analiza aiqbench
This research investigates the learning dynamics of latent Transformers when predicting clean endpoints versus velocities in flow-based sampling, revealing that algebraically equivalent objectives induce distinct computational behaviors within finite networks. By shifting from velocity prediction to direct clean regression, the study demonstrates significant improvements in ImageNet generation quality, reducing FID scores substantially at both Base and Large scales. The findings highlight how architectural choices regarding output interfaces directly impact model efficiency and performance in high-dimensional generative tasks.
Equivalent Flows, Unequal Learning: Clean-Latent Prediction in Transformers
arXiv:2605.27102v3 Announce Type: replace-cross Abstract: Flow samplers consume velocity, but the neural network can predict the clean endpoint and convert it to velocity through a fixed affine readout. We study this choice with JLT, a latent Transformer in a frozen variational autoencoder (VAE) representation. For squared error, the optimal clean and velocity predictors are algebraically equivalent; a finite Transformer assigns different computation to its learned output under the two interfaces. A local Gaussian analysis identifies a known residual response supplied by the readout and isotropic target variance added by velocity prediction. Measured FLUX.2 channel spectra support this geometric distinction: 90% of target variance occupies 83 of 128 clean directions versus 109 velocity directions. Under a matched velocity objective, clean prediction improves ImageNet FID-50K from 6.56 to 2.70 at Base scale and from 2.12 to 1.47 at Large scale, with lower FID at every measured Large checkpoint. Scaling clean prediction to 951M parameters reaches FID-50K 1.19 and IS 271.96. In addition, an objective ablation at Base scale shows that direct clean regression reaches FID-50K 2.38 without time-dependent error weighting. These results show how moving known computation outside the network changes learning under algebraically equivalent flow interfaces. Code: https://github.com/akatsuki-neo/JLT/blob/main/README.md
Abstract:Flow samplers consume velocity, but the neural network can predict the clean endpoint and convert it to velocity through a fixed affine readout. We study this choice with JLT, a latent Transformer in a frozen variational autoencoder (VAE) representation. For squared error, the optimal clean and velocity predictors are algebraically equivalent; a finite Transformer assigns different computation to its learned output under the two interfaces. A local Gaussian analysis identifies a known residual response supplied by the readout and isotropic target variance added by velocity prediction. Measured FLUX.2 channel spectra support this geometric distinction: 90% of target variance occupies 83 of 128 clean directions versus 109 velocity directions. Under a matched velocity objective, clean prediction improves ImageNet FID-50K from 6.56 to 2.70 at Base scale and from 2.12 to 1.47 at Large scale, with lower FID at every measured Large checkpoint. Scaling clean prediction to 951M parameters reaches FID-50K 1.19 and IS 271.96. In addition, an objective ablation at Base scale shows that direct clean regression reaches FID-50K 2.38 without time-dependent error weighting. These results show how moving known computation outside the network changes learning under algebraically equivalent flow interfaces. Code: this https URL
Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
arXiv:2605.27102 [cs.CV]
(or arXiv:2605.27102v3 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2605.27102 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 292 cuvinte
Subiecte conexe
Build a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod
AWS · 4 sept. 2026
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · 8 oct. 2026
SymTRELLIS: Symmetry-Enforced Voxel Latents for 3D Generation
arXiv · 5 iun. 2026
SymTRELLIS: Symmetry-Enforced Voxel Latents for 3D Generation
arXiv · 9 sept. 2026