---
library_name: pytorch
tags:
- bidirectional-flow-matching
- text-to-image
- image-to-text
- text-to-speech
- automatic-speech-recognition
---

# BVFM model weights

This model repository contains the deployment weights for the image/text and
speech/text branches of Bidirectional Variational Flow Matching (BVFM). Source
code, training/inference entry points, and environment instructions live in
the separate GitHub repository.

## Layout

```text
image/
  bvfm_image_step40000.pt
  FlowTiTok_512.bin
  README.md
speech/
  bvfm_speech_step299999_inference.pt
  merged_config.json
  semantic_vae_1000k/
    config.json
    metainfo.json
    dac/ema_state_dict.pth
  README.md
```

The repository intentionally excludes training optimizer states, periodic
snapshots, logs, metrics, generated samples, datasets, and caches.

Clone with Git LFS or use `huggingface-cli download`. Point the GitHub code at
the downloaded root with:

```bash
export BVFM_WEIGHTS_ROOT=/path/to/this/model/repository
```

See the branch-specific READMEs for exact file roles and checksums.

## Third-party components

`image/FlowTiTok_512.bin` is the released FlowTiTok image tokenizer required
by image encoding and decoding. `speech/semantic_vae_1000k` is the released
Semantic-VAE decoder required to map the 64-D speech latent to waveform. See
`THIRD_PARTY_NOTICES.md` and retain the corresponding upstream notices when
redistributing these files.