MACRO: Advancing Multi-Reference Image Generation with Structured Long-Context Data

1. Problem Statement

Current state-of-the-art in-context image generation models (e.g., Bagel, OmniGen2) excel at single-reference or few-reference tasks (e.g., identity preservation, style transfer) but suffer from severe performance degradation when the number of input reference images increases beyond 3–5.

• Root Cause: The authors identify a fundamental data bottleneck. Existing datasets are dominated by single- or few-reference pairs and lack structured, long-context supervision. Consequently, models fail to learn the dense inter-reference dependencies required for complex reasoning, such as maintaining spatial consistency across multiple views, tracking temporal dynamics in video sequences, or composing multiple distinct subjects into a coherent scene.
• Evaluation Gap: There is a lack of standardized benchmarks capable of evaluating generative coherence across varying input scales (1 to 10 images) and diverse task dimensions (customization, illustration, spatial, temporal).

2. Methodology

A. MacroData: A Large-Scale Structured Dataset

The core contribution is MacroData, a dataset comprising 400,000 high-quality samples, with each sample containing up to 10 reference images. The dataset is systematically organized into four complementary dimensions to cover the full spectrum of multi-reference generation:

1. Customization (100k samples): Focuses on composing multiple referents (humans, objects, scenes, clothing, styles) into coherent original scenes.
   • Pipeline: Aggregates metadata from diverse sources (OpenSubject, MVImgNet, DL3DV, etc.). Uses VLMs to filter for identity leakage and aesthetic quality. Synthesizes target images using closed-source models, followed by bidirectional VLM assessment to ensure input fidelity and prompt consistency.
2. Illustration (100k samples): Generates context-aware images to complement interleaved text narratives.
   • Pipeline: Extracts "anchor images" from massive interleaved corpora (OmniCorpus). Uses VLMs to rewrite noisy preceding contexts into concise, coherent narratives and filters out samples with poor semantic alignment.
3. Spatial (100k samples): Synthesizes novel viewpoints from multiple input views.
   • Subtasks:
     • Outside-in: Multi-view 3D object renderings (Objaverse).
     • Inside-out: Panoramic scene generation (DIT360, Pano360).
   • Pipeline: Samples canonical views, applies strict spatial overlap filters to ensure physical plausibility, and designates specific target views for generation.
4. Temporal (100k samples): Forecasts future keyframes based on historical video sequences.
   • Pipeline: Extracts keyframes from video clips (OmniCorpus-YT) using shot boundary detection and DINO-based sequence grouping. VLMs generate summary prompts and quality scores to ensure temporal coherence.

Data Construction Strategy: Unlike previous works relying on noisy web scraping, MacroData employs a distillation pipeline using advanced closed-source models (e.g., Nano Banana Pro, Gemini-3-Flash) for synthesis and rigorous VLM-based filtering to ensure high fidelity and logical consistency.

B. MacroBench: Standardized Evaluation Benchmark

To address the lack of evaluation protocols, the authors propose MacroBench, containing 4,000 samples (1,000 per task).

• Structure: Evaluates performance across the four task dimensions and graded input scales (1–3, 4–5, 6–7, 8–10 images).
• Evaluation Mechanism: Uses an LLM-as-Judge paradigm (specifically Gemini-3-Flash, selected for its superior correlation with human judgment compared to GPT-4.1).
• Metrics: Task-specific metrics (e.g., Image Consistency Score, Prompt Following Score, View Transformation Score) are aggregated using a geometric mean to penalize severe failures in any single dimension.

C. Training and Optimization Strategies

• Fine-tuning: The authors fine-tune open-source models (Bagel, OmniGen2, Qwen-Image-Edit) on MacroData.
• Token Selection: To handle the computational overhead of long contexts (up to 10 images), they explore token selection strategies. Text-aligned selection (pruning tokens based on text-image attention scores) and Block-wise selection proved effective, outperforming simple image-wise selection which caused information loss.
• Data Scaling: Experiments show that performance scales consistently with data volume, with the most significant gains occurring between 1K and 10K samples.

3. Key Contributions

1. MacroData: The first large-scale dataset specifically designed for long-context multi-reference image generation, featuring 400k samples with up to 10 inputs, covering Customization, Illustration, Spatial, and Temporal tasks.
2. MacroBench: A comprehensive benchmark that fills the evaluation void, allowing for systematic assessment of model performance across varying input scales and task complexities.
3. Empirical Validation: Demonstrated that fine-tuning on MacroData significantly narrows the performance gap between open-source models and closed-source SOTA models (e.g., Nano Banana Pro, GPT-Image-1.5).
4. Insights on Long-Context Learning: Identified that cross-task co-training yields synergistic benefits and that specific token selection strategies are crucial for managing long-context complexity without sacrificing generation quality.

4. Experimental Results

• Quantitative Performance:
  • Models fine-tuned on MacroData (e.g., Bagel + MacroData) achieved an average score of 5.71 on MacroBench, ranking third overall behind only closed-source models (Nano Banana Pro: 7.12, GPT-Image-1.5: 7.50) and vastly outperforming base open-source models (Bagel: 3.03).
  • Robustness: MacroData-trained models showed significantly less performance degradation as input counts increased from 1–5 to 6–10 images compared to baselines.
  • Generalization: The models also achieved strong results on the existing OmniContext benchmark, outperforming datasets specifically built for that task (e.g., Echo4o).
• Qualitative Results: Visualizations show that MacroData-trained models can successfully integrate features from 10+ images, synthesize novel 3D views with high geometric consistency, and maintain temporal coherence in video prediction, whereas baselines often fail to preserve identity or spatial logic with high input counts.
• Ablation Studies:
  • Cross-task synergy: Training on the full dataset outperformed training on individual subsets, suggesting that learning diverse inter-reference dependencies benefits all tasks.
  • Data Ratio: Progressive tasks (Customization) benefit from upweighting high-input samples (2:2:3:3 ratio), while non-progressive tasks perform best with uniform sampling.

5. Significance

This paper addresses a critical bottleneck in the field of generative AI: the inability of current models to reason over dense, multi-image contexts. By providing a structured dataset and a rigorous benchmark, MacroData and MacroBench:

• Enable the development of models capable of complex real-world applications like multi-subject composition, narrative illustration, and 3D reconstruction.
• Establish a new standard for evaluating long-context visual reasoning, moving beyond simple few-shot customization.
• Provide actionable insights (e.g., token selection, data scaling ratios) for researchers aiming to build more capable, context-aware generative models.

The work suggests that the path to advanced in-context generation lies not just in better architectures, but in structured, high-quality data that explicitly teaches models to understand the relationships between multiple visual inputs.