cjwbw / kandinskyvideo

text-to-video generation model

Cold

Public
1.2K runs
A100 (80GB)
GitHub
Paper
License

Run with an API

Playground API Examples README Versions

Input

prompt

string

Shift + Return to add a new line

a red car is drifting on the mountain road, close view, fast movementa red car is drifting on the mountain road, close view, fast movement

Input prompt.

Default: "a red car is drifting on the mountain road, close view, fast movement"

negative_prompt

string

Shift + Return to add a new line

Specify things to not see in the output

width

integer

Width of output video. Lower the setting if out of memory.

Default: 640

height

integer

Height of output video. Lower the setting if out of memory.

Default: 384

num_inference_steps

integer

Number of denoising steps

Default: 50

guidance_scale

number

Scale for classifier-free guidance

Default: 5

interpolation_guidance_scale

number

Scale for interpolation guidance

Default: 0.25

interpolation_level

string

An enumeration.

Default: "low"

fps

integer

fps for the output video.

Default: 10

Run this model in Node.js with one line of code:

npx create-replicate --model=cjwbw/kandinskyvideo

or set up a project from scratch

Install Replicate’s Node.js client library:

npm install replicate

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Import and set up the client:

import Replicate from "replicate";

const replicate = new Replicate({
  auth: process.env.REPLICATE_API_TOKEN,
});

Run cjwbw/kandinskyvideo using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

const output = await replicate.run(
  "cjwbw/kandinskyvideo:849b70f3e300a650aa8b78d0f8f24d104824b832ea7f61c79bd2c7e78a4ad545",
  {
    input: {
      fps: 10,
      width: 640,
      height: 384,
      prompt: "a red car is drifting on the mountain road, close view, fast movement",
      guidance_scale: 5,
      interpolation_level: "low",
      num_inference_steps: 50,
      interpolation_guidance_scale: 0.25
    }
  }
);

// To access the file URL:
console.log(output.url()); //=> "http://example.com"

// To write the file to disk:
fs.writeFile("my-image.png", output);

To learn more, take a look at the guide on getting started with Node.js.

Install Replicate’s Python client library:

pip install replicate

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Import the client:

import replicate

Run cjwbw/kandinskyvideo using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

output = replicate.run(
    "cjwbw/kandinskyvideo:849b70f3e300a650aa8b78d0f8f24d104824b832ea7f61c79bd2c7e78a4ad545",
    input={
        "fps": 10,
        "width": 640,
        "height": 384,
        "prompt": "a red car is drifting on the mountain road, close view, fast movement",
        "guidance_scale": 5,
        "interpolation_level": "low",
        "num_inference_steps": 50,
        "interpolation_guidance_scale": 0.25
    }
)
print(output)

To learn more, take a look at the guide on getting started with Python.

Set the REPLICATE_API_TOKEN environment variable:

export REPLICATE_API_TOKEN=<paste-your-token-here>

Find your API token in your account settings.

Run cjwbw/kandinskyvideo using Replicate’s API. Check out the model's schema for an overview of inputs and outputs.

curl -s -X POST \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Content-Type: application/json" \
  -H "Prefer: wait" \
  -d $'{
    "version": "cjwbw/kandinskyvideo:849b70f3e300a650aa8b78d0f8f24d104824b832ea7f61c79bd2c7e78a4ad545",
    "input": {
      "fps": 10,
      "width": 640,
      "height": 384,
      "prompt": "a red car is drifting on the mountain road, close view, fast movement",
      "guidance_scale": 5,
      "interpolation_level": "low",
      "num_inference_steps": 50,
      "interpolation_guidance_scale": 0.25
    }
  }' \
  https://api.replicate.com/v1/predictions

To learn more, take a look at Replicate’s HTTP API reference docs.

Output

{
  "completed_at": "2023-11-25T17:30:06.441672Z",
  "created_at": "2023-11-25T17:17:32.380766Z",
  "data_removed": false,
  "error": null,
  "id": "jnvfjstbxqlangdvnwz3pqlh6u",
  "input": {
    "fps": 10,
    "width": 640,
    "height": 384,
    "prompt": "a red car is drifting on the mountain road, close view, fast movement",
    "guidance_scale": 5,
    "interpolation_level": "low",
    "num_inference_steps": 50,
    "interpolation_guidance_scale": 0.25
  },
  "logs": "0%|          | 0/50 [00:00<?, ?it/s]\n  2%|▏         | 1/50 [00:00<00:31,  1.57it/s]\n  4%|▍         | 2/50 [00:01<00:26,  1.81it/s]\n  6%|▌         | 3/50 [00:01<00:24,  1.91it/s]\n  8%|▊         | 4/50 [00:02<00:23,  1.96it/s]\n 10%|█         | 5/50 [00:02<00:22,  1.99it/s]\n 12%|█▏        | 6/50 [00:03<00:21,  2.01it/s]\n 14%|█▍        | 7/50 [00:03<00:21,  2.02it/s]\n 16%|█▌        | 8/50 [00:04<00:20,  2.03it/s]\n 18%|█▊        | 9/50 [00:04<00:20,  2.04it/s]\n 20%|██        | 10/50 [00:05<00:19,  2.04it/s]\n 22%|██▏       | 11/50 [00:05<00:19,  2.03it/s]\n 24%|██▍       | 12/50 [00:06<00:18,  2.04it/s]\n 26%|██▌       | 13/50 [00:06<00:18,  2.04it/s]\n 28%|██▊       | 14/50 [00:06<00:17,  2.04it/s]\n 30%|███       | 15/50 [00:07<00:17,  2.05it/s]\n 32%|███▏      | 16/50 [00:07<00:16,  2.05it/s]\n 34%|███▍      | 17/50 [00:08<00:16,  2.05it/s]\n 36%|███▌      | 18/50 [00:08<00:15,  2.05it/s]\n 38%|███▊      | 19/50 [00:09<00:15,  2.04it/s]\n 40%|████      | 20/50 [00:09<00:14,  2.04it/s]\n 42%|████▏     | 21/50 [00:10<00:14,  2.04it/s]\n 44%|████▍     | 22/50 [00:10<00:13,  2.05it/s]\n 46%|████▌     | 23/50 [00:11<00:13,  2.04it/s]\n 48%|████▊     | 24/50 [00:11<00:12,  2.03it/s]\n 50%|█████     | 25/50 [00:12<00:12,  2.03it/s]\n 52%|█████▏    | 26/50 [00:12<00:11,  2.04it/s]\n 54%|█████▍    | 27/50 [00:13<00:11,  2.04it/s]\n 56%|█████▌    | 28/50 [00:13<00:10,  2.04it/s]\n 58%|█████▊    | 29/50 [00:14<00:10,  2.04it/s]\n 60%|██████    | 30/50 [00:14<00:09,  2.04it/s]\n 62%|██████▏   | 31/50 [00:15<00:09,  2.04it/s]\n 64%|██████▍   | 32/50 [00:15<00:08,  2.05it/s]\n 66%|██████▌   | 33/50 [00:16<00:08,  2.05it/s]\n 68%|██████▊   | 34/50 [00:16<00:07,  2.05it/s]\n 70%|███████   | 35/50 [00:17<00:07,  2.05it/s]\n 72%|███████▏  | 36/50 [00:17<00:06,  2.05it/s]\n 74%|███████▍  | 37/50 [00:18<00:06,  2.05it/s]\n 76%|███████▌  | 38/50 [00:18<00:05,  2.05it/s]\n 78%|███████▊  | 39/50 [00:19<00:05,  2.05it/s]\n 80%|████████  | 40/50 [00:19<00:04,  2.05it/s]\n 82%|████████▏ | 41/50 [00:20<00:04,  2.05it/s]\n 84%|████████▍ | 42/50 [00:20<00:03,  2.05it/s]\n 86%|████████▌ | 43/50 [00:21<00:03,  2.05it/s]\n 88%|████████▊ | 44/50 [00:21<00:02,  2.05it/s]\n 90%|█████████ | 45/50 [00:22<00:02,  2.05it/s]\n 92%|█████████▏| 46/50 [00:22<00:01,  2.05it/s]\n 94%|█████████▍| 47/50 [00:23<00:01,  2.05it/s]\n 96%|█████████▌| 48/50 [00:23<00:00,  2.05it/s]\n 98%|█████████▊| 49/50 [00:24<00:00,  2.05it/s]\n100%|██████████| 50/50 [00:24<00:00,  2.05it/s]\n100%|██████████| 50/50 [00:24<00:00,  2.03it/s]",
  "metrics": {
    "predict_time": 28.07745,
    "total_time": 754.060906
  },
  "output": "https://replicate.delivery/pbxt/ESLBl1hGLVZiJ5NEMoksjLY0Rf9CG3Xwb5jZBGrn8hoOCf7RA/output.mp4",
  "started_at": "2023-11-25T17:29:38.364222Z",
  "status": "succeeded",
  "urls": {
    "get": "https://api.replicate.com/v1/predictions/jnvfjstbxqlangdvnwz3pqlh6u",
    "cancel": "https://api.replicate.com/v1/predictions/jnvfjstbxqlangdvnwz3pqlh6u/cancel"
  },
  "version": "849b70f3e300a650aa8b78d0f8f24d104824b832ea7f61c79bd2c7e78a4ad545"
}

Generated in

28.1 seconds

Tweak itReport

0%|          | 0/50 [00:00<?, ?it/s]
  2%|▏         | 1/50 [00:00<00:31,  1.57it/s]
  4%|▍         | 2/50 [00:01<00:26,  1.81it/s]
  6%|▌         | 3/50 [00:01<00:24,  1.91it/s]
  8%|▊         | 4/50 [00:02<00:23,  1.96it/s]
 10%|█         | 5/50 [00:02<00:22,  1.99it/s]
 12%|█▏        | 6/50 [00:03<00:21,  2.01it/s]
 14%|█▍        | 7/50 [00:03<00:21,  2.02it/s]
 16%|█▌        | 8/50 [00:04<00:20,  2.03it/s]
 18%|█▊        | 9/50 [00:04<00:20,  2.04it/s]
 20%|██        | 10/50 [00:05<00:19,  2.04it/s]
 22%|██▏       | 11/50 [00:05<00:19,  2.03it/s]
 24%|██▍       | 12/50 [00:06<00:18,  2.04it/s]
 26%|██▌       | 13/50 [00:06<00:18,  2.04it/s]
 28%|██▊       | 14/50 [00:06<00:17,  2.04it/s]
 30%|███       | 15/50 [00:07<00:17,  2.05it/s]
 32%|███▏      | 16/50 [00:07<00:16,  2.05it/s]
 34%|███▍      | 17/50 [00:08<00:16,  2.05it/s]
 36%|███▌      | 18/50 [00:08<00:15,  2.05it/s]
 38%|███▊      | 19/50 [00:09<00:15,  2.04it/s]
 40%|████      | 20/50 [00:09<00:14,  2.04it/s]
 42%|████▏     | 21/50 [00:10<00:14,  2.04it/s]
 44%|████▍     | 22/50 [00:10<00:13,  2.05it/s]
 46%|████▌     | 23/50 [00:11<00:13,  2.04it/s]
 48%|████▊     | 24/50 [00:11<00:12,  2.03it/s]
 50%|█████     | 25/50 [00:12<00:12,  2.03it/s]
 52%|█████▏    | 26/50 [00:12<00:11,  2.04it/s]
 54%|█████▍    | 27/50 [00:13<00:11,  2.04it/s]
 56%|█████▌    | 28/50 [00:13<00:10,  2.04it/s]
 58%|█████▊    | 29/50 [00:14<00:10,  2.04it/s]
 60%|██████    | 30/50 [00:14<00:09,  2.04it/s]
 62%|██████▏   | 31/50 [00:15<00:09,  2.04it/s]
 64%|██████▍   | 32/50 [00:15<00:08,  2.05it/s]
 66%|██████▌   | 33/50 [00:16<00:08,  2.05it/s]
 68%|██████▊   | 34/50 [00:16<00:07,  2.05it/s]
 70%|███████   | 35/50 [00:17<00:07,  2.05it/s]
 72%|███████▏  | 36/50 [00:17<00:06,  2.05it/s]
 74%|███████▍  | 37/50 [00:18<00:06,  2.05it/s]
 76%|███████▌  | 38/50 [00:18<00:05,  2.05it/s]
 78%|███████▊  | 39/50 [00:19<00:05,  2.05it/s]
 80%|████████  | 40/50 [00:19<00:04,  2.05it/s]
 82%|████████▏ | 41/50 [00:20<00:04,  2.05it/s]
 84%|████████▍ | 42/50 [00:20<00:03,  2.05it/s]
 86%|████████▌ | 43/50 [00:21<00:03,  2.05it/s]
 88%|████████▊ | 44/50 [00:21<00:02,  2.05it/s]
 90%|█████████ | 45/50 [00:22<00:02,  2.05it/s]
 92%|█████████▏| 46/50 [00:22<00:01,  2.05it/s]
 94%|█████████▍| 47/50 [00:23<00:01,  2.05it/s]
 96%|█████████▌| 48/50 [00:23<00:00,  2.05it/s]
 98%|█████████▊| 49/50 [00:24<00:00,  2.05it/s]
100%|██████████| 50/50 [00:24<00:00,  2.05it/s]
100%|██████████| 50/50 [00:24<00:00,  2.03it/s]

Examples

View more examples

Run time and cost

This model costs approximately $0.23 to run on Replicate, or 4 runs per $1, but this varies depending on your inputs. It is also open source and you can run it on your own computer with Docker.

This model runs on Nvidia A100 (80GB) GPU hardware. Predictions typically complete within 3 minutes. The predict time for this model varies significantly based on the inputs.

Readme

Kandinsky Video — a new text-to-video generation model

SoTA quality among open-source solutions

Kandinsky Video is a text-to-video generation model, which is based on the FusionFrames architecture, consisting of two main stages: keyframe generation and interpolation. Our approach for temporal conditioning allows us to generate videos with high-quality appearance, smoothness and dynamics.

Pipeline

The encoded text prompt enters the U-Net keyframe generation model with temporal layers or blocks, and then the sampled latent keyframes are sent to the latent interpolation model in such a way as to predict three interpolation frames between two keyframes. A temporal MoVQ-GAN decoder is used to get the final video result.

Architecture details

Text encoder (Flan-UL2) - 8.6B
Latent Diffusion U-Net3D - 4.0B
MoVQ encoder/decoder - 256M

BibTeX

If you use our work in your research, please cite our publication:

@article{arkhipkin2023fusionframes,
  title     = {FusionFrames: Efficient Architectural Aspects for Text-to-Video Generation Pipeline},
  author    = {Arkhipkin, Vladimir and Shaheen, Zein and Vasilev, Viacheslav and Dakhova, Elizaveta and Kuznetsov, Andrey and Dimitrov, Denis},
  journal   = {arXiv preprint arXiv:2311.13073},
  year      = {2023}, 
}