MACRO: Structured Multi-Reference Image Data¶
Scope checked: 2026-09-04. MACRO is a research release for multi-reference image generation, not a new universal inference architecture. The authors publish structured data, a benchmark, training material, and checkpoints for selected open models so that a system can learn to use more than one reference image without relying solely on short-context training examples.
What Is Released¶
The project describes a 400K-sample data release with up to ten reference images per sample. It organizes examples across four complementary task families:
| Task family | What the references are intended to provide |
|---|---|
| Customization | subjects, identity, objects, or visual attributes |
| Illustration | multimodal context for a requested composition |
| Spatial | view, geometry, or scene relationships |
| Temporal | visual state across time |
MacroBench evaluates those scenarios, and the repository supplies model-specific paths for Bagel, OmniGen2, and Qwen-Image-Edit. The data and benchmark make a useful evaluation substrate; they do not establish that an arbitrary model accepts ten images or that it will preserve every reference faithfully.
Choose a Compatible Route¶
Use the exact route maintained by the project:
- choose one supported base model and its documented revision;
- follow that model's preprocessing, resolution budget, prompt syntax, and checkpoint loading path;
- use the matching MACRO checkpoint or training configuration rather than mixing adapter formats;
- keep a baseline run from the unmodified base model for the same task fixture.
The reference count changes the image-token budget. MACRO's implementation dynamically reduces input-image resolution as more references are supplied. Do not apply a fixed pixel budget from an old tutorial to a different base model; inspect the current configuration and measure whether critical details survive the resize.
Evaluate the Real Task¶
Multi-reference generation has several independent failure modes: subject mixing, lost small details, incorrect relationships, selective copying, and a plausible but unfaithful output. Build an acceptance fixture that names which reference controls which result attribute, then inspect the output against that mapping.
MacroBench uses an LLM-based scoring workflow. It can be a repeatable research signal, but it is not a replacement for human review, rights checks, or task-specific validators in a production workflow. Retain prompts, input ordering, model revision, random seed where applicable, and generated artifacts with each comparison.
Data, Rights, and Release Terms¶
The project documents a data-construction pipeline and provides released artifacts, but the provenance and terms of every input, base model, checkpoint, and downstream output remain separate questions. In particular:
- read the license and model card for the selected base model and MACRO artifact;
- verify whether the intended deployment and redistribution are allowed;
- use only reference images that the operator is authorized to process;
- do not infer commercial permission for every component from the availability of the repository.