Emu3.5-Image-FlashAR

This repository hosts the FlashAR checkpoint for Emu3.5-Image.

FlashAR is introduced in "FlashAR: Efficient Post-Training Acceleration for Autoregressive Image Generation". It accelerates a pretrained raster-scan autoregressive image generator by adding a vertical prediction branch and a learnable fusion gate. Decoding proceeds by anti-diagonal steps, reducing the serial image-token decoding length from H * W to H + W - 1.

• Paper: arXiv:2605.09430
• Project page: FlashAR
• Code: Emu3.5-NAR

Checkpoint

This checkpoint corresponds to:

• Base model family: Emu3.5-Image
• Checkpoint name: Emu3.5-Image-FlashAR
• Training step: 74000
• Default visual-token grid used by the release scripts: 32 x 32
• Default CFG setting used for the released checkpoint metadata: 5.0

The weight shards are stored at the repository root:

model-00001-of-00016.safetensors
...
model-00016-of-00016.safetensors
model.safetensors.index.json
config.json
checkpoint_meta.json
configuration.json

Usage

Use this repository as the FlashAR checkpoint path together with:

• the base Emu3.5-Image model;
• the Emu3.5 vision tokenizer;
• the FlashAR code from the project repository.

Example layout:

weights/Emu3.5-Image/
weights/Emu3.5-VisionTokenizer/
checkpoints/Emu3.5-Image-FlashAR/

Example generation command from the code repository:

MODEL_PATH=./weights/Emu3.5-Image \
TOKENIZER_PATH=./src/tokenizer_emu3_ibq \
VQ_PATH=./weights/Emu3.5-VisionTokenizer \
CKPT_PATH=./checkpoints/Emu3.5-Image-FlashAR \
PROMPT="a red car parked next to a blue mailbox" \
CFG_SCALE=5.0 \
OUT_PATH=./outputs/sample.png \
bash generate.sh

Citation

@article{zhou2026flashar,
  title={FlashAR: Efficient Post-Training Acceleration for Autoregressive Image Generation},
  author={Zhou, Junkang and He, Yefei and Chen, Feng and Wang, Weijie and Zhuang, Bohan},
  journal={arXiv preprint arXiv:2605.09430},
  year={2026}
}