Analyzing Learning-Dynamics Across Loss and Prediction Choices for Diffusion Models
A new arXiv paper (2611.17967v1) analyzes how the choice of prediction target and training loss affects learning dynamics in denoising diffusion models. The work examines models that predict clean data (x), noise (ε), or velocity (v) during training. The abstract indicates the study spans different loss formulations measured against these prediction choices. No experimental results, benchmarks, or numerical findings are available in the provided materials, and the paper is listed as a new announcement with no reported follow-up.
How this story unfolded
Key points4 items Model summary · ordered by when the evidence was published
- Point 01 11/16, 13:00
The paper is posted as arXiv:2611.17967v1 under the "new" announcement type.
Based on 1 - Point 02 11/16, 13:00
It studies learning dynamics of diffusion models across loss and prediction choices.
Based on 1 - Point 03 11/16, 13:00
Denoising generative models are described as predicting clean data (x), noise (ε), or velocity (v).
Based on 1 - Point 04 11/16, 13:00
Training losses are evaluated in relation to these prediction targets.
Based on 1
ReportsBy publish time · newer further down
- November 16, 2026 · Mon
- Start
E0: Expressive Fine-Grained Discrete Action Prediction for Vision-Language-Action Models via Tweedie Discrete Diffusion
Original summary arXiv:2511.21542v3 Announce Type: replace Abstract: Vision-Language-Action (VLA) models offer a unified framework for robotic manipulation by integrating visual perception, languag
arXiv cs.RO Original ↗ - Start
Analyzing Learning-Dynamics Across Loss and Prediction Choices for Diffusion Models
Original summary arXiv:2611.17967v1 Announce Type: new Abstract: Denoising generative models predict clean data ($x$), noise ($\epsilon$), or velocity ($v$), while training losses can be measured i
arXiv cs.LG Original ↗ - November 18, 2026 · Wed
- +2d 0h
Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers
Original summary arXiv:2611.22287v1 Announce Type: new Abstract: Latent diffusion models rely on a decoder to map denoised latents to pixels, adding complexity and inference cost. We introduce Deco
arXiv cs.CV Original ↗ - Still developing · later reports will keep extending this axis