AI Chronicles

text-generation · 623 likes · 233701 downloads

150 likes · 0 downloads

text-generation · 283 likes · 7629 downloads

tabular-classification · 89 likes · 3 downloads

text-generation · 376 likes · 191409 downloads

image-text-to-text · 107 likes · 12635 downloads

text-generation · 120 likes · 901 downloads

image-text-to-text · 1600 likes · 630246 downloads

token-classification · 96 likes · 414 downloads

text-generation · 3192 likes · 159967 downloads

image-text-to-text · 1184 likes · 1113871 downloads

text-generation · 499 likes · 34371 downloads

text-generation · 127 likes · 13038 downloads

image-to-video · 138 likes · 0 downloads

text-generation · 183 likes · 21935 downloads

text-generation · 929 likes · 288741 downloads

image-to-image · 49 likes · 0 downloads

102 likes · 0 downloads

46 likes · 20 downloads

text-generation · 623 likes · 114499 downloads

text-generation · 191 likes · 5792 downloads

text-generation · 130 likes · 196441 downloads

text-generation · 336 likes · 46677 downloads

text-generation · 200 likes · 136933 downloads

text-generation · 181 likes · 2671 downloads

text-generation · 296 likes · 135452 downloads

text-generation · 49 likes · 16824 downloads

text-generation · 154 likes · 511 downloads

text-generation · 53 likes · 33018 downloads

text-to-image · 441 likes · 56953 downloads

text-generation · 138 likes · 1766 downloads

text-generation · 56 likes · 37885 downloads

text-generation · 63 likes · 38140 downloads

text-generation · 490 likes · 212201 downloads

text-generation · 2556 likes · 597090 downloads

text-to-image · 264 likes · 39515 downloads

225 likes · 10 downloads

text-generation · 760 likes · 72715 downloads

Conventional reinforcement learning strategies for visual generation typically employ sample-wise reward functions, yet this practice frequently results in reward hacking that degrades image diversity and introduces visual anomalies. To address these limitations, we present a novel framework that finetunes generative models using distribution-wise rewards, ensuring better alignment with real-world data distributions. Unlike rewards that evaluate samples individually, distribution-wise reward accounts for the data distribution of the samples, mitigating the mode collapse problem that occurs when all samples optimize towards the same direction independently. To overcome the prohibitive computational cost of estimating these rewards, we introduce a subset-replace strategy that efficiently provides reward signals by updating only a small subset of a generated reference set. Additionally, we apply RL to optimize post-hoc model merging coefficients, potentially mitigating the train-inference inconsistency caused by introducing stochastic differential equation (SDE) in regular RL practices. Extensive experiments show our approach significantly improves FID-50K across various base models, from 8.30 to 5.77 for SiT and from 3.74 to 3.52 for EDM2. Qualitative evaluation also confirms that our method enhances perceptual quality while preserving sample diversity.

Representation alignment has become an effective way to accelerate diffusion transformer training and improve generation quality. Recent self-alignment methods, such as SRA and Self-Flow, further remove the dependency on external pretrained encoders by constructing alignment within the diffusion model itself. However, the mechanism behind the improvement from SRA to Self-Flow, dual-time scheduling, remains under-examined: Self-Flow attributes its gain to interactions between tokens at different noise levels, where cleaner tokens help infer noisier ones. In this work, we revisit this explanation and ask whether the gain instead comes from data augmentation along the noise dimension. To disentangle these factors, we introduce Attention Separation, which preserves the same dual-timestep input as Self-Flow while blocking attention between tokens assigned to different noise levels. Surprisingly, removing such interaction does not degrade performance and can even improve it, suggesting that the improvement from SRA to Self-Flow mainly comes from data augmentation. Furthermore,We show that Attention Separation itself provides an augmentation effect by splitting a single image into multiple effective training parts to expand the training data. Based on these observations, we combine self-representation alignment with dual-timestep and attention-separation augmentation, and demonstrate the effectiveness of this design on ImageNet.