Diffusion Intelligence Atlas

A research map of diffusion models and embodied intelligence

A curated map connecting diffusion and flow models with scalable generation, efficient systems, reinforcement learning, world models, embodied AI, and native multimodal intelligence. DiT remains a central architectural thread, while adjacent work is included where it helps explain how these areas converge.

156 papers 139 in the last 3 years 7 research topics exports updated Aug 2026

research landscape

Papers are positioned from SPECTER embeddings of their titles and full abstracts, with UMAP preserving local semantic neighborhoods. Colors show seven interpretable research topics, while distance and links reflect semantic similarity. Select a topic, tag, or date range to update the whole view. Pinch or use the mouse wheel to zoom; drag or use two-finger scrolling to pan.

publication window All dates

topic cloud

Larger terms appear more often in the current selection. Select a term to filter both the map and paper list.

cross-topic bridges

Pairs from different clusters that remain close in the similarity space—useful starting points for cross-topic reading.

papers

Search titles, authors, topics, or summaries. Open the source page for metadata or PDF for the paper.

Loading the paper catalog…

about this collection

Time coverage

The main window covers papers first released from 2 Aug 2023 to 2 Aug 2026. Earlier work is included when it is needed to understand how diffusion-based intelligence evolved.

Inclusion criteria

The collection favors papers that introduce a reusable architectural, training, systems, or evaluation idea. Each paper is also labeled by how directly it relates to DiT.

Sources & maintenance

Metadata and full English abstracts are checked against arXiv or the original paper. Chinese abstracts are translated with GPT-5.6-sol and cached in the catalog. Semantic embeddings and the map are rebuilt when the catalog changes.