CCCCCC/VPO-5B

VPO: Aligning Text-to-Video Generation Models with Prompt Optimization

• Repository: https://github.com/thu-coai/VPO
• Paper: VPO: Aligning Text-to-Video Generation Models with Prompt Optimization
• Data: https://huggingface.co/datasets/CCCCCC/VPO

VPO

VPO is a principled prompt optimization framework grounded in the principles of harmlessness, accuracy, and helpfulness. VPO employs a two-stage process that first constructs a supervised fine-tuning dataset guided by safety and alignment, and then conducts preference learning with both text-level and video-level feedback. As a result, VPO preserves user intent while enhancing video quality and safety.

Model Details

Video Generation Model

This model is trained to optimize user prompt for CogVideoX-5B. VPO-2B is for CogVideoX-2B.

Data

Our dataset can be found here.

Language

English

Intended Use

Prompt Template

We adopt a prompt template as

In this task, your goal is to expand the user's short query into a detailed and well-structured English prompt for generating short videos.

Please ensure that the generated video prompt adheres to the following principles:

1. **Harmless**: The prompt must be safe, respectful, and free from any harmful, offensive, or unethical content.  
2. **Aligned**: The prompt should fully preserve the user's intent, incorporating all relevant details from the original query while ensuring clarity and coherence.  
3. **Helpful for High-Quality Video Generation**: The prompt should be descriptive and vivid to facilitate high-quality video creation. Keep the scene feasible and well-suited for a brief duration, avoiding unnecessary complexity or unrealistic elements not mentioned in the query.

User Query:{user prompt}

Video Prompt:

Inference code

Here is an example code for inference:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = ''

prompt_template = """In this task, your goal is to expand the user's short query into a detailed and well-structured English prompt for generating short videos.

Please ensure that the generated video prompt adheres to the following principles:

1. **Harmless**: The prompt must be safe, respectful, and free from any harmful, offensive, or unethical content.  
2. **Aligned**: The prompt should fully preserve the user's intent, incorporating all relevant details from the original query while ensuring clarity and coherence.  
3. **Helpful for High-Quality Video Generation**: The prompt should be descriptive and vivid to facilitate high-quality video creation. Keep the scene feasible and well-suited for a brief duration, avoiding unnecessary complexity or unrealistic elements not mentioned in the query.

User Query:{}

Video Prompt:"""

device = 'cuda:0'
model = AutoModelForCausalLM.from_pretrained(model_path).half().eval().to(device)
# for 8bit
# model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device, load_in_8bit=True)
tokenizer = AutoTokenizer.from_pretrained(model_path)

text = "a cute dog on the grass"
messgae = [{'role': 'user', 'content': prompt_template.format(text)}]

model_inputs = tokenizer.apply_chat_template(messgae, add_generation_prompt=True, tokenize=True, return_tensors="pt").to(device)
output = model.generate(model_inputs, max_new_tokens=1024, do_sample=True, top_p=1.0, temperature=0.7, num_beams=1)
resp = tokenizer.decode(output[0]).split('<|start_header_id|>assistant<|end_header_id|>')[1].split('<|eot_id|>')[0].strip()

print(resp)

See our Github Repo for more detailed usage (e.g. Inference with Vllm).