LongBanana

An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis

Code · Download and use · License

90 tasks per release · 9 categories · 32 reference images per task

LongBanana evaluates image synthesis from long, heterogeneous reference contexts. Systems must retrieve relevant evidence, reject distractors, preserve identities and attributes, bind relations, and satisfy scene-wide constraints in one final image. Each task pairs a natural-language instruction with reference images and an atomic evaluation checklist grounded in the task's entity–relation graph.

LongBanana-Gen supports agentic image generation; LongBanana-Eval scores individual checklist obligations using a direct visual model or an agentic Judge. The accompanying code supports sandboxed OpenClaw, Codex, and Claude Code workflows. This repository contains the datasets, not model weights or agent runtimes.

Dataset Releases

Release	Cases	References	Evaluation checks
open_v1	90	2,880	4,670
open_v1p5	90	2,880	5,170

open_v1p5 extends open_v1 with 90 additional tasks featuring non-overlapping content. Both releases share the same data format, task taxonomy, and generation and evaluation workflows, allowing users to expand their experiments with the same pipeline. Select a release with DATASET_MANIFEST. Each release has ten tasks per category.

Task Taxonomy

Suite	Category	Capability	Tasks per release
Scene Composition	SCC	Semantic Coverage & Consistency	10
Scene Composition	RGC	Relational Graph Consistency	10
Scene Composition	LAC	Layout Arrangement Consistency	10
Visual Appearance	GAC	Global Appearance Conditioning	10
Visual Appearance	LAB	Local Appearance Binding	10
Visual Appearance	AHC	Appearance Harmonization & Coherence	10
Reference Reasoning	SEL	Selection & Retrieval	10
Reference Reasoning	TSR	Temporal / State Reasoning	10
Reference Reasoning	AGG	Counting & Aggregation	10

Categories identify the dominant task demand, not mutually exclusive visual content. Evaluation spans Content Fidelity, Relational Binding, and Global Constraint Satisfaction.

Download and Use

Download open_v1.tar.zst or open_v1p5.tar.zst, together with SHA256SUMS. A tar implementation with Zstandard support extracts each archive into its own release directory:

sha256sum --ignore-missing -c SHA256SUMS
tar --zstd -xf open_v1.tar.zst

For the other release, extract open_v1p5.tar.zst instead. After setting up the LongBanana code, validate the artifact and use its manifest in the experiment configuration:

export DATASET_MANIFEST="$PWD/open_v1/manifest.json"
python -m longbanana validate-data "$DATASET_MANIFEST" --evaluation

These are file-based benchmark bundles; use the LongBanana loader rather than assuming a tabular datasets.load_dataset() configuration. The open-1018 code branch is the sandboxed implementation for both releases; legacy preserves the earlier code. See the code repository for branch publication status and execution instructions.

Data Structure

File or asset	Contents
manifest.json	Release identity, revision, file paths, counts, and checksums
public/tasks.jsonl	One generation task per line
Referenced image files	Ordered reference inventory; public backgrounds in open_v1 only
evaluation/checks.jsonl	Atomic check text, answer options, and fixed reference bindings
LICENSE.txt, ATTRIBUTION.md	Dataset license and attribution

open_v1p5 additionally includes evaluation/binding_reviews.jsonl, binding metadata outside the Judge task inputs. Image paths are relative to the artifact root; filenames, reference order, and SHA-256 hashes are part of the data contract.

A generation task has case_id, category, prompt, an ordered references array, and an optional background. Each reference has an id, path, and sha256. A public background is additional to the 32 reference images and is placed first in generation inputs.

An evaluation check has case_id, check_id, category, question, guide, options, option_desc, and ordered reference_ids. An empty reference binding requires a candidate_only_reason. The dataset fixes the supplied reference set; an agentic Judge may decide which of those images it needs to open.

Generation agents receive only the instruction and public images. Evaluation checklists and binding metadata are not generation inputs. The archives contain no simulated targets, expected scores, human judgments, credentials, or experiment traces. Candidate images are supplied by the system being evaluated.

Evaluation Protocol

For each check, judge the candidate against its requirement and bound references. Use Fail = 0, Uncertain = 1, and Pass = 2. Invalid responses are not judgments and have no numeric score.

Case score = 100 × mean(check scores) / 2
Category score = mean(case scores in that category)
Overall score = mean(all case scores)

This is unweighted case-level aggregation, not pooling all checks across cases. Report coverage and invalid counts; incomplete scores are provisional. Record the release and revision, generation and Judge configurations, and any explicit missing-value or difficulty-weighting policy. Historical paper tables use their stated aggregation and cohort, not necessarily this unweighted summary.

Dataset Construction

open_v1 retains its source instructions, image bytes, and check definitions. open_v1p5 uses graph-designed tasks with synthetic reference-image production and iterative refinement. Its check requirements are projected into the shared Judge format using the versioned graph-check-projection-1 profile. Each check has fixed reference bindings or an explicit candidate-only justification.

Intended Use

Use the datasets to study long-context multi-reference synthesis, reference retrieval and binding, iterative agentic generation, and check-level visual evaluation.

License and Attribution

Both dataset releases are licensed under Creative Commons Attribution 4.0 International. See LICENSE.txt for the full terms and ATTRIBUTION.md for attribution information. These files are also included inside each archive. Sharing and adaptation, including commercial use, are permitted under the license; retain attribution, link the license, and indicate modifications. This does not grant third-party privacy, publicity, patent, or trademark rights.

The accompanying code is separately licensed under MIT. Brand marks retain their owners' rights; artwork sources describe the figures and resource badges. Scholarly citation is appreciated and is not an additional license condition.

Citation

arXiv Paper: Coming soon. The paper link will be added when available.

@misc{cao2026longbanana,
  title={LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis},
  author={Cao, Haoxiang and Zhang, Yuxuan and Du, Penghui and others},
  year={2026},
  url={https://github.com/Hooccee/LongBanana_open}
}