Ex-Omni-2D: Expressive Omni-Modal Dialogue Models with Native Visual Presence
Paper • 2608.10720 • Published • 15
Ex-Omni-2D
Paper: Ex-Omni-2D: Expressive Omni-Modal Dialogue Models with Native Visual Presence
Project page: https://logo-cuhksz.github.io/Ex-Omni-2D
Code: https://github.com/LOGO-CUHKSZ/Ex-Omni-2D-Code
This repository contains the weights required by the public Ex-Omni-2D inference pipeline. The current release includes the full-sequence Teacher; the causal Student checkpoint is not included yet.
Contents
model-*.safetensors: Ex-Omni-2D language, vision, and speech model.video_generator/teacher.safetensors: Teacher video-generator LoRA.- tokenizer, generation, model configuration, and shard index files.
The inference code downloads the remaining public dependencies from their canonical Hugging Face repositories: Wan2.1-T2V-1.3B, Qwen3-VL-2B-Instruct, Qwen3-TTS-Tokenizer-12Hz, Qwen3-TTS-12Hz-0.6B-Base, Whisper-large-v3, and OmniAvatar-1.3B.
Usage
Use the release configuration in the Ex-Omni-2D inference repository. Weight
references use hf:// URIs and are downloaded lazily into the standard
Hugging Face cache.
- Downloads last month
- 51
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support