AI BriefingGEAR: An End-to-End Training Method That Boosts Image Generation Efficiency by 10x
Key point
A new study unveiled GEAR, which combines a tokenizer and a generative model to improve image generation training speed by 10x.
Details
Existing image generation models use a two-stage approach where the tokenizer is trained first, followed by the generative model, which causes a mismatch problem between the two models.
The newly proposed GEAR(Guided End-to-end AutoRegression) jointly trains a VQ tokenizer and an autoregressive (AR) generator end-to-end through Representation Alignment. The non-differentiability issue that arises in this process is resolved using a Dual Read-out(Hard/Soft branch) approach, which guides the tokenizer in a direction that is easier for the generative model to predict.
Key achievements include:
- Training Speed Innovation: ImageNet gFID convergence speed is up to 10x faster compared to the existing LlamaGen-REPA
- Feature Alignment Optimization: Guides the generative model's features to become more similar to DINOv2, securing spatial consistency
- Versatility: Applicable to various quantization methods such as VQVAE, LFQ, IBQ, as well as Text-to-Image generation tasks
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
MultiMatte, a Promptable Background Removal Model Based on SAM 3, Released, Achieving S-measure of 0.901 on DIS-VD BenchmarkThree.js r186 Introduces Native 3D Gaussian Splatting Support and WebGPU Rendering