NOVA

Haoge Deng1,4*, Ting Pan2,4*, Haiwen Diao3,4*, Zhengxiong Luo4*, Yufeng Cui4 Huchuan Lu3, Shiguang Shan2, Yonggang Qi1†, Xinlong Wang4†

BUPT1, ICT-CAS2, DLUT3, BAAI4 * Equal Contribution, † Corresponding Author

We present NOVA (NOn-Quantized Video Autoregressive Model), a model that enables autoregressive image/video generation with high efficiency. NOVA reformulates the video generation problem as non-quantized autoregressive modeling of temporal frame-by-frame prediction and spatial set-by-set prediction. NOVA generalizes well and enables diverse zero-shot generation abilities in one unified model.

🚀News

• [Oct 2025] Released our next video generation model 🐻 URSA.
• [Jul 2025] Codebase refactor with Accelerate, OmegaConf and Wandb.
• [Feb 2025] Released Evaluation Guide.
• [Feb 2025] Released Training Guide
• [Jan 2025] Accepted by ICLR 2025. [OpenReview] & [Poster].
• [Dec 2024] Released Project Page
• [Dec 2024] Released 🤗 Online Demo (T2I, T2V)
• [Dec 2024] Released paper, weights, and Quick Start guide and Gradio Demo local code .

✨Hightlights

• 🔥 Novel Approach: Non-quantized video autoregressive generation.
• 🔥 State-of-the-art Performance: High efficiency with state-of-the-art t2i/t2v results.
• 🔥 Unified Modeling: Multi-task capabilities in a single unified model.

🗄️Model Zoo

See detailed description in Model Zoo

Text to Image

Model	Parameters	Resolution	Data	Weight	GenEval	DPGBench
NOVA-0.6B	0.6B	512x512	16M	🤗 HF link	0.75	81.76
NOVA-0.3B	0.3B	1024x1024	600M	🤗 HF link	0.67	80.60
NOVA-0.6B	0.6B	1024x1024	600M	🤗 HF link	0.69	82.25
NOVA-1.4B	1.4B	1024x1024	600M	🤗 HF link	0.71	83.01

Text to Video

Model	Parameters	Resolution	Data	Weight	VBench
NOVA-0.6B	0.6B	33x768x480	20M	🤗 HF link	80.12

📖Table of Contents

• 1. Installation
  • 1.1 From Source
  • 1.2 From Git
• 2. Quick Start
  • 2.1 Text to Image
  • 2.2 Text to Video
  • 2.3 Image to Video
• 3. Gradio Demo
• 4. Train
• 5. Inference
• 6. Evaluation

1. Installation

1.1 From Source

Clone this repository to local disk and install:

pip install diffusers transformers accelerate imageio-ffmpeg omegaconf wandb
git clone https://github.com/baaivision/NOVA.git
cd NOVA && pip install .

1.2 From Git

You can also install from the remote repository if you have set your Github SSH key:

pip install diffusers transformers accelerate imageio-ffmpeg omegaconf wandb
pip install git+ssh://[email protected]/baaivision/NOVA.git

2. Quick Start

2.1 Text to Image

import torch
from diffnext.pipelines import NOVAPipeline

model_id = "BAAI/nova-d48w768-sdxl1024"
model_args = {"torch_dtype": torch.float16, "trust_remote_code": True}
pipe = NOVAPipeline.from_pretrained(model_id, **model_args)
pipe = pipe.to("cuda")

prompt = "a shiba inu wearing a beret and black turtleneck."
image = pipe(prompt).images[0]
    
image.save("shiba_inu.jpg")

2.2 Text to Video

import os
import torch
from diffnext.pipelines import NOVAPipeline
from diffnext.utils import export_to_image, export_to_video
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

model_id = "BAAI/nova-d48w1024-osp480"
low_memory = False

model_args = {"torch_dtype": torch.float16, "trust_remote_code": True}
pipe = NOVAPipeline.from_pretrained(model_id, **model_args)

if low_memory:
    # Use CPU model offload routine and expandable allocator if OOM.
    pipe.enable_model_cpu_offload()
else:
    pipe = pipe.to("cuda")

# Text to Video
prompt = "Many spotted jellyfish pulsating under water."
video = pipe(prompt, max_latent_length=9).frames[0]
export_to_video(video, "jellyfish.mp4", fps=12)

# Increase AR and diffusion steps for better video quality.
video = pipe(
  prompt,
  max_latent_length=9,
  num_inference_steps=128,  # default: 64
  num_diffusion_steps=100,  # default: 25
).frames[0]
export_to_video(video, "jellyfish_v2.mp4", fps=12)

# You can also generate images from text, with the first frame as an image.
prompt = "Many spotted jellyfish pulsating under water."
image = pipe(prompt, max_latent_length=1).frames[0, 0]
export_to_image(image, "jellyfish.jpg")

2.3 Image to Video

import os, torch, PIL.Image, numpy as np
from diffnext.pipelines import NOVAPipeline
from diffnext.utils import export_to_image, export_to_video
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

model_id = "BAAI/nova-d48w1024-osp480"
low_memory = False

model_args = {"torch_dtype": torch.float16, "trust_remote_code": True}
pipe = NOVAPipeline.from_pretrained(model_id, **model_args)

if low_memory:
    # Use CPU model offload routine and expandable allocator if OOM.
    pipe.enable_model_cpu_offload()
else:
    pipe = pipe.to("cuda")

prompt = "Many spotted jellyfish pulsating under water."

# Step1: Generate or select an image that matches the resolution 768x480.
image = pipe(prompt, max_latent_length=1).frames[0, 0]
export_to_image(image, "jellyfish.jpg")

# Step2: Use this image to generate subsequent frames.
video = pipe(prompt, image=np.array(PIL.Image.open("jellyfish.jpg")), max_latent_length=9).frames[0]
export_to_video(video, "jellyfish.mp4", fps=12)

3. Gradio Demo

# For text-to-image demo
python scripts/app_nova_t2i.py --model "BAAI/nova-d48w1024-sdxl1024" --device 0

# For text-to-video demo
python scripts/app_nova_t2v.py --model "BAAI/nova-d48w1024-osp480" --device 0

4. Train

• See Training Guide

5. Evaluation

• See Evaluation Guide

6. Inference

• See Inference Guide

📋Todo List

• Model zoo
• Quick Start
• Gradio Demo
• Training guide
• Evaluation guide
• Inference guide
• Prompt Writer
• Larger model size
• Additional downstream tasks: Image editing, Video editing, Controllable generation

Citation

If you find this repository useful, please consider giving a star ⭐ and citation 🦖:

@article{deng2025ursa,
  title={Uniform Discrete Diffusion with Metric Path for Video Generation},
  author={Deng, Haoge and Pan, Ting and Zhang, Fan and Liu, Yang and Luo, Zhuoyan and Cui, Yufeng and Shen, Chunhua and Shan, Shiguang and Zhang, Zhaoxiang and Wang, Xinlong},
  journal={arXiv preprint arXiv:2510.24717},
  year={2025}
}

@article{deng2024nova,
  title={Autoregressive Video Generation without Vector Quantization},
  author={Deng, Haoge and Pan, Ting and Diao, Haiwen and Luo, Zhengxiong and Cui, Yufeng and Lu, Huchuan and Shan, Shiguang and Qi, Yonggang and Wang, Xinlong},
  journal={arXiv preprint arXiv:2412.14169},
  year={2024}
}

Acknowledgement

We thank the repositories: MAE, MAR, MaskGIT, DiT, Open-Sora-Plan, CogVideo, FLUX, OpenMuse and CodeWithGPU.

License

Code and models are licensed under Apache License 2.0.