Efficient Pretraining of Text-to-Image Models
summary
A concrete method claims 42x compression of megapixel images into 24x24 latents while reconstructing visual detail, using a VAE-style encoder plus a diffusion autoencoder (stage B) and a latent diffusion generator (stage C). Pipeline trains on CLIP/UNCLIP embeddings, uses low‑res diffusion upsamplers to avoid 64x64 limits, and yields faster sampling — the released stable cascade models reportedly outperform SDXL in inference time while still producing high‑quality images.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:23Language to visual models
- 1:27Text-to-image techniques
- 2:11DALL·E 1 visual tokenization
- 6:05Diffusion models
- 8:56DALL·E 2 CLIP embeddings
- 12:03Stable Diffusion latent
- 15:16VersaN compression limits
- 19:51Stable Cascade efficiency
- 20:53Beyond text-to-image
- 21:37Q&A
related talks
30mHow to train a large diffusion model
41mBuilding Generative Image & Video models at Scale
01h 17mDeep generative models for image...
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
28mPosterior Collapse in Deep Generative models
17mApproximate Caching for Efficiently Serving Text-to-Image Diffusion Models
41mBridging Pixels and Words: The Transformer Vision Revolution
45mHey, Look, I'm an Artist Now © ML Image Generator