Deploy Your Own VideoGen Model
summary
Walks through training a DiT/LTX-2.5 pipeline using LoRA fine-tuning on just 20 five-second clips, runs on Modal with B200 per-second billing, and deploys a serverless GPU endpoint that cold-starts in seconds. Claims fine-tune finishes in ~50 minutes for a few dollars, reuses base model knowledge, and yields a recognizable, era-matched recreation of a teenage likeness rather than a generic subject. Shows code, storage layout, training config, and a working generate endpoint that returns videos.
AI-written summary · about
chapters (12)
- 0:00Welcome
- 1:52Motivation for custom video generation
- 3:57History of video generation models
- 5:23How diffusion models work
- 7:28ML training and inference
- 10:57Customizing with LoRA
- 13:24Introducing Modal
- 15:51Setting up Modal for training
- 17:36Training the LoRA model
- 18:39Deploying inference server
- 19:22Demo results
- 20:45Future ideas and Q&A
related talks
22mMovie Gen Video: State-of-the-art Video Generation
25mLights, Camera, Inference! Video Generation as a Service With VLLM-O...
33mHow Genmo Scales Video Generation with
17mThe State of Generative Media
24mState of Open Video Generation Models
51mSpeeding up deployment of machine learning models
14mEfficient Training of Video Generation Foundation Model at ByteDance
32mVideo Action Models