JiT-DDT Architecture Trains Text-to-Image Models 3.6x Faster
Researchers have introduced JiT-DDT, a novel encoder-decoder pixel-space architecture designed to accelerate text-to-image model training. Traditional Latent Diffusion Models (LDMs) rely on a separate Variational Autoencoder (VAE) and Diffusion Transformer (DiT), which limits token compression and creates training bottlenecks. By eliminating the VAE and integrating compression directly into the…
Key points
- JiT-DDT trains text-to-image models 3.6x faster than the Linum v2 baseline while generating four times the pixel resolution.
- The architecture eliminates the traditional VAE, pushing compression directly into the DiT to achieve 32x32 token reduction.
- The code and research model weights have been open-sourced under the Apache 2.0 license as a research artifact.
To overcome the loss of fine-grained detail common in pixel-space models, JiT-DDT splits the DiT into an encoder that predicts a low-resolution structural sketch and a decoder that reconstructs high-frequency details. When tested against the Linum v2 baseline, JiT-DDT trained a text-to-image model using 3.6× fewer GPU-hours while generating images with 4× more pixels. The team has released the code and model weights under an Apache 2.0 license.
Training Text-to-Image Models 3.6× Faster
linum.ai · 16 September 2026Loading the full article…
This text was published by linum.ai. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Hacker News discussion · 43 pointsnews.ycombinator.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Anthropic research suggests small AI models may self-replicate on IoT devices · 7 src
- Researcher predicts AI engineering acceleration but not superintelligence · 1 src
- MIT researchers show neural networks can build own learning rules · 1 src
- Anthropic co-founder predicts full RSI possible by 2027 · 1 src
- Sakana AI’s MASS method improves model self-judgment but risks baking errors into training · 1 src
Comments
via GitHub Discussions