EvoAwaken-Workshop/DreamX-Creator-gguf

DreamX-Creator GGUF

Community GGUF conversion of GD-ML/DreamX-Creator for the native CPU runtime in rust-model-inference.

DreamX-Creator generates synchronized video and audio from a first frame and a text prompt. These files also contain the optional 2x video refiner components.

Files

Both files are required and must be used together. The runtime verifies their shared pair ID before inference.

File	Size	Contents
DreamX-Creator-Q8_0.gguf	12.83 GB	Q8_0 Creator video/audio/joint models and SR-DiT refiner; selected tensors remain BF16/F32
mmproj-DreamX-Creator-BF16.gguf	15.56 GB	BF16/F32 UMT5 text encoder, video/audio VAEs, latent upsamplers, and LightVAE

GGUF metadata:

• Source model: GD-ML/DreamX-Creator
• Main architecture: dreamx
• Projector architecture: clip / dreamx_creator
• Exporter version: 1
• Pair ID: 8a08f2e3bcafde55da1853e109bdc83f39f5e279eb3b8c02352637fc50993bbb

Download

hf download EvoAwaken-Workshop/DreamX-Creator-gguf \
  --local-dir models/DreamX-Creator

Usage

Build rust-model-inference, then run:

cargo build --release --bin rust-model-inference

target/release/rust-model-inference \
  --dreamx \
  --model models/DreamX-Creator/DreamX-Creator-Q8_0.gguf \
  --mmproj models/DreamX-Creator/mmproj-DreamX-Creator-BF16.gguf \
  --image first-frame.png \
  --prompt "A man speaking while seated on a yellow couch." \
  --negative-prompt "low quality" \
  --duration 5 --fps 24 --steps 50 --seed 0 \
  --target-spatial-tokens 880 \
  --refine --refiner-kv-len 9 \
  --latent-upsample flash --refiner-decoder lightvae \
  --threads 8 \
  --out dreamx.mp4

Add --dry-run to validate the model pair and inspect the memory estimate without running generation. Full inference produces base video/audio/mux outputs and, with --refine, refined video/mux outputs.

Conversion

The pair is exported from the original DreamX-Creator checkpoint with the repository's DreamX converter:

python3 tools/converter/dreamx/convert_dreamx_creator.py \
  models/DreamX-Creator \
  --out-dir models/DreamX-Creator \
  --outtype q8_0

Validation and limitations

• The matched real GGUF pair has passed loading/preflight and a reduced 64x64, one-frame, one-step native CPU end-to-end run.
• The runtime currently supports DreamX-Creator on CPU only.
• The reduced validation run does not establish official 2K output quality.
• CUDA/Python Oracle numerical parity has not been completed, and the refiner shifted scheduler has not been compared checkpoint by checkpoint. This runtime path is therefore marked Experimental.
• Review the runtime's current support status before production use.

References

• Official DreamX-Creator implementation
• Original model weights
• rust-model-inference
• DreamX-Creator technical report

Citation

@misc{zhu2026dreamxcreatordemocratizingnativeaudiovideo,
  title={DreamX-Creator: Democratizing Native Audio-Video Generation at 2K Resolution},
  author={Jiashu Zhu and Yanhao Zheng and Ruitian Tian and Rujing Dang and Shen Zhang and Bingze Song and Jiachen Lei and Ruimin Lin and Jiahong Wu and Xiangxiang Chu},
  year={2026},
  eprint={2608.31106},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.31106}
}

License

The original project and model are released under the Apache License 2.0. Follow the original project's terms and applicable laws when using generated content.