Generative retrieval has demonstrated strong performance by directly generating product semantic identifiers (SIDs). Extending this paradigm to image search, however, is nontrivial because real-world query images contain diverse information, including the search target, useful auxiliary evidence, and irrelevant visual content. This requires the model to identify and focus on the search target while selectively utilizing auxiliary evidence. In this paper, we propose PailitaoGR, a Latent Think-with-Images method for generative image retrieval, which internalizes target-focused perception and selective auxiliary-evidence utilization into a the generative retrieval model, enabling Zooming without Cropping and Reading without OCR. Specifically, we design a target-focused perception mechanism that identifies and enhances visual tokens of the search target, consisting of a target Enhancer and a learning strategy based on on-policy distillation and attention guidance loss, enabling the model to focus on search-target regions. We also design a selective auxiliary-evidence utilization mechanism that identifies and enhances visual tokens of auxiliary evidence, including an auxiliary enhancer and an in-capacity incremental contrastive distillation strategy, enabling the model to exploit auxiliary evidence. We construct training and validation sets sampled from real-world online image-search logs. Experiments show that our method outperforms existing baselines by an average of 13.8%, validating its effectiveness.
Figures & tables
Figure 1. Comparison of direct, tool-assisted, and latent Think-with-Images paradigms for generative image retrieval. Our approach internalizes tool-enabled visual capabilities without extra inference latency.
Figure 2. Framework of our method.
Method
Input
CLICK
PURCHASE
H@1
H@5
H@10
R@5
R@10
H@1
H@5
H@10
R@5
R@10
Traditional Retrieval
DINOv3
Crop
36.38
58.61
64.36
32.60
39.87
14.64
36.98
44.35
35.84
43.27
CLIP
Crop
30.57
53.90
61.69
28.83
36.73
12.24
32.57
41.11
31.54
40.05
Existing Generative Retrieval
IRGen
Crop
29.72
55.02
60.06
28.87
36.03
11.54
32.81
39.85
32.00
39.14
Table 1. Comparison of retrieval performance (%) on the test set.
ROT Ratio
Method
CLICK
PURCHASE
H@1
H@5
H@10
R@5
R@10
H@1
H@5
H@10
R@5
R@10
0–5%
SFT Baseline
18.01
36.53
42.42
19.12
26.44
6.23
20.71
29.12
20.51
28.85
Ours
36.53
65.49
72.05
37.82
49.92
15.32
42.59
54.38
41.83
53.85
Δ
+18.52
+28.96
+29.63
+18.71
+23.49
+9.09
+21.89
+25.25
+21.31
+25.00
5–10%
SFT Baseline
26.02
50.52
57.25
26.86
35.88
9.85
29.06
38.11
28.71
37.81
Ours
42.11
75.18
81.18
45.20
57.25
17.29
50.68
62.05
49.66
61.30
Table 2. Retrieval performance (%) under different search-target ROT area ratios. Δ denotes the absolute improvement of our method over the SFT Baseline.
OCR
Method
CLICK
PURCHASE
H@1
H@5
H@10
R@5
R@10
H@1
H@5
H@10
R@5
R@10
With OCR
SFT Baseline
24.50
49.84
57.40
25.69
34.87
9.45
28.76
38.04
27.87
37.20
Crop Teacher
32.95
63.66
70.07
34.88
45.61
12.53
39.96
49.76
38.92
48.93
Ours
37.95
70.84
77.22
40.56
52.19
15.37
46.54
57.71
45.46
56.73
ΔSFT
+13.44
+21.00
+19.82
+14.87
+17.32
+5.92
+17.78
+19.68
+17.59
+19.53
ΔCrop
+5.00
+7.18
+7.15
+5.67
+6.58
+2.84
+6.58
+7.96
+6.54
+7.80
Table 3. Retrieval performance (%) on queries with and without OCR information. ΔSFT and ΔCrop denote the absolute improvements of our method over the SFT Baseline and Crop Teacher, respectively.
Figure 3. Visualization of answer-to-visual attention. From left to right: query image, attention map of the SFT Baseline, and attention map of our model.
Method
Components
CLICK
PURCHASE
Target
Auxiliary
H@10
R@10
H@10
R@10
SFT Baseline
63.24
40.74
44.57
44.05
+ Focusing Internalization
✓
76.53
52.93
57.33
56.75
PailitaoGR
✓
✓
79.19
55.27
60.12
59.48
Table 4. Ablation study of the core components in PailitaoGR.
Method
Objective
CLICK
PURCHASE
OPD
ROT
Entropy
H@10
R@10
H@10
R@10
w/o OPD
✓
✓
73.05
49.40
53.94
53.25
w/o ROT
✓
✓
73.47
50.22
54.91
54.27
w/o Entropy
✓
✓
75.64
52.07
56.72
55.98
Full
✓
✓
✓
76.51
52.93
57.52
56.86
Table 5. Ablation of target-focused perception mechanism.
Method
Criterion
CLICK
PURCHASE
Utility
Accessibility
H@10
R@10
H@10
R@10
Direct Auxiliary Transfer
77.04
53.05
57.84
57.08
+ Utility
✓
78.71
54.91
59.70
59.06
+ Accessibility
✓
✓
79.19
55.27
60.12
59.48
Table 6. Ablation of selective auxiliary-evidence utilization mechanism
Composed image retrieval requires identifying a target image from a gallery by integrating a reference image with a textual modification instruction. In a training-free zero-shot setting, this task relies on constructing a retrieval-oriented textual query within a frozen vision--language embedding space at inference time. Existing approaches predominantly rely on a single-pass generation strategy that fuses the reference context and modification text into a unified description. This strategy makes it difficult to detect or correct semantic distortions and omissions during generation. Consequently, the preservation of reference attributes and the integration of textual requirements interfere with each other, which degrades retrieval precision. To address these challenges, we introduce PEC-CIR, a training-free framework that structures query construction as a multi-stage reasoning pipeline. The framework operates through a Planner--Executor--Critic architecture where the Planner extracts explicit constraints, the Executor generates multiple candidate target descriptions, and the Critic evaluates these candidates based on constraint compliance. By reframing query construction as a staged inference process instead of a single-pass output, PEC-CIR reduces the propagation of generative errors by explicitly evaluating candidate queries before retrieval, thereby improving retrieval stability.
Gunho Jung, Jeong-Woo Park, Seon Bin Kim +1
Department of Artificial Intelligence, Korea University, Seoul, 02841, Republic of Korea
Text-to-image models can generate visual concepts that their prompts never mention, and recent work shows that such associations can be implanted so that ordinary prompts trigger them. Auditing this behavior requires recovering the description that a prompt omits. Existing prompt optimizers are not well suited to this task. Gradient-based methods require backward access to the scoring encoder. Discrete evolutionary search treats semantically unrelated tokens as equally close. Language-model-guided search requires one inference call per candidate, which limits exploration. This paper presents EmbGA, an embedding-space genetic algorithm that recovers concepts present in a set of images but absent from the prompt that generated them. EmbGA keeps the population structure of genetic search but applies its operators in the token embedding space. Crossover interpolates between parent embeddings, and mutation moves an embedding toward a randomly drawn vocabulary embedding. Each offspring is then projected back to discrete tokens before scoring, so every evaluated candidate is a valid token sequence. We evaluate EmbGA on two benchmarks against six optimizers under two exploration budgets. EmbGA achieves the highest alignment when the exploration budget is large, recovers the most grounded tokens on natural images, and runs much faster than language-model-guided search.
Recent image generation models have shown strong capabilities in generating high-fidelity and photorealistic images. However, they are fundamentally constrained by frozen internal knowledge, thus often failing on real-world scenarios that are knowledge-intensive or require up-to-date information. In this paper, we present Gen-Searcher, as the first attempt to train a search-augmented image generation agent, which performs multi-hop reasoning and search to collect the textual knowledge and reference images needed for grounded generation. To achieve this, we construct a tailored data pipeline and curate two high-quality datasets, Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k, containing diverse search-intensive prompts and corresponding ground-truth synthesis images. We further introduce KnowGen, a comprehensive benchmark that explicitly requires search-grounded external knowledge for image generation and evaluates models from multiple dimensions. Based on these resources, we train Gen-Searcher with SFT followed by agentic reinforcement learning with dual reward feedback, which combines text-based and image-based rewards to provide more stable and informative learning signals for GRPO training. Experiments show that Gen-Searcher brings substantial gains, improving Qwen-Image by around 16 points on KnowGen and 15 points on WISE. We hope this work can serve as an open foundation for search agents in image generation, and we fully open-source our data, models, and code.