| --- |
| title: BitDance-14B-64x |
| emoji: 🚀 |
| colorFrom: red |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: 6.5.1 |
| app_file: app.py |
| pinned: false |
| license: apache-2.0 |
| short_description: Open-source autoregressive model with binary visual tokens. |
| --- |
| |
| # 🚀 BitDance-14B-64x |
|
|
| BitDance is a scalable autoregressive (AR) foundation model with **14 billion parameters**. It introduces a novel approach to image generation by predicting **binary visual tokens** instead of standard codebook indices. |
|
|
|
|
|
|
| ## 🌟 Key Features |
| - **Binary Visual Tokenizer:** Scales token entropy to $2^{256}$ states, providing a highly expressive yet compact discrete representation. |
| - **Binary Diffusion Head:** Replaces standard categorical classification with continuous-space diffusion for high-precision sampling in massive discrete spaces. |
| - **Next-Patch Diffusion:** A parallel decoding paradigm that predicts up to **64 tokens per step**, achieving a 30x speedup over traditional AR models for 1024x1024 resolution. |
| - **Multimodal Foundation:** Trained on large-scale multimodal data, excelling in prompt adherence, spatial reasoning, and high-fidelity photorealistic rendering. |
|
|
| ## 🛠️ Performance |
| | Model | Tokens/Step | Speedup (vs. standard AR) | Target Resolution | |
| | :--- | :--- | :--- | :--- | |
| | BitDance-14B-16x | 16 | ~8x | 512px & 1024px | |
| | **BitDance-14B-64x** | **64** | **~30x** | **1024px** | |
|
|
|
|
|
|
| ## 🚀 Quick Start (Local Setup) |
|
|
| If you wish to run the model locally using the `diffusers` library: |
|
|
| ```python |
| import torch |
| from diffusers import DiffusionPipeline |
| |
| pipe = DiffusionPipeline.from_pretrained( |
| "shallowdream204/BitDance-14B-64x", |
| custom_pipeline="shallowdream204/BitDance-14B-64x", |
| torch_dtype=torch.bfloat16 |
| ).to("cuda") |
| |
| prompt = "A cinematic portrait of a futuristic explorer in a neon-lit cyberpunk city, ultra-detailed, 8k." |
| image = pipe(prompt=prompt, height=1024, width=1024).images[0] |
| image.save("output.png") |
