Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

TOPIC · ENTITY #14663

diffusion models

Related event timeline, sources and context from the news index.

EVENT TIMELINE

4

RESEARCH · 1 SOURCE · arXiv cs.AI

StruPO: structure-aware prompt editor that improves faithfulness in text-to-image generation

The authors introduce StruPO, a structure-aware prompt editor that conditions on a calibrated scene graph to guide token-level prompt edits (deletions, reordering, insertions) and uses image-grounded supervision and smooth reward shaping to train a reusable editing policy. Evaluated on three disjoint 2K-prompt test sets across four frozen generation backbones and validated by automated metrics plus a 15-annotator study, StruPO reportedly improves the trade-off between aesthetics and relational faithfulness without iterative critique–rewrite loops.

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Correction Training for Diffusion Models (arXiv:2610.14136v2)

The paper introduces a learned "corrector" that is trained on the base denoiser's own predictions to close the training–inference gap caused by exposure bias in iterative diffusion sampling. At each denoising step the corrector iteratively refines the denoiser's clean-sample prediction; evaluated on DiT, SiT, and JiT this approach substantially improves FID for unguided and unconditional generation, is competitive with classifier-free guidance (CFG), can be combined with CFG to improve diversity, and its inference cost can be removed by distilling refined predictions back into the base model.

7.0

RESEARCH · 1 SOURCE · Apple Machine Learning Research

Normalizing Trajectory Models (NTM) introduced as a few-step, likelihood‑based generative method

Researchers (Jiatao Gu et al.) propose Normalizing Trajectory Models (NTM), which model each reverse sampling step as a conditional normalizing flow and are trained with exact trajectory likelihood. NTM combines shallow invertible blocks per step with a deep parallel predictor across the trajectory, can be trained from scratch or initialized from pretrained flow-matching models, and enables a self-distillation step that yields high-quality samples in four sampling steps on text-to-image benchmarks while retaining exact likelihood over the trajectory.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Inference-time projection for physically valid AlphaFold 3–style biomolecular diffusion models (arXiv:2610.07037v1)

The paper proposes treating physical validity as a constrained inference problem and introduces two closed-form projection operators—an inter-chain van der Waals projection and a ligand distance-geometry projection—applied to the diffusion model's denoised coordinate estimate (x̂0). The operators are local, sparse, displacement-capped, require no network evaluations or gradients, and — when applied to Boltz-2 and OpenFold-3 across five benchmarks (CASP15, CASP16, PoseBusters monomer and complex sets, and the Boltz physical-validity set) — recover perfect physical validity while preserving structural-accuracy and ligand-placement metrics with negligible runtime and memory overhead.

7.0