Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing

TL;DR

Pico-Banana-400K is an Apple-released ~400K-triplet (image, edit instruction, edited image) dataset for text-guided image editing, distilled from Nano-Banana (Gemini-2.5-Flash-Image) over real OpenImages photos. The pipeline uses Gemini-2.5-Flash to write a long detailed prompt per image (Qwen2.5-7B-Instruct produces a parallel short user-style version), Nano-Banana to edit, and Gemini-2.5-Pro as a four-criterion judge whose ~0.7 threshold separates successes from failures. The dataset is structured by a 35-edit-type taxonomy in 8 categories, with three subsets — 258K single-turn SFT, 56K preference pairs (success vs failure on the same input), and 72K multi-turn sequences with coreferential instructions. Cost is ~$100K; license is CC BY-NC-ND 4.0. The paper’s most useful artifact for the team is its per-edit-type success-rate table, which exposes what Nano-Banana can and cannot do reliably.

Key claims

The dataset is distilled from one frontier editor (Nano-Banana) over real OpenImages photos, with a fine-grained 35-type taxonomy in 8 categories (Pixel & Photometric, Object-Level Semantic, Scene Composition, Stylistic, Text & Symbol, Human-Centric, Scale, Spatial/Layout) and explicit per-category sample counts [Table 1].
Each example carries two parallel instructions: a long, content-aware detailed prompt from Gemini-2.5-Flash and a short, user-style summary produced by Qwen2.5-7B-Instruct conditioned on human-written in-context examples — supporting controlled study of prompt-granularity effects [§2.2].
Gemini-2.5-Pro serves as automated judge using a fixed four-criterion rubric — Instruction Compliance (40%), Seamlessness (25%), Preservation Balance (20%), Technical Quality (15%) — aggregated to a single score; a strict ~0.7 threshold splits the 258K successes from the 56K failures retained as preference negatives [§2.3].
The multi-turn subset is built by sampling 100K single-turn items, extending each with 1–4 additional edit types (yielding 2–5-turn sessions), and prompting Gemini-2.5-Pro to write instructions that coreferentially link to prior edits (e.g. “change the color of it”) so chains exercise both compositionality and pragmatic reference [§2.4].
Per-edit-type success rates are reported and exhibit a consistent pattern: global photometric/stylistic edits succeed most often (artistic style transfer, film grain/vintage, modern↔historical restyle), object-level remove/replace are mid-range, and fine spatial control / layout extrapolation / typography are the weakest — relocate-object, change size/shape/orientation, outpainting, font-style change, and Pixar/Disney-3D stylization all fail substantially more often [§3, Fig. 6].
The dataset positions itself against synthesizing-from-models predecessors (Echo-4o-Image 180K, GPT-Image-Edit-1.5M) on source-image realism (Real photos from OpenImages, not synthetic) and against real-pair predecessors (GIER, MagicBrush 10K, UltraEdit, OmniEdit, HQ-Edit) on scale and dual instruction format [Table 3].
Quality-driven scope decisions are documented: brightness/contrast/saturation edits and sharpen/blur were excluded because Nano-Banana produced negligible or unstable visual change relative to source; two-image composition edits were excluded because results were not reliable enough for training pairs [§2.1].
Total stated production cost is approximately $100K USD, dataset released under CC BY-NC-ND 4.0 (research-only, no derivatives), source images bound by the OpenImages CC BY 2.0 license; due to legal constraints source images are not redistributed — users download OpenImages packed files and join via the provided URL mapping [§1, repo README].

Method

The pipeline is a two-stage MLLM cascade with an automatic judge. For each OpenImages source image, Gemini-2.5-Flash is prompted (with a system message demanding content-aware grounding in visible objects/colors/positions) to write a long, detailed editing instruction. Qwen2.5-7B-Instruct, given human-written instruction exemplars as in-context demonstrations, rewrites the long instruction into a concise user-style command. Nano-Banana (Gemini-2.5-Flash-Image) then executes the edit on the source image. Gemini-2.5-Pro inspects the (source, instruction, edit) triplet and emits a structured score over four weighted dimensions; the aggregated scalar is compared to a strict ~0.7 threshold. Above-threshold edits enter the single-turn SFT corpus (~258K); below-threshold edits are retained as the negative member of preference pairs (~56K), explicitly for DPO / reward-model training. For the multi-turn subset, 100K single-turn samples are extended with 1–4 additional turns; Gemini-2.5-Pro writes the per-turn instructions conditioned on image + edit-type history, with explicit instruction to use coreferential language tying back to earlier edits, and the same per-turn Nano-Banana + judge step is applied. Human-centric edit types apply category-specific filtering so face/clothing/pose edits are only attempted on appropriate images.

Results

Scale and composition: ~258K single-turn SFT triplets, ~56K success/failure preference pairs on the same source-instruction inputs (the failures retried automatically and preserved), ~72K multi-turn sessions averaging 2–5 turns each [§1].
Per-edit-type success-rate ordering [§3, Fig. 6]: strong artistic style transfer is the most reliable Nano-Banana operation; film grain/vintage, modern↔historical restyle, photo↔cartoon/sketch, and seasonal change are in the next band; object remove/replace are mid-band; relocate-object is the weakest object operation; change size/shape/orientation similarly low; outpainting struggles with boundary continuity; change-font-style is the single lowest-success edit type, with translate/replace/add text only modestly better; Pixar/Disney-3D and caricature stylizations of people produce identity drift.
Taxonomy coverage [Table 1]: 35 edit types instantiated, with the largest count buckets being Add-film-grain (15,443), Add-new-scene-context (14,830), Strong-artistic-style-transfer (15,285), Modern↔historical (14,856), Add-new-object (14,190), Remove-object (15,111), Replace-object-category (14,549), Change-object-attribute (13,813), and Zoom-in (13,729). Text-edit and human-centric buckets are smaller (often 1–4K) reflecting category-eligibility filtering.
Reported total production cost: approximately $100K USD across generation, editing, and judging passes [§1, footnote 1].

Why it’s interesting

Pico-Banana-400K is the cleanest filed instance of the “frontier editor + frontier judge + permissive image source = open editing-data corpus” recipe — it slots directly into the same lineage as Ditto/Editto for video editing (Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset (Ditto / Editto)) and Señorita-2M (Señorita-2M: A High-Quality Instruction-based Dataset for General Video Editing by Video Specialists) on the image side, but with Nano-Banana (Gemini-2.5-Flash-Image) as both generator and Apple’s distillation target rather than a stable of expert specialists. The preference-pair subset is the most directly reusable artifact: it generates the success/failure pairs by running the same prompt on the same source image until one passes the judge and one does not, which is structurally what reward models like EditReward (EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing) need for training but historically had to be hand-collected. The per-edit-type success-rate breakdown also serves as an honest weakness map for Nano-Banana itself — text rendering and fine spatial control are quantified failure modes, not just anecdotes — and aligns with the same “VLMs are global-stylization-strong, geometry-weak” pattern documented from the consumer side in tweets like Astropulse: sending special noise patterns alongside images to manipulate FLUX Kontext. The methodological move is identical to Synthetic Training Data‘s thesis: the verification gate (Gemini-2.5-Pro at threshold 0.7) is the load-bearing piece, not the raw scale.