Training a foundation image model and deploying it to millions of users
summary
A custom diffusion model is trained from scratch with a bespoke dataset and architecture, running on 256 H100s and consuming energy comparable to 600 French households; data pipeline uses a vector database and Go exporter (open source) to compress and serve up to ~10,000 images/s for training while inference traffic peaks at ~300–400 requests/s. Performance wins come from step distillation and TensorRT compilation, yielding about 3x faster generation and deployment to millions of users.
AI-written summary · about
chapters (12)
- 0:00Image generation models intro
- 1:28Why PhotoRoom trained custom model
- 2:35Live demo: AI background generation
- 4:21Diffusion models in one minute
- 6:32Key ingredients for training
- 7:16Compute requirements
- 9:23Model size and architecture
- 12:14Data collection and filtering
- 15:47Deploying to millions
- 17:33Optimizing model speed
- 18:59Future plans and open-source
- 19:41Conclusion and Q&A
related talks
30mHow to train a large diffusion model
28mComputing Large Foundational Models
04mProcessing 100+ million images per month with deep learning
41mBuilding Generative Image & Video models at Scale
27mFoundation Models: Blueprints for success
34mDeep Learning at Scale
19mDeploying ML models to mobile devices
17mAll AI Roads Lead to Distribution