Modelpublic

Efficient-Large-Model/H3-to-LTX-Latent-Adapter

sourceHugging Faceupdated 16d agoView on Hugging Face

1likes957downloads

Model Card

H3-to-LTX Latent Adapter

An original BF16 inference adapter that converts normalized MiniMax-H3 video latents into normalized LTX-2.5 Conv VideoVAE latents. It is used by Sol-H3-Spark in Sana.

The model applies the released temporal/spatial packing transform and a 194,759,504-parameter Conv3D residual adapter (width 752, 22 blocks). It does not contain MiniMax-H3 or LTX model weights and does not generate video alone.

Boundary	Tensor
Input	Normalized H3 latent, `[B, 24, T_h3, H/16, W/16]`
Output	Normalized LTX latent, `[B, 128, T_ltx, H/32, W/32]`

Pass the original pixel-frame count and pixel geometry to the adapter's convert method. The inference library in Sana performs the matching packing and normalization. Changing that transform is not interchangeable with this checkpoint.

Files:

—model.safetensors: BF16 adapter weights, 389,541,112 bytes.
—config.json: architecture, geometry and normalization contract.

Weight SHA-256: 170199a390c40ac97f5895bc9c8cc29817e74fb9193c858a85d8c0f1f30724ac.

This model card does not assign a new weight license. MiniMax-H3 and LTX components used with the adapter remain subject to their respective terms.