---
license: apache-2.0
---


# <span style="font-family: 'Courier New', monospace; font-weight: bold">T2V-Turbo</span>: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback

## 4-step Text-to-video Generation

<table class="center">
  <td>
  <video poster="" autoplay controls muted loop height="150%" playbackRate=2.0><source src="https://rg-lcd.s3.amazonaws.com/t2v-turbo-demo/t2v-turbo-vc2/4steps/0273.mp4" type="video/mp4"></video>
  </td>
  <td>
  <video poster="" autoplay controls muted loop height="150%" playbackRate=2.0><source src="https://rg-lcd.s3.amazonaws.com/t2v-turbo-demo/t2v-turbo-vc2/4steps/0054.mp4" type="video/mp4"></video>
  </td>
  <td>
  <video poster="" autoplay controls muted loop height="150%" playbackRate=2.0><source src="https://rg-lcd.s3.amazonaws.com/t2v-turbo-demo/t2v-turbo-vc2/4steps/0262.mp4" type="video/mp4"></video>
  </td>

  <tr>
  <td style="text-align:center;" width="320">With the style of low-poly game art, A majestic, white horse gallops gracefully across a moonlit beach.</td>
  <td style="text-align:center;" width="320">medium shot of Christine, a beautiful 25-year-old brunette resembling Selena Gomez, anxiously looking up as she walks down a New York street, cinematic style</td>
  <td style="text-align:center;" width="320">a cartoon pig playing his guitar, Andrew Warhol style</td>
  <tr>
</table >

<table class="center">
  <td>
  <video poster="" autoplay controls muted loop height="150%" playbackRate=2.0><source src="https://rg-lcd.s3.amazonaws.com/t2v-turbo-demo/t2v-turbo-vc2/4steps/0023.mp4" type="video/mp4"></video>
  </td>
  <td>
  <video poster="" autoplay controls muted loop height="150%" playbackRate=2.0><source src="https://rg-lcd.s3.amazonaws.com/t2v-turbo-demo/t2v-turbo-vc2/4steps/0021.mp4" type="video/mp4"></video>
  </td>
  <td>
  <video poster="" autoplay controls muted loop height="150%" playbackRate=2.0><source src="https://rg-lcd.s3.amazonaws.com/t2v-turbo-demo/t2v-turbo-vc2/4steps/0064.mp4" type="video/mp4"></video>
  </td>

  <tr>
  <td style="text-align:center;" width="320">a dog wearing vr goggles on a boat</td>
  <td style="text-align:center;" width="320">Pikachu snowboarding</td>
  <td style="text-align:center;" width="320">a girl floating underwater </td>
  <tr>
</table >

## Model description 🚀

This repository contains `unet_lora.pt` that can turn [VideoCrafter2](https://ailab-cvc.github.io/videocrafter2/) into our <span style="font-family: 'Courier New', monospace; font-weight: bold">T2V-Turbo</span> (VC2). Our <span style="font-family: 'Courier New', monospace; font-weight: bold">T2V-Turbo</span> (VC2) can achieve both fast and high-quality T2V generation. On [VBench](https://vchitect.github.io/VBench-project/), the 4-step generation from our <span style="font-family: 'Courier New', monospace; font-weight: bold">T2V-Turbo</span> (VC2) even outperform proprietary systems, including [Gen-2](https://research.runwayml.com/gen2) and [Pika](https://pika.art/). Please refer to our [GitHub repo](https://github.com/Ji4chenLi/t2v-turbo) for detailed instructions.

## Usage 👓

*This checkpoint is obtained by merging the UNet [LoRA weight](https://huggingface.co/jiachenli-ucsb/T2V-Turbo-VC2) to the UNet of [VideoCrafter2](https://huggingface.co/VideoCrafter/VideoCrafter2/). Therefore, the checkpoint here is also under the apache-2.0 license.*

You need to first clone our [GitHub repo](https://github.com/Ji4chenLi/t2v-turbo). Here are the codes to load the checkpoint.
```py
from utils.common_utils import load_model_checkpoint
from utils.utils import instantiate_from_config


config = OmegaConf.load("configs/inference_t2v_512_v2.0.yaml")
model_config = config.pop("model", OmegaConf.create())
pretrained_t2v = instantiate_from_config(model_config)

unet_config = model_config["params"]["unet_config"]
unet_config["params"]["time_cond_proj_dim"] = 256
unet = instantiate_from_config(unet_config)
pretrained_t2v.model.diffusion_model = unet
pretrained_t2v = load_model_checkpoint(pretrained_t2v, "checkpoints/t2v_turbo_vc2.pt")

```


## Misuse, Malicious Use and Excessive Use 📖
Our model is meant for research purposes.

- It is prohibited to generate content that is demeaning or harmful to people or their environment, culture, religion, etc.
- Prohibited for pornographic, violent and bloody content generation.
- Prohibited for error and false information generation.