Modal logo
FeaturedGetting started Hello, worldSimple web scraperServing web endpointsLarge language models (LLMs) Deploy an OpenAI-compatible LLM service with vLLMRun DeepSeek-R1 and Phi-4 with llama.cppLow-latency, serverless TensorRT-LLMRun Vision-Language Models with SGLangRun a multimodal RAG chatbot to answer questions about PDFsFine-tune an LLM with AxolotlReplace your CEO with an LLMImages, video, & 3D Fine-tune Wan2.1 video models on your faceRun Flux fast with torch.compileFine-tune Flux with LoRAAnimate images with video generation modelsGenerate video clips with MochiRun Stable Diffusion with a CLI, API, and web UIDeploy ControlNet demos with GradioAudio Deploy a Moshi voice chatbotCreate music with MusicGenComputational biology Fold proteins with Chai-1Build a protein-folding dashboardFold proteins with Boltz-1Sandboxed code execution Run a LangGraph agent's code in a secure GPU sandboxBuild a stateful, sandboxed code interpreterRun Node.js, Ruby, and more in a SandboxRun a sandboxed Jupyter notebookParallel processing and job scheduling Transcribe podcasts with WhisperDeploy a Hacker News SlackbotRun a Document OCR job queueServe a Document OCR web appTraining models from scratch Train an SLM with early-stopping grid search over hyperparametersRun long, resumable training jobsHosting popular libraries FastHTML: Deploy 100,000 multiplayer checkboxesYOLO: Fine-tune and serve computer vision modelsMultiOn: Create an agent for AI newsBlender: Build a 3D render farmStreamlit: Run and deploy Streamlit appsComfyUI: Run Flux on ComfyUI as an APISQLite: Publish explorable data with DatasetteAlgolia: Build docsearch with a crawlerConnecting to other APIs MongoDB: Vector and geospatial search over satellite imagesDiscord: Deploy and run a Discord BotGoogle Sheets: Sync databases and APIs to a Google SheetLangChain: Run a RAG Q&A chatbotTailscale: Add Modal Apps to your VPNPrometheus: Publish custom metrics with PushgatewayManaging data Mount S3 buckets in Modal appsBuild your own data warehouse with DuckDB, DBT, and ModalCreate a LoRA Playground with Modal, Gradio, and S3Miscellaneous

Deploy a personalized music video generation service on Modal

Music videos are cool, but unless you are famous or pay a lot of money, you don’t get to star in them.

Until now!

The repo includes all the code you need to deploy a custom music video generator on Modal, a serverless infrastructure platform for data, ML, and AI applications.

Below is a sample video, generated by Modal Developer Advocate @charles_irl.

And because Modal is generic serverless infrastructure, you can customize this custom music video generator however you wish — it’s just code and containers!

Setup

In the Python environment of your choosing, run pip install modal.

If you run into trouble with Python environments, we suggest using this Google Colab notebook, where we’ve set the environment up for you. It’s a bit of work to get used to running terminal commands in a notebook if you haven’t done that before, but the Python setup works and running the notebook in Colab is free! All you need is a Google account.

Then, if you’ve never used Modal on the computer you’re using, run modal setup to create an account on Modal (if you don’t have one) and set up authentication.

Data Prep

Create a folder inside data/, parallel to the sample data, data/sample. You can name it whatever you want.

Place at least four images of yourself in that folder — ideally eight or more. Images should be in .png or .jpg format and around 400 to 800 pixels on each side. For best results, we recommend putting a variety of images, in particular where you are wearing different clothes and making different faces, and including some images that have other people in them. But you can also just take a few photos of yourself right now!

Optionally, add captions in .txt files in that same folder. They should look something like "[trigger] smiling at the camera, outdoor scene, close-up, selfie". See the sample data for more example image-caption pairs.

Training

Start up a JupyterLab server on Modal with

modal run train_from_notebook.py

Click the modal.host URL that appears in the output to open Jupyter in the browser.

Open the training notebook, training.ipynb.

Read the notebook and run it, following the instructions to edit cells as needed.

In particular, change the dataset path to the folder you created — it has been mounted on the remote cloud machine where the notebook is running.

You can also directly upload data to the /root/data folder on the remote machine. You can even edit caption files inside of JupyterLab! This data will stick around between runs, and you can find it with

modal volume ls finetune-video-data

See the help for modal volume and its subcommands for details.

The notebook will kick off training, which takes a few minutes. Take note of the name given to your training run. By default, it’s a hash like 38c67a92f6ce87882044ab53bf94cce0, but you can customize it in the notebook. This is your finetune-id.

If you forget it, you can show all of your finetune-ids by running

modal volume ls finetune-video-models

Inference

Test out your new fine-tuned model by running:

modal run inference.py --finetune-id {your-finetune-id} --num-frames 15

You can also provide a --prompt to customize the generation.

You can deploy the video generator onto Modal with

modal deploy inference.py

Modal is serverless, so this won’t cost you any money when it isn’t serving any traffic.

Music video generation

Once you’ve deployed an inference endpoint, you can generate a music video starring yourself by running

modal run music_video_generator.py --finetune-id {your-finetune-id}

With the default settings, this will create a thirty second video in about five minutes by running generation in parallel on seven H100s.

The music can be changed by passing in a different song via the --mp3-file argument. The default is a Modal-themed song in data/coding-up-a-storm.mp3. This song was created with Suno, a music generation service — that runs on Modal! If you want to DIY music generation as well, see this example in the Modal docs.

The generated clips can be changed by passing a different list of prompts via the --prompt-file argument. The default is a set of prompts created with OpenAI’s GPT-4.5 system. You can write your own or generate them with a language model. If you want to serve your own language model, see this example in the Modal docs.