Hi everyone,

We are excited to share Agentic Cosmos, which moves image generation beyond a one-shot prompt-to-pixels call. An agent reasons over the request, plans the composition, and calls specialized tools to search, generate, render, and refine the image. The agent applies world knowledge, renders accurate text, and delivers a more polished layout.

Agentic Cosmos vs. GPT-Image 2

Below are two representative side-by-side results generated from the same prompts. Click either comparison to open the full 28-example PDF.

SAME PROMPT · SIDE-BY-SIDE

Introduction

One-shot image generators can create compelling pixels, but complex design requests need more than visual appearance. They may involve a known person or product, multiple coordinated regions, readable copy, consistent hierarchy, precise placement, and the ability to revise one defective area without regenerating everything.

Agentic Cosmos addresses this gap by representing the target image as a structured visual program. A reasoning model creates the complete initial composition as one tree, and every downstream stage executes that shared plan.

Highlight 1: Image as a Structured Layout Tree

The core representation is one layout-aware tree that defines the visual hierarchy, semantic regions, parent-child relationships, bounding boxes, layer order, typography, asset requirements, background intent, and overlap constraints.

The agent first plans the full composition, then breaks it into smaller tasks for each region and visual element. Every task records what should appear, where it belongs, and how it will be created—even when an element will be generated directly into the background.

Search, generation, text rendering, and composition all follow the same tree, so one region can be corrected without redesigning the rest.

Highlight 2: Grounded Search and Image Synthesis

For identity-sensitive leaves, the tree records both the visual intent and a concrete search query. The agent runs search, selects usable reference assets, and passes those references—together with the tree's spatial intent—to an image-generation model.

In the livestream example, the presenter, product, and brand are grounded separately, then synthesized into one coherent background rather than left for one-shot generation to invent.

Highlight 3: Bbox-Guided Composition

Once the background, visual assets, and component subtrees are ready, each output is placed using the bounding box and z-order defined in the Structured Layout Tree.

This creates a clear division of labor: the reasoner decides the structure and spatial relationships, search and generation create the visual content, and the executor composes the ready pieces.

Highlight 4: Localized Refinement Loop

After composition, visual review feeds issues found in the final pixels back to the same Structured Layout Tree. The agent maps each issue to the responsible node or subtree, repairs only that branch, recomposes the image, and reviews it again.

In the example below, the loop compacts and moves the itinerary-panel subtree while preserving the background pixels, train scene, and exact copy. The unaffected regions are not regenerated.

What Agentic Cosmos Demonstrates

Contributors & Internal Collaboration