The first model for automatic text-to-video prompt completion: Given a few words as input, the model will generate a few whole text-to-video prompts.

Details

It is fine-tuned on the VidProM dataset using Mistral-7B-v0.1 and 8 A100 GPUs.

Usage

Download the model

from transformers import pipeline
import torch
pipe = pipeline("text-generation", model="WenhaoWang/AutoT2VPrompt", model_kwargs={"torch_dtype": torch.bfloat16}, device_map="cuda:0")

Set the Parameters

input = "An underwater world"      # The input text to generate text-to-video prompt.
max_length = 50                    # The maximum length of the generated text.
temperature = 1.2                  # Controls the randomness of the generation. Higher values lead to more random outputs.
top_k = 8                          # Limits the number of words considered at each step to the top k most likely words.
num_return_sequences = 10          # The number of different text-to-video prompts to generate from the same input.

Generation

all_prompts = pipe(input, max_length = max_length, do_sample = True, temperature = temperature, top_k = top_k, num_return_sequences=num_return_sequences)

def process(text):
    text = text.replace('\n', '.')
    text = text.replace('  .', '.')
    text = text[:text.rfind('.')]
    text = text + '.'
    return text

for i in range(num_return_sequences):
    print(process(all_prompts[i]['generated_text']))

You will get 10 text-to-video prompts, and you can pick one you like most.

An underwater world, 25 ye boy, with aqua-green eyes, dk sandy blond hair, from the back, and on his back a fish, 23 ye old, weing glasses,ctoon chacte.
An underwater world, the video should capture the essence of tranquility and the beauty of nature.. a woman with short hair weing a green dress sitting at the desk.
An underwater world, the ocean is full of discded items, the water flows, and the light penetrating through the water.
An underwater world.. a woman with red eyes and red lips  is looking forwd.
An underwater world.. an old man sitting in a chair, smoking a pipe, a little smoke coming out of the chair, a man is drinking a glass.
An underwater world. The ocean is filled with bioluminess as the water reflects a soft glow from a bioluminescent phosphorescent light source. The camera slowly moves away and zooms in..
An underwater world. the girl looks at the camera and smiles with happiness..
An underwater world, 1960s horror film..
An underwater world.. 4 men in 1940s style clothes walk ound a gothic castle. night, fe. A girl is running, and there e some flowers along the river.
An underwater world,  -camera pan up . A girl is playing with her cat on a sunny day in the pk. A man is running and then falling down and dying.

License

The model is licensed under the CC BY-NC 4.0 license.

Citation

@article{wang2024vidprom,
  title={VidProM: A Million-scale Real Prompt-Gallery Dataset for Text-to-Video Diffusion Models},
  author={Wang, Wenhao and Yang, Yi},
  journal={arXiv preprint arXiv:2403.06098},
  year={2024}
}

Acknowledgment

The fine-tuning process is helped by Yaowei Zheng.

Contact

If you have any questions, feel free to contact Wenhao Wang ([email protected]).

Downloads last month
38
Safetensors
Model size
7.24B params
Tensor type
BF16
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train WenhaoWang/AutoT2VPrompt