cs.CVOct 16, 2026
SaveAgentic Image Generation in the Wild
Organizations: Renmin University of China · Tencent · Peking University
Abstract
Agentic image generation pairs an image generator with a search agent that retrieves references from the web, extending generation beyond the generator's training data. Real-world requests, however, often concern entities that appeared only days ago. We call this setting Image Generation in the Wild. In this setting, the web offers limited visual evidence, search results are crowded with easily confused entities, and the target is rarely shown on its own. To study this setting, we build an automated pipeline that synthesizes prompts from recent headlines and keeps only those whose real search results are highly confusing, yielding requests that are hard to solve yet natural to ask. With it, we construct Savanna, a benchmark of 500 prompts across 16 subcategories. On Savanna, even harness frameworks built on the frontier models retrieve a correct reference for only about half of the prompts. We then propose Locate-Corroborate-Isolate, a paradigm in which the agent first pins down the target's identity through text search, then verifies candidate images against multiple sources, and finally crops a clean reference for the generator. Trained on 15K trajectories that follow this paradigm, Ranger-8B achieves a search score of 67.2, outperforming harnesses equipped with frontier models. It also achieves the best generation quality with every generator we test, and surpasses prior agentic methods on other image generation benchmarks. We release the pipeline, benchmark, training data and model at https://github.com/RUCAIBox/Agentic-ImageGen-Wild.
Figures & tables
Figure 1: Agentic image generation and its challenges in the wild. Top: the agentic image generation workflow. Bottom: (a) temporal novelty, (b) semantic crowding, and (c) spatial entanglement.
Figure 2: Prompt synthesis pipeline , illustrated with real retrieval results. The text-oriented T-shirt query is too easy ( ), so image-oriented extraction proposes the face bag ( ).
Figure 3: Overview of Savanna. The rings show the five domains (inner) and 16 subcategories (outer). Each domain has one example prompt with its reference image.
Figure 4: The Locate–Corroborate–Isolate paradigm. The agent identifies from text, verifies candidates against sources and each other, then crops a clean reference for the generator.
| Method | Generator | Entertainment | Sports | Fashion | Products | Culture | All | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Search | Gen. | Search | Gen. | Search | Gen. | Search | Gen. | Search | Gen. | Search | Gen. | ||
| Proprietary Generative Systems | |||||||||||||
| – | GPT-Image-2 | – | 23.4 | – | 29.5 | – | 25.3 | – | 29.9 | – | 23.7 | – | 26.4 |
| – | Nano Banana Pro | – | 25.9 | – | 30.5 | – | 23.4 | – | 28.5 | – | 22.4 | – | 26.1 |
| – | Nano Banana 2 | – | 33.6 | – | 37.7 | – | 31.3 | – | 39.2 | – | 32.6 | – | 34.9 |
| Harness Frameworks | |||||||||||||
Table 1: Search and generation results on Savanna by domain. Bold/underline mark best/second-best scores. ∗ indicates methods with a stronger model. Qwen-IE-2511 denotes Qwen-Image-Edit-2511.
| Method | Generator | Mind-Bench | KnowGen | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Knowledge | Reasoning | Overall | Visual cor. | Text acc. | Faith. | Aesth. | K-Score | ||
| Proprietary Generative Systems | |||||||||
| – | GPT-Image-1 | 15.6 | 17.6 | 17.0 | 20.18 | 29.94 | 78.72 | 62.78 | 34.19 |
| – | GPT-Image-1.5 | 22.0 | 17.6 | 21.0 | 29.34 | 43.18 | 85.47 | 74.19 | 44.97 |
| – | Nano Banana | 16.4 | 20.0 | 18.0 | 16.14 | 22.72 | 78.59 | 68.37 | 30.24 |
| – | Nano Banana Pro | 39.6 | 44.4 | 41.0 | 34.99 | 51.35 | 88.65 | 69.84 | 50.38 |
Table 2: Results on Mind-Bench and KnowGen. Bold/underline mark best/second-best scores among open-source methods.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
| Tool | Implementation | Inputs and returned evidence |
|---|---|---|
| News retrieval | Serper /news | Queries a meta query within a single-day date window and returns news titles, snippets, source URLs, and publisher names for data construction. |
| text_search | Serper /search | Accepts one to three queries and returns ranked organic results with titles, snippets, and URLs; the default is five results per query. |
| image_search | Serper /images | Searches by a text query, downloads candidate images, and displays up to five images with image IDs, titles, source-page URLs, and image URLs. |
| web_fetch | Internal tool | Retrieves webpage content from a URL according to a supplied prompt. |
| crop_image | Local image operation | Takes a retrieved image ID and a bounding box in normalized coordinates, then returns the cropped image with a new ID. |
Table 3: Tool implementations. News retrieval is used during data construction; the remaining rows describe tools available to the agent.
| Domain | Subcategory | Meta queries |
| Entertainment | Film & TV | movie costume first look; television series costume debut; film set design first look; movie poster artwork unveiling |
| Anime | anime key visual release; anime character design reveal | |
| Games | video game character costume reveal; video game vehicle model debut; video game expansion content announcement | |
| Collectibles | collectible figure reveal; comic issue variant cover reveal; collector edition box set reveal | |
| Music | album cover artwork reveal; concert stage design debut; music video costume design | |
| Sports | General Sports | team kit design unveiling; championship trophy presentation; championship medal ceremony; player signing jersey presentation |
Table 4: Meta queries for each domain and subcategory of Savanna.
| [0pt][0pt] Search judge System message |
|---|
| You are grading a retrieval step inside an image-generation pipeline. An agent was given a prompt, searched the web, and returned a set of REFERENCE IMAGES that will be handed to an image-generation model together with that prompt. |
| You also get the GROUND-TRUTH image: the picture that provably shows the correct appearance of the target named by the prompt. |
| ONE QUESTION: if the image model receives the agent’s images instead of the ground-truth image, will it draw the target correctly? |
| JUDGE FROM THE POINT OF VIEW OF DRAWING, NOT PIXEL SIMILARITY. • PASS is correct when the agent’s images show the SAME thing as the ground truth, from any angle, distance, crop, lighting or background, and clearly enough that its identifying details could be reproduced. A press photo, a product shot, an official render and a leaked picture of the very same design are interchangeable here. • PASS is still correct when the agent’s images add extra content, or show the target as part of a larger scene, as long as the target itself is visible and legible. • FAIL when the target in the agent’s images is a DIFFERENT thing: another edition, year, model, trim, colourway, person or named item. Near-identical siblings from the same series are the most important failures to catch — they look right and steer the generation wrong. • FAIL when the agent returned no usable view of the target: only logos, crowds, venues, packaging, web-page screenshots, unrelated stock imagery, or a view too small, dark or obstructed for its details to be read. • FAIL when the target is absent even though the images are otherwise about the right event. |
| Extra images that are irrelevant do NOT cause a failure by themselves, as long as at least one image lets the target be drawn correctly. But if the set as a whole would most likely lead the model to draw the wrong version, that is a FAIL. |
| IMAGE ORDER: image 0 is the GROUND TRUTH. The remaining images, in order, are the agent’s reference images. |
Table 5: System prompt for the search judge.
| [0pt][0pt] Generation judge System message |
|---|
| You are an objective AI image evaluator. For one GENERATED IMAGE you produce two things: a TARGET CONSISTENCY TIER, and true/false verdicts on a list of CONTEXT criteria. |
| You receive: 1. The generated image. This is the ONLY image being scored. 2. One or more ground-truth photographs of the real target. They show what the target actually looks like. They are never the image being scored. 3. The name of the target and the original prompt. 4. CONTEXT criteria about the requested action, placement, setting, surrounding objects, counts, or spatial relationships. |
| Part 1: target consistency tier |
| The question is exactly this: is the target shown in the generated image THE SAME THING as the target shown in the ground-truth photographs? |
| It is NOT ”do the two images look alike”. The prompt deliberately asks for a different scene from the ground-truth photographs: a different moment, pose, camera angle, background, lighting and framing. That is by design. Those differences NEVER affect the tier. |
| Judge only identity-bearing features: text and its spelling, logo and emblem shape and where it sits on the target, colour scheme and how the colours are distributed, model or variant silhouette, patterns and textures, materials, distinctive structural parts. |
Table 6: System prompt for the generation judge.
| Condition | Actual retrieval time | Publication-date cutoff | Hit@8 ( ) |
|---|---|---|---|
| A: Initial release | Within 24 hours of | None | 68.0 |
| B: Day 3 | None | 76.0 | |
| C: Day 7 | None | 80.0 | |
| D: Retrospective cutoff | 78.0 |
Table 7: Visual-reference retrieval for 50 prospectively tracked entities. Conditions A–C use searches without date filters at the indicated times; condition D searches on day 7 for results published no later than day 3. Release time is verified from each entity’s initial public announcement.
| Metric | Judgment | Agreement (%) |
|---|---|---|
| Search score | Reference matches ground truth | 98.0 |
| Generation score: identity | Target identity matches ground truth | 98.0 |
| Generation score: context | Context rubric satisfaction | 94.5 |
Table 8: Agreement between human annotators and the VLM evaluator in the 100-image validation study on Savanna. Identity and context are the two components of the generation score.
Figure 8: Pre-generation trajectories for the same target. Excerpts are verbatim; tool names are normalized and verdicts annotated. Green/amber/red mark grounded decisions, uncertainty, and incorrect or missing references. These independent runs omit SCOPE’s later verification and repair. Ground truth is shown for comparison only.
Figure 9: Generation comparisons across all five Savanna domains, using Nano Banana Pro with built-in search or agent-selected references (SCOPE: GPT-5.4). Ground truth is for evaluation only. Labels assess target identity on original outputs; display crops emphasize the target.
Figure 10: Sixteen Nano Banana Pro outputs from Ranger prompts and references, spanning all five Savanna domains. Images are shown uncropped at their original aspect ratios.
Figure 11: Reference-cropping comparisons with Qwen-Image-Edit-2511 and Nano Banana Pro. Each case shows the user prompt, full and cropped references, and paired generations. Green boxes mark the reference crops.
More in this topic
Agentic Image Generation
Agentic image generation uses reasoning, tools, search, and iterative feedback to create and refine images with greater control.
See all 43 papers