
• Time resolution is one latent frame, roughly four rendered frames. A span cannot be selected more finely, and the frames at either end of a selection move with it.
• Length must satisfy length % 17 == 5 when a latent is produced by encoding a video. An invalid length silently comes back shorter.

Licence

PolyForm Small Business 1.0.0 - free for individuals and for companies with fewer than 100 people and less than USD 1,000,000 (2019, inflation-adjusted) revenue in the prior tax year. LICENSE has the exact wording; the summary here is not the licence.

Above that threshold, see COMMERCIAL.md.

This library imports nothing from ComfyUI, which is why it can be licensed separately from the ComfyUI nodes that use it. Those live in ComfyUI-MiniMax-H3-Inpaint-Tools under GPL-3.0.

Contributing

Issues only, no pull requests - CONTRIBUTING.md explains why.