AI research wiki
vision-audio-video · 2026-05-01 · Tier 3

Visual Generation in the New Era: Atomic to Agentic World Modeling

Visual Generation in the New Era: Atomic to Agentic World Modeling

arXiv: 2604.28185 · paper · HF Tier: 3 — visual generation survey / taxonomy Raw: ../../raw/huggingface/2026-05-01-visual-generation-new-era-evolution-atomic-mapping-agentic-world.md

TL;DR

A taxonomy paper that argues visual generation has progressed from appearance synthesis to intelligent generation — visuals grounded in structure, dynamics, domain knowledge, and causal relations. Five-level taxonomy: Atomic Generation → Conditional Generation → In-Context Generation → Agentic Generation → World-Modeling Generation. Each level nests prior capabilities and adds a qualitatively new one. Drivers: diffusion-to-flow-matching, unified understanding+generation, improved visual representations, SFT and preference-based post-training, reward modeling, large-scale data curation, sampling acceleration. Pairs the taxonomy with stress tests and case studies that map failure modes to taxonomy levels.

Why this is useful

The taxonomy is genuinely informative as a navigational structure for the visual-generation space. Most current models are stuck at Conditional or In-Context Generation; "Agentic" and "World-Modeling" are aspirational. Useful for placing today's papers (PhyCo physics priors → Conditional+; X-WAM 4D world action → World-Modeling beginnings).

Connection to prior wiki

  • PhyCo (05-01) lands at Conditional Generation with physical-property maps as conditioning.
  • X-WAM (04-30) is closer to World-Modeling Generation with 4D dynamics and action coupling.
  • Seedance 2 (04-16) / GLM-5V-Turbo (04-30) sit at In-Context Generation with multimodal inputs.
  • Agentic Generation is largely empty in today's literature — the level the field has not credibly reached.

Research angle

Taxonomy papers are useful for placing work, less useful for driving it. The interesting consequence: the gap between current models (Conditional / In-Context) and Agentic / World-Modeling is the explicit research target the survey identifies. Whoever builds the first credible Agentic Generation system (visual generator with persistent state and closed-loop feedback) sets the next level.