DreamID-Omni — Quantized Model Mirror

Quantized checkpoints for DreamID-Omni by Xu Guo et al.

Original model: XuGuo699/DreamID-Omni

Available Checkpoints

File	Precision	Size	Notes
dreamid_omni_fp8.safetensors	FP8 (float8_e4m3fn)	~12 GB	Fastest, lowest VRAM. Recommended for ≤16 GB GPUs.
dreamid_omni_bf16.safetensors	BF16 (bfloat16)	~23 GB	Best quality. Recommended for ≥24 GB GPUs.

The original FP32 checkpoint is 46.6 GB. These converted versions are functionally equivalent but require significantly less disk space and VRAM.

Usage with ComfyUI-FFMPEGA

These models are automatically downloaded and used by the ComfyUI-FFMPEGA extension.

Set dreamid_precision to:

• auto — prefers FP8 if available, else BF16
• fp8 — use FP8 quantized weights
• bf16 — use BF16 weights

Model Details

DreamID-Omni is a unified framework for controllable human-centric audio-video generation. It generates identity-preserving talking-head videos with synchronized speech from face images and reference audio.

• Architecture: Fusion DiT (dual video + audio diffusion transformer)
• Base model: Built on Wan2.2-TI2V-5B + MMAudio
• Paper: arXiv:2602.12160
• Code: github.com/Guoxu1233/DreamID-Omni

License

Apache 2.0 — see LICENSE file. Same license as the original model.

Citation

@misc{guo2026dreamidomni,
  title={DreamID-Omni: Unified Framework for Controllable Human-Centric Audio-Video Generation},
  author={Xu Guo and Fulong Ye and Qichao Sun and Liyang Chen and Bingchuan Li and Pengze Zhang and Jiawei Liu and Songtao Zhao and Qian He and Xiangwang Hou},
  year={2026},
  eprint={2602.12160},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2602.12160},
}

Ethics Statement

This project is intended for academic research and technical demonstration purposes only. Users are strictly prohibited from generating illegal, defamatory, pornographic, or harmful content. Generated videos should be labeled as "AI-Generated" to prevent misinformation.