Researchers at TikTok have released the image generation AI 'BitDance,' which they used to create an autoregressive model that is faster and higher quality than Z-Image.
A research team including ByteDance, the developer of TikTok, and the Chinese University of Hong Kong released the image generation AI model ' BitDance ' on February 17, 2026. BitDance uses
BitDance: Scaling Autoregressive Generative Models with Binary Tokens
https://bitdance.csuhan.com/
BitDance is an image generation model with 14 billion parameters that was developed to solve the problem of slow generation processing, a weakness of autoregressive models.
The graph below shows the image generation speed on the horizontal axis and the benchmark score on the vertical axis. It can be seen that it is 4.3 times faster and of higher quality than GLM-Image , which is also an autoregressive model. It also has faster processing speeds than the diffusion models Qwen-Image and Z-Image .
There is also
Anime-style images can also be generated.
The sample also included an image of Doraemon.
A demo app that can generate images using BitDance has also been released, so I will try using it. First, click the link below.
BitDance-14B-64x - a Hugging Face Space by shallowdream204
https://huggingface.co/spaces/shallowdream204/BitDance-14B-64x
This time, in order to generate a 'life-like image of a maid making a peace sign in a cafe,' I put together the following prompt using the examples on the gallery page as a reference.
A high-resolution ultra-detailed photorealistic portrait of a young East Asian girl indoors, with fair smooth skin and a natural soft glow, large round dark brown eyes with clear reflections, subtle under-eye softness, a small delicate nose, light pink innocent slightly glossy lips, and a gentle expression, featuring a tiny beauty mark on her cheek for realism. She has dark brown shiny hair styled into two long thick braided pigtails falling over her shoulders, with slightly loose braids showing individual strands, wispy straight bangs softly covering her forehead, and a few natural flyaway hairs. She is wearing a maid clothes. Her pose is playful and casual, head slightly tilted toward the camera, leaning forward, both hands raised near the frame making a peace sign, creating a friendly and intimate feeling. The background is a cozy modern cafeteria with a wooden table and chair, softly blurred with shallow depth of field. Soft warm indoor lighting evenly illuminates her face, no harsh shadows, highlighting skin texture and hair shine. Shot with a high-end DSLR or mirrorless camera, 50mm lens, f/1.8, cinematic bokeh, sharp focus on the face, natural color grading, high dynamic range, realistic proportions, Japanese/Korean portrait photography aesthetic, cozy winter vibe, candid snapshot feeling, extremely detailed, 8k quality.
Fill in the prompts and click 'Generate'.
A 1024 x 1024 pixel image was generated in about 30 seconds.
The generated image is below. Click to see the original image before it was reduced.
As shown above, BitDance can generate images according to the prompts in English, but it does not support Japanese. Even when I entered the prompt 'A picture of a bear eating an apple,' a completely different image was generated. It seems that BitDance specializes in English and Chinese.
BitDance model data is available at the following link:
BitDance - a shallowdream204 Collection
https://huggingface.co/collections/shallowdream204/bitdance
The related code is also available on GitHub.
GitHub - shallowdream204/BitDance: BitDance: Open-source autoregressive model with binary visual tokens. A research project for building powerful multimodal autoregressive model.
https://github.com/shallowdream204/BitDance?tab=readme-ov-file
Related Posts:
