cs.CVMar 26, 2026
SaveMACRO: Advancing Multi-Reference Image Generation with Structured Long-Context Data
Organizations: HKU MMLab · Meituan
Abstract
Generating images conditioned on multiple visual references is critical for real-world applications such as multi-subject composition, narrative illustration, and novel view synthesis, yet current models suffer from severe performance degradation as the number of input references grows. We identify the root cause as a fundamental data bottleneck: existing datasets are dominated by single- or few-reference pairs and lack the structured, long-context supervision needed to learn dense inter-reference dependencies. To address this, we introduce MacroData, a large-scale dataset of 400K samples, each containing up to 10 reference images, systematically organized across four complementary dimensions -- Customization, Illustration, Spatial reasoning, and Temporal dynamics -- to provide comprehensive coverage of the multi-reference generation space. Recognizing the concurrent absence of standardized evaluation protocols, we further propose MacroBench, a benchmark of 4,000 samples that assesses generative coherence across graded task dimensions and input scales. Extensive experiments show that fine-tuning on MacroData yields substantial improvements in multi-reference generation, and ablation studies further reveal synergistic benefits of cross-task co-training and effective strategies for handling long-context complexity. The dataset and benchmark will be publicly released.
Figures & tables
Figure 1: Overview of MacroData. MacroData contains 400K high-quality samples with up to 10 input images across four long-context multi-reference image generation tasks: (a) Customization : generating compositions conditioned on multiple reference images, (b) Illustration : producing illustrative images based on multimodal context, (c) Spatial : predicting novel view images given multiple views, specifically including outside-in objects and inside-out scenes, and (d) Temporal : forecasting future frames based on historical sequence. Each task is composed of 100k samples, split into different numbers of reference images, including 1-3, 4-5, 6-7, and 8-10 input images.
Figure 2: Statistics of MacroData. MacroData comprises four tasks, each containing 100k samples. (a) The number of input images in the customization subtask averaged 5.84 images per sample, with a maximum of 10. (b) The number of input images across all tasks averaged 5.44 images per sample. (c) Comparison among different datasets. (d) The distribution of data composition for each task.
Figure 3: Customization Subset Pipeline composites preprocessed metadata via rule-based and VLM-reasoned sampling and applies a bidirectional assessment to ensure reference fidelity and prompt consistency.
Figure 6: Temporal Subset Pipeline extracts video keyframes, groups them into coherent sequences using DINOv2, and utilizes a VLM to generate a summary prompt and quality score for predicting the final frame.
| Model | Customization | Illustration | Spatial | Temporal | Average | ||||
| 1-5 | 6-10 | 1-5 | 6-10 | 1-5 | 6-10 | 1-5 | 6-10 | ||
| Closed-source models | |||||||||
| Nano Banana Pro [ 9 ] | 9.41 | 7.60 | 9.12 | 8.88 | 3.28 | 3.20 | 8.21 | 7.25 | 7.12 |
| GPT-Image-1.5 [ 30 ] | 9.62 | 8.76 | 8.90 | 8.84 | 3.32 | 4.24 | 8.48 | 7.84 | 7.50 |
| Open-source models | |||||||||
| BAGEL [ 11 ] | 5.37 | 2.53 | 4.51 | 4.34 | 0.67 | 0.53 | 3.36 | 2.93 | 3.03 |
Table 1: Quantitative Comparison Results on MacroBench. “+ X” denotes the baseline model specifically fine-tuned on the corresponding dataset. Results of “1–3” and “4–5” are combined into “1–5”, and “6-7” and “8-10” are combined into “6-10”.
Figure 7: Qualitative Results on All Tasks. In each row, results of different datasets are fine-tuned from the same base model.
| Model | SINGLE | MULTIPLE | SCENE | Average | |||||
| Character | Object | Character | Object | Char. + Obj. | Character | Object | Char. + Obj. | ||
| Closed-source models | |||||||||
| Nano Banana Pro [ 9 ] | 9.07 | 9.43 | 8.91 | 9.23 | 8.79 | 8.94 | 8.61 | 8.70 | 8.96 |
| GPT-Image-1.5 [ 30 ] | 9.33 | 9.44 | 9.40 | 9.43 | 9.11 | 9.22 | 9.14 | 8.90 | 9.25 |
| Open-source models | |||||||||
| BAGEL [ 11 ] | 7.45 | 7.40 | 5.54 | 6.56 | 7.17 | 4.59 | 5.49 | 5.95 | 6.27 |
Table 2: Quantitative Comparison Results on OmniContext. “Char. + Obj.” denotes Character + Object. Black “+ Ours Customization” indicates training exclusively on customization data (similar to Echo4o [ 57 ] , MICo [ 45 ] , and OpenSubject [ 27 ] ), while gray “+ Ours All” indicates training on the full dataset. “ † ” denotes results reported in previous works, others are reproduced under setting in Sec. 5.1 .
| Model | Customization | Illustration | Spatial | Temporal | Average | ||||
|---|---|---|---|---|---|---|---|---|---|
| 1-5 | 6-10 | 1-5 | 6-10 | 1-5 | 6-10 | 1-5 | 6-10 | ||
| Bagel | 5.37 | 2.53 | 4.51 | 4.34 | 0.67 | 0.53 | 3.36 | 2.93 | 3.03 |
| Bagel + All | 8.92 | 7.12 | 5.69 | 5.56 | 3.32 | 3.48 | 6.23 | 5.40 | 5.71 |
| Bagel + Customization | 8.61 | 6.43 | 5.19 | 4.92 | 0.81 | 0.64 | 4.43 | 3.17 | 4.27 |
| Bagel + Illustration | 5.46 | 2.16 | 5.93 | 5.70 | 1.21 | 1.22 | 5.14 | 4.59 | 3.92 |
| Bagel + Spatial | 5.08 | 2.24 | 4.89 | 4.78 | 2.72 | 2.98 | 4.99 | 4.13 | 3.97 |
Table 3: Comparison of Models Trained on Different Data Subsets. “+ Customization/Illustration/Spatial/Temporal” refers to models trained exclusively on the corresponding subset, while “+ All” denotes the model trained on the complete dataset.
Figure 8: Impact of Data Ratio. (a) Performance curves on Customization tasks of our MacroBench. (b) Performance curves on Temporal tasks of our MacroBench. The x-axis represents the different image count categories (1–3/4–5/6–7/8–10). Each line represents a different data ratio for training data of different input counts.
Figure 9: Data Scaling Analysis. (a) Performance curves on the Customization task of our MacroBench. (b) Performance curves on OmniContext [ 48 ] . The x-axis represents the number of training steps. Each line represents different number of training samples.
Figure 10: Text-to-Image Data Ratio. (a) The performance curve on the subset of our MacroBench. (b) The performance curve on OmniContext [ 48 ] . The x-axis is the GenEval Score [ 15 ] for text-to-image generation evaluation. Each line represents different ratio of text-to-image data in training.
Figure 11: Token Selection Strategies. (a) Block-wise Selection, (b) Image-wise Selection, (c) Text-aligned Selection.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
| Judge | Setting | Pearson | Spearman | Kendall |
|---|---|---|---|---|
| Gemini-3-Flash [ 8 ] | Overall | 0.821 | 0.795 | 0.665 |
| Human | 0.770 | 0.798 | 0.613 | |
| GPT-4.1 [ 29 ] | Overall | 0.555 | 0.598 | 0.470 |
| Human | 0.447 | 0.510 | 0.354 |
Table 6: Judge–human agreement measured by Pearson ( ), Spearman ( ), and Kendall ( ) correlations. Overall : all 280 samples (GT human scores assigned 10). Human : non-GT annotated samples only (160 total). Best per column in bold .
| Model | 1–3 | 4–5 | 6–7 | 8–10 | Avg |
|---|---|---|---|---|---|
| Closed-source models | |||||
| Nano Banana Pro [ 9 ] | 9.58 | 9.24 | 8.18 | 7.01 | 8.50 |
| GPT-Image-1.5 [ 30 ] | 9.67 | 9.57 | 9.30 | 8.23 | 9.19 |
| Open-source models | |||||
| BAGEL [ 11 ] | 6.78 | 3.96 | 2.87 | 2.19 | 3.95 |
| BAGEL + Echo4o [ 57 ] | 7.89 | 6.01 | 4.71 | 3.24 | 5.46 |
Table 7: Detailed Results on Customization Task. Columns 1–3, 4–5, 6–7, 8–10 denote difficulty bins; Avg is the macro-average over all bins.
Figure 13: Qualitative Results of MacroBench Customization tasks for Bagel [ 11 ] fine-tuned on MacroData.
| Model | 1–3 | 4–5 | 6–7 | 8–10 | Avg |
|---|---|---|---|---|---|
| Closed-source models | |||||
| Nano Banana Pro [ 9 ] | 9.39 | 8.85 | 8.87 | 8.89 | 9.00 |
| GPT-Image-1.5 [ 30 ] | 9.41 | 8.39 | 8.77 | 8.91 | 8.87 |
| Open-source models | |||||
| BAGEL [ 11 ] | 4.66 | 4.35 | 4.54 | 4.14 | 4.42 |
| BAGEL + Echo4o [ 57 ] | 4.71 | 4.32 | 4.14 | 4.13 | 4.33 |
Table 8: Detailed Results on Illustration Task. Columns 1–3, 4–5, 6–7, 8–10 denote difficulty bins; Avg is the macro-average over all bins.
Figure 14: Qualitative Results of MacroBench Illustration tasks for Bagel [ 11 ] fine-tuned on MacroData.
| Model | 1–3 | 4–5 | 6–7 | 8–10 | Avg |
|---|---|---|---|---|---|
| Closed-source models | |||||
| Nano Banana Pro [ 9 ] | 3.09 | 3.47 | 3.34 | 3.05 | 3.24 |
| GPT-Image-1.5 [ 30 ] | 3.01 | 3.63 | 4.25 | 4.23 | 3.78 |
| Open-source models | |||||
| BAGEL [ 11 ] | 0.87 | 0.47 | 0.53 | 0.53 | 0.60 |
| BAGEL + Echo4o [ 57 ] | 0.88 | 0.63 | 0.90 | 0.70 | 0.78 |
Table 9: Detailed Results on Spatial Task. Columns 1–3, 4–5, 6–7, 8–10 denote difficulty bins; Avg is the macro-average over all bins.
Figure 15: Qualitative Results of MacroBench Spatial tasks for Bagel [ 11 ] fine-tuned on MacroData.
| Model | 1–3 | 4–5 | 6–7 | 8–10 | Avg |
|---|---|---|---|---|---|
| Closed-source models | |||||
| Nano Banana Pro [ 9 ] | 8.62 | 7.79 | 7.28 | 7.21 | 7.73 |
| GPT-Image-1.5 [ 30 ] | 8.73 | 8.23 | 8.06 | 7.62 | 8.16 |
| Open-source models | |||||
| BAGEL [ 11 ] | 3.90 | 2.82 | 3.13 | 2.72 | 3.14 |
| BAGEL + Echo4o [ 57 ] | 3.71 | 2.67 | 2.60 | 2.21 | 2.80 |
Table 10: Detailed Results on Temporal Task. Columns 1–3, 4–5, 6–7, 8–10 denote difficulty bins; Avg is the macro-average over all bins.
Figure 16: Qualitative Results of MacroBench Temporal tasks for Bagel [ 11 ] fine-tuned on MacroData.
Figure 17: Quantitative Results of different token selection strategies.
Figure 18: Failure Cases of Bagel [ 11 ] fine-tuned on our MacroData. Blue images are input references, green ones are ground truth, and red ones are generated failure results.
Figure 19: Visualization of MacroData Customization subset.
Figure 20: Visualization of MacroData Illustration subset.
Figure 21: Visualization of MacroData Spatial subset.
Figure 22: Visualization of MacroData Temporal subset.
Figure 23: Evaluation prompt for the Customization task.
Figure 24: Evaluation prompt for the Illustration task.
Figure 25: Evaluation prompt for the Spatial task.
Figure 26: Evaluation prompt for the Temporal task.
More in this topic
Image Generation
Image generation studies models that create or edit images from text, examples, layouts, and other structured controls.
See all 463 papers