Describe a shot.
Watch it render.
Type a scene and LTX-Video, a 13B-parameter video diffusion model, turns it into a four-second clip in about 8 seconds. It runs on a single GPU that Chalk deploys from one Python file.
Direct a shot
Describe the subject, the motion, the light and the camera, or start from one of the suggestions. One GPU renders one clip at a time, so if others are ahead of you, you will see your place in the queue.
How it works
A video diffusion model starts from pure noise and removes a little of it at every step, guided by your words, until a coherent clip is left.
Describe
Write a prompt like a shot list: the subject, how it moves, the light, and the camera. A T5 text encoder turns it into the embedding the video model is conditioned on.
Draft
LTX-Video's 13B distilled transformer denoises 97 frames of latent video in 8 steps at a reduced resolution, which settles composition and motion quickly.
Refine
A latent upsampler doubles the resolution, four more steps restore detail, and the VAE decodes 768×512 frames that are encoded to a 24 fps H.264 clip you can play in the browser.
Deploy your own in three commands
Clone the example, log in, and run the deploy script. Chalk builds the image, provisions an L40S, and the replica pulls the model weights as it boots. You get a URL that serves this exact demo.
The same replica exposes a small JSON API: submit a prompt, poll the job, download the MP4. Wire it into a product, a batch job, or an agent that storyboards its own shots.
chalk-ai/text-to-videogit clone https://github.com/chalk-ai/text-to-video
cd text-to-video
# Or export CHALK_CLIENT_ID and CHALK_CLIENT_SECRET.
chalk login
# Builds the image, provisions an L40S, prints the URL of your live demo.
./deploy.py --gpu nvidia-l40sWhat Chalk handles
The model is the interesting part. Everything around it is the part Chalk does for you.
The right GPU for the model
Ask for gpu="nvidia-l40s" and Chalk provisions a 48 GB card, schedules the replica, and gives it an HTTPS URL.
Images from Python
Image.debian_slim().pip_install(...) builds the container with torch, diffusers and ffmpeg. No Dockerfile, no registry to manage.
Fast enough to iterate
A 4-second, 768×512 clip renders in about 8 seconds on a single L40S, so you can try another prompt while the idea is still fresh.
Large weights, quick builds
The replica pulls the ~45 GB checkpoint at boot instead of baking it into the image, and a readiness probe keeps traffic away until the model is on the GPU.
Runs in your cloud
Prompts and rendered clips stay in your Chalk environment. Nothing passes through a third-party generation API.
Three files, open source
The server, the browser client and the deploy script are Apache-2.0 on GitHub. Swap in your own pipeline and keep the queue and the UI.
Run generative models on your own GPUs
Video, images, speech or language. Talk to an engineer about serving them in your cloud.