v0.3.0

PaperBanana v0.3.0

The feature release that closes out the entire issue tracker — every open issue and PR resolved. Highlights below; the dataset, benchmarking, and venue work make this the most research-credible release yet.

Highlights

Official benchmarking — paperbanana benchmark --split test runs the official PaperBananaBench test split (292 diagram / 240 plot cases, with the official per-example aspect ratios), and --mode vanilla provides the no-pipeline ablation baseline — results are now directly comparable to the paper (arXiv:2601.23265).

Dataset, one command — paperbanana data download fetches the full reference set + test split from a project-controlled mirror (bench-data-v1, credit to the upstream authors) with SHA256 verification. The broken curated option is gone (#200), and Unicode-filename entries import correctly on macOS (#234).

Multi-candidate generation (#237) — --num-candidates N (1–8): planning runs once, refinement fans out in parallel with seed offsets. Cost estimates and --budget account for the fan-out; one failed candidate never kills the rest.

Polish mode (#238) — paperbanana polish --input figure.png: a VLM audits your existing figure against the venue style guide and proposes up to 10 concrete improvements, then a guided image edit applies them to the original figure (Gemini image-conditioned editing). Failed refinement rounds throughout the pipeline now roll back to the previous best image instead of shipping a degraded result.

Sketch input (#223) — --image sketch.png (repeatable): hand-drawn sketches, whiteboard photos, or prior figures guide the Planner alongside retrieved exemplars. The Critic stays blind to the sketch and judges against your text only.

Venue style packs (#89) — venues are now a directory convention: built-in NeurIPS/ICML/ACL/IEEE packs plus user packs under ~/.config/paperbanana/venues/ (paperbanana venues init <name> scaffolds one; venues list shows all). venue.yaml carries per-venue defaults (aspect ratio, fonts). All venue format facts — column widths, fonts, caption rules — are grounded in the venues' current official author instructions with cited sources (#250), and two-column venues default to column-friendly 4:3 figures.

Style-guide synthesis (#236) — paperbanana guidelines synthesize generates a corpus-grounded style guide from any reference set via VLM map-reduce (hex palettes, line/shape semantics, domain-conditional rules), budget-guarded.

Distribution (#239) — "Open in Colab" quickstart notebook (free-tier Gemini path) and a slim Dockerfile.

Behavioral changes

• Unknown --venue names now error with the list of available venues (previously fell back silently to NeurIPS).
• --venue icml|acl|ieee runs default to 4:3 aspect ratio via venue.yaml (override with --aspect-ratio).
• Venue font preferences now reflect each venue's actual mandated Times-family fonts.
• DatasetManager.download() no longer takes a dataset parameter (the curated option was removed).

Quality

Suite grew to 909 passing tests, including end-to-end coverage for every feature above. MCP tool surface (11 tools) is pinned by smoke tests.

Contributors

Thanks to @xraywu, @haosenwang1018, @YB0y, @JUSTSUJAY, @statxc, @simplaj, and everyone who filed the issues this release closes.