Image generation and editing models have advanced rapidly, yet remain unreliable when prompts require external world knowledge. Bounded and long-tail parametric knowledge prevents direct or reason-then-generate approaches from recovering the required facts and visual appearances. Existing agentic generation and editing methods mitigate this limitation with retrieval tools, yet remain constrained by insufficient visual verification, overloaded policy models, and weak integration of retrieved textual and visual evidence. To address these limitations, we present WeAgent-MMGenEdit, a full-stack recipe including a multimodal harness, a scalable data construction pipeline, a comprehensive benchmark, and post-training methods for the agent policy and image backend. We first introduce WeAgent-Harness, a multimodal runtime with persistent evidence management and dedicated verification and integration tools that organize retrieved multimodal evidence into a dense carrier. Upon this, we develop a scalable pipeline for prompt synthesis and agentic trajectory collection, yielding 23K supervised trajectories and 14.7K RL tasks with three-layer verifiable checklists. We further introduce WeBench-MMGenEdit, a bilingual benchmark covering both knowledge-intensive image generation and multi-image editing. Finally, a two-sided post-training recipe based on SFT and RL improves the agent policy and image backend. Together, WeAgent-MMGenEdit enables a 30B-total/3B-active policy to outperform similarly sized policy models and approach the performance of a 1T-parameter agent.
Figure 2: WeAgent-MMGenEdit on knowledge-intensive generation (top) and multi-reference editing (bottom).
Figure 3: Three paradigms for multimodal knowledge-intensive image generation and editing. (A) Closed-book generation relies solely on parametric knowledge and suffers from severe factual hallucinations. (B) Existing agentic systems retrieve external evidence but fail to effectively verify and integrate it because of insufficient visual inspection, an overloaded policy model, or weak entity–attribute–layout binding. (C) WeAgent-MMGenEdit explicitly verifies retrieved visual evidence and organizes multimodal evidence into a dense rendered carrier that provides both content and layout guidance for final generation, yielding the most accurate and faithful generation results.
Figure 4: WeAgent-Harness: A Multimodal Runtime for Retrieval, Verification, and Integration. Search tools acquire external textual and visual evidence, which is stored under stable identifiers. Dedicated visual verification and code-based integration modules inspect candidate images and organize verified multimodal evidence into a dense carrier for final image generation and editing. The final response interleaves the generated image with concise textual information; all interactions and artifacts are recorded for training and evaluation.
Role
Tool
Function
Retrieve
Web_Search
Queries the open web and returns titles, links, snippets, and dates.
Web_Extract
Fetches a webpage and extracts information relevant to the current need.
Search_for_Image
Retrieves candidate images using a textual query.
Search_by_Image
Searches the web using an image for identification and provenance.
Verify
Vision_Analyze
Answers focused questions about registered images.
Integrate
Execute_Code
Integrates verified multimodal evidence and renders it into a draft image.
Table 1: Tool interfaces in WeAgent-Harness , organized by the retrieve–verify–integrate–deliver workflow.
Figure 5: Construction of WeDataset-MMGenEdit. Top: bilingual multimodal knowledge is organized into a concept bank and knowledge graph, from which multi-hop tasks and three layers of verifiable checklists are synthesized simultaneously. Bottom: tasks are executed in WeAgent-Harness, and the resulting agentic process, generation input, and final image are independently graded and routed to SFT, RL, or evaluation pools.
Figure 6: Composition of WeBench-MMGenEdit. The benchmark contains 300 human-audited cases, exactly balanced between generation and editing and between English and Chinese. It emphasizes deep multimodal retrieval and knowledge-intensive generation and editing: 94% of cases require at least five retrieval hops, and 69.3% of editing cases contain multiple user-provided images.
Figure 7: Checklist-grounded agentic RL. Grouped trajectories are sampled in WeAgent-Harness and scored by isolated process and generation-input judges. Valid trajectories are normalized within each prompt group and optimized with GSPO using only policy-generated tokens. Rollout, reward evaluation, and actor optimization run asynchronously, with updated parameters periodically synchronized to rollout workers.
Figure 8: Two-stage image editing post-training. (a) Multi-reference Image Editing SFT learns stable conditioning over heterogeneous references using independent resolution buckets and target-only supervision. (b) Diffusion-NFT RL stage samples multiple candidates, evaluates them with a decomposed multimodal reward, and optimizes group-relative preferences while remaining anchored to the SFT model.
Model
Policy Model
Gen/Edit Model
Harness
Image Generation
Image Editing
IA
TKA
VKA
VQ
W_Avg
IA
TKA
VKA
CP
VQ
W_Avg
Direct Gen/Edit
GPT-Image-2
N/A
GPT-Image-2
N/A
63.20
33.83
42.18
85.13
48.79
53.86
23.67
47.95
81.06
82.33
48.99
Gemini3.1-Flash
N/A
Gemini3.1-Flash
N/A
67.18
46.50
45.70
81.41
55.12
56.35
31.49
54.60
75.08
77.72
51.79
Gemini3.1-Flash-Lite
N/A
Gemini3.1-Flash-Lite
N/A
65.77
46.22
43.84
79.06
53.81
53.07
30.52
47.95
66.71
76.27
48.51
Seedream-5.0-pro
N/A
Seedream-5.0-pro
N/A
58.07
29.91
35.56
80.01
43.58
54.09
26.37
47.72
77.34
78.18
48.60
Table 2: Quantitative Results on Image Generation and Editing. Comparison of direct generation/editing models, reasoning-based pipelines, existing agentic methods, agentic generation/editing within WeAgent-Harness, and our WeAgent-MMGenEdit models. TKA and VKA , highlighted in light yellow, are the primary knowledge-oriented evaluation dimensions.
Models
Science & Knowledge
Pop Culture & News
Overall
Visual cor.
Text acc.
Faithfulness
Aesthetics
Visual cor.
Text acc.
Faithfulness
Aesthetics
Direct Gen/Edit
GPT-Image-1
20.92
27.89
72.79
63.95
19.43
31.98
84.64
61.60
34.19
GPT-Image-1.5
29.25
40.14
81.29
77.21
29.43
46.22
89.64
71.17
44.97
Gemini 2.5 Flash Image
18.03
19.39
72.79
65.82
14.24
26.04
84.39
70.91
30.24
Gemini 3 Pro Image
39.46
49.32
86.22
70.92
30.51
53.37
91.07
68.75
50.38
Table 3: Quantitative Results on the KnowGen Benchmark. We compare direct image generation models and agentic generation methods on the Science & Knowledge and Pop Culture & News subsets.
Model Name
Knowledge-Driven
Reasoning-Driven
Overall
SE
Weather
MC
IP
WK
SL
Poem
Life Reason
GU
Math
Direct Gen/Edit
GPT-Image-1
0.32
0.06
0.22
0.02
0.16
0.32
0.10
0.24
0.10
0.12
0.17
GPT-Image-1.5
0.36
0.18
0.22
0.04
0.30
0.34
0.08
0.34
0.10
0.02
0.21
Gemini 2.5 Flash Image
0.30
0.10
0.12
0.00
0.30
0.32
0.36
0.20
0.04
0.08
0.18
Gemini 3 Pro Image
0.50
0.36
0.40
0.16
0.56
0.62
0.68
0.30
0.16
0.46
0.41
Table 4: Quantitative Results on Mind-Bench. We compare direct image generation models and agentic generation methods across knowledge-driven and reasoning-driven tasks.
Figure 9: Qualitative comparison on knowledge-intensive generation and editing. WeAgent-MMGenEdit more effectively retrieves, verifies, and integrates multimodal evidence, yielding more accurate and consistent outputs across different image-generation backends than direct and existing agentic baselines.
Model Name
Harness
Image Generation
Image Editing
Agentic Chain Score
Prompt Score
Agentic Chain Score
Prompt Score
TKR
VKR
MEI
Avg
TKS
VKS
Avg
TKR
VKR
MEI
Avg
TKS
VKS
Avg
Open-Source Agentic Gen/Edit
Mind Brush (GPT-5.1)
Mind Brush
50.31
46.95
49.27
48.84
52.91
44.10
48.51
20.76
20.17
36.10
25.68
26.71
10.01
18.36
GenSearcher-8B
GenSearcher
45.41
44.17
41.54
43.71
43.89
38.73
41.31
18.66
19.13
27.54
21.78
22.55
8.42
15.49
GenEvolve-8B
GenEvolve
42.14
53.66
42.71
46.17
46.03
46.47
46.25
22.38
21.59
35.92
26.63
29.92
17.37
23.65
Table 5: Comparison of Agentic Chain and Prompt Scores. We evaluate existing agentic systems and different policies within WeAgent-Harness on generation and editing.
Ablation
Performance
Search Tools
Verify Tools
Integrate Tools
Agent SFT
Agent RL
Process Reward
Gen. W-Avg
Edit. W-Avg
Harness Ablation
✓
×
×
–
–
–
47.22
42.14
✓
✓
×
–
–
–
49.03
43.30
✓
×
✓
–
–
–
48.70
42.91
✓
✓
✓
–
–
–
51.86
45.80
Table 6: Ablation study of WeAgent-MMGenEdit. Contributions of WeAgent-Harness components and different stages of agent post-training are evaluated on image generation and editing using weighted-average performance.
Agent SFT
Agent RL
Initialization
Qwen3-VL-30B-A3B
WeAgent-SFT
Training data
∼23 K expert trajectories
∼14.7 K RL task pool
Batching
Global batch 32
64 prompts × 8 trajectories
Maximum context
49,152 tokens
32,768-token episode context
Optimizer
Adam
Adam
Learning rate
1×10−5
5×10−7
Table 7: Main configurations for agent-side post-training.
WeEdit-M-SFT
WeEdit-M-RL
Initialization
Qwen-Image-Edit-2509
WeEdit-M-SFT
Training data
∼50 K samples
Hard subset from ∼15 K RL candidates
Trainable parameters
LoRA, rank 256
LoRA, rank 256
Reference images
1–5
1–5
Learning rate
5×10−5
5×10−5
Candidates / condition
–
K=12
Table 8: Main configurations for image-side post-training.
Figure 10: Agentic trajectory for Case 1 (Part 1 of 4).
Figure 11: Agentic trajectory for Case 1 (Part 2 of 4).
Figure 12: Agentic trajectory for Case 1 (Part 3 of 4).
Figure 13: Agentic trajectory for Case 1 (Part 4 of 4).
Figure 14: Agentic trajectory for Case 2 (Part 1 of 4).
Figure 15: Agentic trajectory for Case 2 (Part 2 of 4).
Figure 16: Agentic trajectory for Case 2 (Part 3 of 4).
Figure 17: Agentic trajectory for Case 2 (Part 4 of 4).
Figure 18: Agentic trajectory for Case 3 (Part 1 of 6).
Figure 19: Agentic trajectory for Case 3 (Part 2 of 6).
Figure 20: Agentic trajectory for Case 3 (Part 3 of 6).
Figure 21: Agentic trajectory for Case 3 (Part 4 of 6).
Figure 22: Agentic trajectory for Case 3 (Part 5 of 6).
Figure 23: Agentic trajectory for Case 3 (Part 6 of 6).
Figure 24: Agentic trajectory for Case 4 (Part 1 of 3).
Figure 25: Agentic trajectory for Case 4 (Part 2 of 3).
Figure 26: Agentic trajectory for Case 4 (Part 3 of 3).
Unified multimodal models provide a natural and promising architecture for understanding diverse and complex real-world knowledge while generating high-quality images. However, they still rely primarily on frozen parametric knowledge, which makes them struggle with real-world image generation involving long-tail and knowledge-intensive concepts. Inspired by the broad success of agents on real-world tasks, we explore agentic modeling to address this limitation. Specifically, we present Unify-Agent, a unified multimodal agent for world-grounded image synthesis, which reframes image generation as an agentic pipeline consisting of prompt understanding, multimodal evidence searching, grounded recaptioning, and final synthesis. To train our model, we construct a tailored multimodal data pipeline and curate 143K high-quality agent trajectories for world-grounded image synthesis, enabling effective supervision over the full agentic generation process. We further introduce FactIP, a benchmark covering 12 categories of culturally significant and long-tail factual concepts that explicitly requires external knowledge grounding. Extensive experiments show that our proposed Unify-Agent substantially improves over its base unified model across diverse benchmarks and real world generation tasks, while approaching the world knowledge capabilities of the strongest closed-source models. As an early exploration of agent-based modeling for world-grounded image synthesis, our work highlights the value of tightly coupling reasoning, searching, and generation for reliable open-world agentic image synthesis.
Shuang Chen, Quanxin Shou, Hangting Chen +16
University of California, Los Angeles · Tencent Hunyuan · The Hong Kong University of Science and Technology +1
Recent years have witnessed remarkable progress in image generation and editing, particularly regarding instruction following and visual fidelity. However, when handling ambiguous intentions, logical reasoning, and Out-of-Distribution (OOD) knowledge, existing image models often yield sub-optimal results due to a lack of deep reasoning capabilities and real-time external information. Although emerging unified understanding-and-generation models attempt to bridge this gap, they remain constrained by their intrinsic parameter scales and static knowledge gaps. Inspired by agentic paradigms, we propose RS-Gen: a plug-and-play, training-free, multi-stage image agentic framework. RS-Gen innovatively introduces a "Questioning-and-Solving" closed-loop mechanism to accurately identify logical issues and knowledge gaps, autonomously planning actions to bridge information deficits and execute deep logical reasoning. Extensive experiments demonstrate that RS-Gen significantly expands the capability boundaries of foundational image generation and editing models. Specifically, on the WISE Verified and RISEBench benchmarks, RS-Gen yields substantial absolute performance gains of 0.313 for Qwen-Image and 19.70 for Qwen-Image-Edit-2511, respectively, successfully elevating both to the state-of-the-art (SOTA) level among open-source models.
Natural language has become a powerful interface for image generation and editing, yet text-guided visual systems remain highly sensitive to prompt formulation. Semantically similar requests can produce different outputs depending on wording, specificity, and how explicitly visual constraints are stated, motivating prompt enhancement as a trainable component rather than a peripheral user choice. Existing strong enhancers often rely on large, proprietary LLMs such as ChatGPT or Gemini, adding cost, latency, and deployment dependence to the visual generation pipeline. We propose Agentic Prompt Enhancer (APE), a lightweight framework that post-trains small language models (SLMs) as prompt-enhancement agents. APE supports both single-agent rewriting and role-specialized multi-agent enhancement. Its single-agent instantiation, SAPE, rewrites the prompt in one pass, while its multi-agent instantiation, MAPE, decomposes enhancement into a router--rewriter--composer process for handling compositional constraints over objects, attributes, spatial relations, and edits. With task-aware rewards and post-training protocols, APE improves visual alignment and prompt following without modifying the downstream visual model. Experiments on challenging image generation and editing benchmarks demonstrate that post-trained small prompt enhancers reliably outperform their base counterparts, narrowing the gap to closed-source prompt enhancers; in addition, MAPE proves particularly strong on complex compositional tasks within these benchmarks.