Reproduce “Generative Representation Learning for Scalable Latent Tokenization” on OpenResearch

About this paper

Alan Baade, Kyle Sargent, Vincent Sitzmann, Eric Ryan Chan, Justin Johnson, Li Fei-Fei, Gordon Wetzstein

RePR trains image representations to capture how context improves a generative predictor’s predictions of image patches. It uses this signal to guide a tokenizer, then generates pixels from that representation. On ImageNet, 50,000 generations yield FID 2.18 for RePR-78M plus JiT-L versus 2.36 for pixel JiT-L at nearly matched training compute. This compares training FLOPs, not inference cost.

Key insights

• Generative likelihood-based supervision offers an alternative to contrastive objectives and may reduce feature suppression; in a watermark stress test, RePR retained class-probing performance better than LeJEPA and DINOv3.
• Across tested compute ranges on ImageNet and GPIC, larger or more extensively trained RePR encoders generally improved generation, with reported power-law scaling trends.
• Representation training can be worthwhile when its cost is counted: the reported RePR system used nearly the same total FLOPs as pixel-space JiT-L while attaining a lower FID-50K.

Summary by alphaXiv.