New Study Quantifies What Diffusion Models Generate First: Only RAE Space Shows Abstract-to-Concrete Order
sedielem · x · 2026-10-12
When does diffusion generate each feature?
Researchers from Westlake University and collaborators introduce feature information dynamics, a framework that uses optimal denoising losses to quantitatively trace when each feature emerges during diffusion generation in arbitrary representation spaces.
- Builds on the "diffusion as spectral autoregression" view (low frequencies emerge first in pixel space), which relies on natural images' power-law spectra — an argument that doesn't transfer to learned latent representations.
- Key finding: the choice of representation space changes the implicit generation order. In their comparison, only the diffusion model in RAE latent space exhibits an abstract-to-concrete hierarchy, and it also converges fastest.
- An interactive visualization accompanies the paper, showing denoising losses, loss gaps, and accumulated class information across log-SNR.
Finding the optimal generation order and representation per modality is left as future work.
More from Multimodal
- Rigging in Blender via MCP: six-step recipe for Opus 5.5 to build and animate 3D scenes — sidahuj · 2026-10-12
- "Fuzzbeast style" AI illustrations produce delightfully weird red-eyed characters — OVolosin82152 · 2026-10-12
- First look at ByteDance's new World Model: text-to-navigable 3D worlds with real-time editing — Uncanny_Harry · 2026-10-12
- Stanford's WonderZoom generates infinite zoomable 3D detail from a single image — import_jmr · 2026-10-12
- AI-generated orange cat band 'The Five Meow Band' performs a full bar show — Reasonable_Damage_18 · 2026-10-12
- 'When Sand Speaks' Ep 1: an AI-voiced series from stone tools to silicon — projectforty2 · 2026-10-12