Nano-consistent-150K — Identity-Consistent Image Editing Dataset Built with Nano-Banana

TL;DR

Nano-consistent-150K is a 159,492-sample synthetic dataset for identity-consistent portrait editing, constructed by running Google’s Nano-Banana image model against ~4,500 base portrait identities (Pixabay photos + anime illustrations) across six editing sub-tasks (Action, Background, Hairstyle, Temporal, Human Interaction, OOTD). The pipeline filters base identities with GPT-5-mini for clear facial features, generates Nano-Banana outputs from task-specific instructions, then re-runs GPT-5-mini as a quality gate to reject instruction non-compliance and cut-and-paste artifacts. Each portrait gets >35 distinct edits, so the dataset is dense per-identity rather than wide-and-shallow — explicitly designed to teach unified image-editing models to preserve facial identity across action, scene, era, and outfit changes. Authors are releasing the dataset openly plus a Qwen-Image LoRA fine-tuned on it.

Key claims

Nano-Banana substantially outperforms GPT-4o and the strongest open-weight image model (Qwen-Image) on the specific axis of facial identity preservation across edits — this gap is the dataset’s motivation [§Introduction].
The pipeline uses a two-stage GPT-5-mini gate: first to filter source portraits down from ~10K Pixabay images to ~4,000 with clear facial features, then to filter generated outputs to remove instruction non-compliance and visible cut-and-paste artifacts [§Data Construction Pipeline].
Instructions are issued in two formats per sample — a training-oriented long-form description and a user-oriented single-sentence command — so the same edit supervises both robust caption-style training and concise prompt-style inference [§Data Construction Pipeline].
Dataset composition: 120K single-image editing samples + 40K multi-reference generation samples = 159,492 total, spanning eight sub-tasks (six explicitly described: Action, Background, Hairstyle, Temporal, Human Interaction, OOTD) [§Dataset Statistics].
“Anchoring on consistent human identities enables construction of interleaved data that seamlessly link multiple editing tasks, instructions, and modalities around the same individual” [§Introduction] — the >35-edits-per-portrait density is what the authors argue distinguishes this from breadth-first synthetic editing corpora.
Temporal sub-task places the same identity in different historical contexts, requiring clothing and background to co-vary with the era (e.g. 1905 daily life, year-2000 millennial settings) — a control axis most editing datasets don’t cover [§Temporal Task].

Method

The pipeline has four stages. (1) Source identity filtering: ~10K Pixabay portraits run through GPT-5-mini, gated on clear facial features, yielding ~4,000 retained portraits plus 500 manually-collected anime characters. (2) Instruction drafting: task-specific instruction templates are hand-designed per sub-task — e.g. ~250 scene locations for the Background task, action templates like “make a Yes gesture” for the Action task. (3) Nano-Banana generation: instructions paired with source portraits are fed to Nano-Banana to produce edited outputs. (4) GPT-5-mini quality gate + instruction rewriting: outputs are filtered for instruction non-compliance and cut-and-paste artifacts; surviving samples are paired with two reformulated instruction styles (training-oriented detailed; user-oriented concise) conditioned on the input image and the generated result.

The Human Interaction sub-task is structurally different from the other five: instead of editing a single identity, it samples 2–4 portraits from the base identity set and uses GPT to draft interaction instructions (e.g. four people forming a band, two people drinking coffee), then Nano-Banana synthesizes a composite scene preserving all selected identities.

Results

The page reports dataset statistics but no quantitative model-training results — those are deferred to a forthcoming Qwen-Image LoRA release.

159,492 total samples across six described sub-tasks (Action, Background, Hairstyle, Temporal, Human Interaction, OOTD) [§Dataset Statistics].
120K single-image edits + 40K multi-reference generations [§Dataset Statistics].
Per-identity density: each retained portrait yields >35 distinct editing outputs [§Introduction].
~250 scene locations covered in the Background sub-task; ~4,500 base identities (4,000 Pixabay + 500 anime) [§Background Task, §Data Construction Pipeline].
No reported FID, identity-preservation, or downstream-model-improvement numbers — claims are pipeline-and-scale, not benchmark.

Why it’s interesting

This is a “frontier-model-as-data-engine” recipe of exactly the shape Synthetic Training Data tracks: a strong closed model (Nano-Banana) generates training data for a weaker open one (Qwen-Image, via LoRA), with an LLM gate (GPT-5-mini) as the load-bearing verification step. The methodology is closest to Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset (Ditto / Editto) (Ditto/Editto), which used a similar frontier-model + filter loop to scale instruction-based video editing supervision — Nano-consistent-150K does the equivalent for identity-consistent image editing, with a sharper specialization (>35 edits per portrait) that targets the identity-preservation failure mode specifically. It also complements Magic Fixup: Streamlining Photo Editing by Watching Dynamic Videos, which manufactured paired editing supervision from video-frame warping (a non-synthetic-generator path to the same training-data goal). The eventual Qwen-Image LoRA release will be the cleaner test of whether this dataset actually transfers — at filing time, only the pipeline is documented.