MiniMax-H3 Ref2VA for HexAnima (Snapdragon NPU pack)

This is a modified (repacked and converted) version of MiniMaxAI/MiniMax-H3 (Ref2VA, reference-to-video) for the HexAnima Android app (https://github.com/ayanoby/hexanima, Hexagon V69 NPU: Snapdragon 8 Gen 1 / 8+ Gen 1). It is a Model Derivative in the sense of the license. The files are not usable with other tools.

Modifications (Section III.2 of the license)

All weight files in this repository were modified from the originals:

  • blkNN.bin, adalnNN.bin, r0.bin, r1.bin, *.f32, hn.bin, H64.f32, H256.f32: the INT8-ConvRot transformer diffusion_models/minimax_h3_ref2va_int8_convrot.safetensors from Comfy-Org/MiniMax-H3 repacked into HMX V69 int8 tiles (one file per block; the adaLN projections in separate files), plus the small input/output layers as raw floats.
  • lora/: the turbo 4-step LoRA loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors from the same repository, repacked into tiles; it is applied on top of the transformer and is subject to the same license.
  • vae/: the video VAE decoder vae/minimax_h3_video_vae_int8_convrot.safetensors (same repository) repacked into int8 tiles (vbNN.bin, glue/).
  • enc/henc.bin: the image encoder of the video VAE converted to a Qualcomm QNN HTP context binary (fp16).

The text encoder is shared with MiniMax-H3-Image: install ayanoby/hexanima-h3 too (files/h3/te).

Nothing here is meant to be used outside the Applicable Territory. Excluded territories: the United States, the European Union, the United Kingdom, South Korea (see LICENSE, Section I). Acceptable Use Policy: Exhibit A of LICENSE.

Layout

About 35 GB. The files go to files/r2v/ of the app as they are (create the empty marker .installed); manifest.json lists the size and sha256 of every file. A phone with 12 GB of RAM and about 60 GB of free storage (together with the H3-Image text encoder) is needed.

Powered by MiniMax H3.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ayanoby/hexanima-r2v-packs

Quantized
(96)
this model