Picbreeder-VLM Archive

Every image evolved by the swarm of vision-language-model "breeders" in In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models (GECCO 2026), together with the CPPN genomes that produced them, the agents' reasoning transcripts, the lineage graphs, and the analysis artifacts behind the paper and blog.

The original Picbreeder (Secretan et al., 2008) let crowds of humans collaboratively evolve images from CPPN genomes, discovering an open-ended tree of recognizable pictures. This dataset is what happened when we replaced the humans with VLM agents: agents join a shared archive, look at candidate images, pick and mutate them toward whatever they "see," and publish their discoveries for others to branch from.

• 📝 Blog / interactive report: https://pub.sakana.ai/picbreeder-vlm
• 📄 Paper (arXiv): https://arxiv.org/abs/2605.23908
• 🧬 Breed your own: https://pub.sakana.ai/picbreeder-vlm/breed/

How the images were made

Every image here was produced by the same loop, run thousands of times. Its unit is a session.

A session begins when a VLM agent either branches an image already published to the shared archive, or starts from a fresh population of randomly initialized CPPNs. The agent is shown a population of 15 CPPN-images and selects one or several as parents. Random mutation — and crossover, when it selects more than one — produces the next population, and the agent selects again; exact copies of the parents are always carried among the offspring. At the 20th generation the agent must choose one image to publish to the archive and give it a title. That published image, with its genome and its lineage, is one row of this dataset.

At any step the agent may toggle color mode, set mutation strength anywhere in [0.01, 1] (default 0.5), and, in color mode, restrict mutation to color-only or structure-only. Random initial populations start grayscale; branched images inherit the color mode they were published under. These controls mirror the original Picbreeder interface.

Agents act concurrently against one shared archive, so an image published by one agent becomes available for another to branch. That is what makes a run a single collaborative phylogeny rather than a bundle of independent searches. Each run is 2,000 sessions.

The experimental conditions perturb exactly three things about this loop: how often an agent's selection is overridden by a random one (ε), how much of its own history the agent remembers (memory_cl), and how many distinct personalities the agent pool draws from (n_personalities).

The paper's sweep

The results in the paper come from 185,013 images across 81 runs and 14 conditions, all bred by gemini-2.5-pro, six seeds per condition. These are exactly the rows where

canonical == True and model == "gemini-2.5-pro"

Everything else here is supplementary — see Beyond the paper below.

from datasets import load_dataset

ds = load_dataset("picbreeder-vlm/picbreeder-vlm-archive", "images", split="default")
paper = ds.filter(lambda r: r["canonical"] and r["model"] == "gemini-2.5-pro")

paper[0]["image"]      # PIL.Image (128x128)
paper[0]["caption"]    # "An abstract black and white oval shape with a motion blur effect."
paper[0]["memory_cl"]  # 1

Browse the images directly in the Dataset Viewer above — sort by vlm_rating_mean, filter on canonical, or full-text search caption.

Splits

Each split is an experimental condition from the sweep.

ε is the probability that an agent's selection is short-circuited and replaced by a random pick. memory_cl = -1 is a sentinel meaning unbounded, which the paper labels mem_20.

Where the annotations come from

Neither the ratings nor the captions were visible to the agents while they evolved. Both are post-hoc annotations, added after the runs finished, to make the archives measurable. Treat them as model outputs under study, not as ground truth and not as the selection signal.

Ratings (vlm_rating_mean, vlm_rating_count) come from a separate pass in which a VLM is shown batches of published images and asked to score each 1–5 with a one-sentence justification. An image may be scored in more than one pass, so vlm_rating_count is the number of scores its mean is taken over — it varies per image and is not a fixed panel size. Images published earlier tend to accumulate more scores, so vlm_rating_count correlates with publication time; keep that in mind before reading vlm_rating_mean as a like-for-like quality ranking.

Captions come from a gemini-2.5-pro pass over each archive. Their purpose is evaluation: the captions are embedded in a text-embedding space, and the spread of that cloud is the paper's semantic coverage metric. In practice a caption is a single descriptive sentence — "A stylized, black-and-white drawing of a bird's head against a heart-shaped background." — with a median of 11 words.

Some runs were captioned only up to a per-run cap, applied in publication order. Where a run is partly captioned, the captioned rows are a chronological prefix of it, not a random sample — so a caption-based analysis of such a run is an analysis of its earlier sessions. Filter caption is not null and check coverage per run before pooling.

The paper's Semantic Recall metric uses no captions at all: it embeds the images themselves with SigLIP2 and measures cosine distance to the 1,824 deduplicated THINGS class names.

Columns

Columns contain image metadata, including the image's provenance from within the experimental run that generated it and the experiment's hyperparameters.

A caption is a fact about an (image, captioner) pair, so each captioner gets its own column. The column name records the captioner, never the breeder — a gemini-2.5-pro row can carry a qwen3-vl-8b caption.

Image provenance

The image column carries the 128×128 thumbnail — the exact pixels the blog gallery renders, re-encoded to WebP from each run's sprite atlas. These are thumbnails, not the original renders.

Because every image is produced by a CPPN, it is resolution-independent: you can re-render any row at arbitrary resolution from the genome archives (site/<run>/genomes.json.gz, or results/<run>/genomes.tar.gz), joining on genome_key.

---

Beyond the paper

Alongside the sweep, this repo ships 78,589 images from 98 supplementary runs: alternative VLM backends, and non-canonical variants of the sweep conditions. They share the same splits, so filter on canonical and model == "gemini-2.5-pro" when you want the paper's rows and nothing else.

The model_ablation split holds the runs that vary the breeder model, and is therefore entirely supplementary. No supplementary run was captioned — the captioning pass only ever ran over the paper's sweep — so both caption columns are null throughout them, and about half their images (48%) carry a rating. Filter caption is not null before any caption-based analysis.

What's not in the parquet

Per-rater rating vectors and per-rater comments (tens of values per image) are omitted to keep the parquet lean and browsable; vlm_rating_mean / vlm_rating_count summarize them. The full vectors live in results/<run>/archive_metadata.json.

Repository layout

The parquet under data/ is a browsable face of the corpus. The rest of the repo is the raw material, kept in the layout the interactive blog fetches:

Limitations and intended use

• The images are machine-generated abstract art, not natural images, and the captions and ratings are machine-generated too. Treat caption and vlm_rating_mean as model outputs under study, not ground truth — a large part of the paper is about how unreliable they are.
• The random split is a control in which selection carries no signal by construction. Don't pool it with the others without meaning to.
• Selection in the noise ablations is randomized by design — that is the ablation.
• This corpus is intended for research on open-endedness, novelty search, and VLM evaluation behavior. It is not a benchmark, and it has no held-out test set.

Citation

@inproceedings{earle2026picbreedervlm,
  title     = {In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models},
  author    = {Earle, Sam and Arulkumaran, Kai and Dai, Andrew and Kumar, Akarsh and Togelius, Julian and Risi, Sebastian},
  booktitle = {Proceedings of the Genetic and Evolutionary Computation Conference (GECCO)},
  year      = {2026}
}

Please also credit the original Picbreeder: Secretan et al., Picbreeder: Evolving Pictures Collaboratively Online, CHI 2008.

License

CC BY-NC 4.0 — attribution required, non-commercial use only.