Picbreeder Rebuilt With a VLM: Can AI Be Creative?

Picbreeder, the goal-free evolutionary art website

Picbreeder was an evolutionary art website where visitors bred pictures by choosing the ones they liked, with no target to aim for. It opened in 2008 and stopped working around 2016.

A visitor saw a grid of abstract patterns and picked one or more of them. The site then produced a new grid of variations, and anyone could pick up a published image and keep breeding it.

Earlier tools for this kind of art were mostly built for one user at a time, and that user tired out before a run got far. Picbreeder spread the effort across many visitors, so an image could keep changing long after the first person left. Faces, animals, vehicles, and skulls turned up this way, though nobody had set out to draw them.

The surviving records made the new study possible. The paper works from a dataset of 9,758 images published on the Picbreeder website, each with its complete line of ancestors, and rebuilds the whole family tree from it. That human archive is the benchmark every AI run is measured against.

CPPN and the NEAT algorithm: images you can pick but not plan

A CPPN, short for Compositional Pattern-Producing Network, is the small neural network that holds one Picbreeder image. It takes a pixel's position as three values, x and y plus r (the distance from the center), and returns that pixel's color as hue, saturation, and brightness. Because any position can go in, the same network can draw its image at any resolution.

Visitors never wired these networks by hand. When someone picked images, the site bred them with the NEAT algorithm (NeuroEvolution of Augmenting Topologies), a neuroevolution method that mutates networks and crosses them with each other.

Gradient descent adjusts only the weights of a network whose layout is fixed. NEAT changes the weights too, but it can also change the layout itself, meaning the nodes and the links between them.

A Picbreeder image could be picked but not planned. The technical blog says that for the person in the loop, imagining an image in advance and then breeding it is close to impossible. Images came instead from a long run of choices, one grid at a time, and all a visitor could do was pick what looked interesting and see what grew.

Kenneth Stanley and the myth of the objective

Kenneth Stanley created Picbreeder with his collaborators, and Jimmy Secretan led the project. Sakana AI's technical blog calls Stanley's work on CPPNs, NEAT, and open-ended search the base the new experiment builds on. The same page's reading list also points to novelty search, the idea of searching for novelty instead of a fixed objective.

Stanley and Joel Lehman wrote Why Greatness Cannot Be Planned: The Myth of the Objective (Springer, 2015). The technical blog points to the claim Stanley has argued across his career, in that book among other places: the objectives most worth reaching are the ones people can only recognize in retrospect. Picbreeder fits that claim, since a skull or a car appeared first and only then did someone see it as worth keeping.

The paper gives the property a name: open-endedness, the capacity to generate a seemingly endless supply of novel and meaningful new forms. The authors see it in how people produce science, technology, and art. They treat Picbreeder as the standard example of people searching this way with a computer in the loop.

Stanley also wrote closing words for the technical blog. He hopes Sakana AI's decision to revisit Picbreeder with modern AI leads others to take seriously how humans and AIs differ when they explore with no set goal.

Sakana AI replaced Picbreeder users with VLM agents

Sakana AI's blog post of July 10, 2026 describes a GECCO 2026 paper, written with MIT and NYU, in which VLM agents take over the users' side of Picbreeder. Nobody gives the agents a target image or says what progress looks like. They browse, breed, publish, and rate, in a pared-down copy of what people did on the old site.

Sakana AI calls the project the AI Picbreeder Experiment. One session runs in four steps, while a separate critic keeps rating the archive in the background.

An agent first sees 100 images from the shared archive. They sit on five shelves of 20: top-rated, best new, most branched, latest, and random. The original home page had similar shelves, with 8 images each.

The agent branches from one image or starts from a fresh random batch. It then picks parents for 20 generations, with 15 candidates in each, and finally publishes one image under a title of its choosing.

Ten breeding agents run at once and share one archive. Once the archive holds at least 100 images, a new VLM with no memory of earlier sessions rates a sample of them from 1 to 5 each time 5 more are published. The scores then fill the top-rated and best-new shelves that later agents browse.

Interactive evolution with vision language models as users

Interactive evolution is evolution where a person, not a formula, decides which candidates survive each round. In this study the person is replaced by a vision language model, an AI that reads images and text together and answers in text. At every decision, the agent writes one or two sentences explaining its choice.

The agents got the same controls human visitors had. They could switch color on or off and set the mutation strength anywhere from 0.01 to 1, with 0.5 as the default.

Session length had to be fixed, because agents that could decide when to publish did so almost at once and ended the session. The team therefore set every session at 20 generations, just under the human average of 21.96 generations. Human sessions varied widely: the median was 11, the longest ran 765, and 70.9% lasted 20 or fewer.

Anyone handing open-ended tasks to AI agents should expect the same habit. In this study, how long the agents worked had to be set from outside.

The default model was gemini-2.5-pro. The newer gemini-3-pro did worse and fixated on perfecting its favorites even more, while the lighter 2.5 Flash and Flash-Lite models drifted toward busy noise. Each setting ran 2,000 sessions, one published image per session, and every setting was repeated with 6 random seeds.

The GECCO 2026 paper and the archive it measures against

The Picbreeder paper sets out to find the "ingredients" of open-endedness by replicating the old site with large vision language models. Its six authors are Sam Earle and Julian Togelius of New York University, Kai Arulkumaran and Sebastian Risi of Sakana AI, Akarsh Kumar of MIT, and Andrew Dai, who is listed as independent. Earle, Dai, and Kumar did their work during an internship or residency at Sakana AI.

The paper first appeared on arXiv on April 1, 2026, and a revised version followed on May 27, well before Sakana AI's July 10 blog post. It is part of the proceedings of GECCO 2026, the Genetic and Evolutionary Computation Conference, held July 13-17 in San José, Costa Rica.

Sakana AI's post says the work is nominated for a best paper award. The project page now lists "Best Paper (Complex Systems), GECCO 2026" in its header.

Four measures turn the gap between people and agents into numbers. Every AI run is scored on them against the human archive:

• Semantic recall: how well the archive's images match 1,824 common object names from the THINGS dataset.
• Visual coverage: how many different looks the archive spans.
• Semantic coverage: how many different subjects it spans, judged from one-sentence captions that gemini-2.5-pro writes for each image.
• Tree balance: whether new images branch from many parents or pile onto a few favorites.

The technical blog treats the human archive as the upper bound for the AI runs and a baseline that makes every choice at random as the lower bound. The random baseline scores low on every measure except tree balance, where blind branching gives the most even tree by design.

Open-endedness results: VLMs circle back, diversity helps

Measured against the human archive, a single VLM agent kept coming back to the same images and made only small jumps between ideas. Giving agents many different personalities helped the most, and one turn of memory beat both none and ten. A skull the agents evolved also held together better under changes than one built with gradient descent.

The table lists the four scores for the main settings, from Table 1 of the paper. Each AI figure is a mean over 6 seeds.

The human archive keeps the best recall and visual coverage. The 1,000-personality setting edges past it on semantic coverage, though 10-20% of that archive is unreadable noise. It also grows a more balanced tree, and random play has the most balanced tree of all because it branches at random.

Agents keep returning to the same images, so leaps stay small

Under the default settings, the archives filled with dozens of near-identical versions of the same few forms. Shapes that looked like foxes and fishbones kept coming back. Semantic coverage came out barely above the random baseline.

Tree balance tells the same story from another side. The default agents scored 0.235, against 0.363 for the human archive, because they kept branching from the same images and rarely started over from a random batch.

Two car lineages show the difference in pictures: the human car came from a face, which led to an eye, then a frog, then the car. The agents' car came from an abstract bird, then a car seat, a dashboard, and a hood ornament. People jumped between unrelated subjects, while the agents stayed close to cars once the car seat appeared.

The team also forced random picks into the loop at a set rate. More randomness made the archives more varied in look and subject, but the images got less sharp, and above a rate of 0.25 recall fell as well.

Even at a rate of 0.75, tree balance only rose from 0.235 to 0.303 in the paper's results. Among noisy options the agent still always had a clear favorite, and the authors call this stubborn favoritism a possible barrier to using VLMs for open-ended search.

Diverse personalities reach the human archive's semantic coverage

With 1,000 personalities in the pool, the agents matched the human archive on semantic coverage. A personality here is a short trait written by a language model and placed at the top of an agent's instructions. The team generated these in batches of 50, up to 1,000 traits.

At the start of each session, one trait is drawn at random. One read "searching for the dry red color of terracotta clay," and the agent that got it published a large set of near-identical solid swatches.

The 1,000-personality runs scored 0.734 on semantic coverage, against 0.730 for the human archive. Their visual coverage rose to 0.665 but stayed just below the human score of 0.681, while recall barely moved.

With only 10 personalities, visual coverage actually dipped, to 0.605 from 0.614 for a single agent. The paper suggests this is because each agent settled into its own corner of the archive.

There is a catch. In the many-agent archives, 10-20% of the images are busy, mostly grayscale patterns that do not look like anything, and they hardly appear in other settings. Of the kinds of noise in the study, this one gave archives the highest semantic coverage by a wide margin, well above the random baseline or heavy forced randomness.

The authors compare them to adversarial attacks, images that fool a model into reading meaning that is not there. That matters when reading the 0.734 score, because part of the archive behind it is noise rather than new subjects.

A little memory helps; too much memory hurts

One turn of memory worked best, and longer memories did not beat it. Memory in this study is the number of earlier steps from the current session that stay in the agent's context. With none, the agent saw only the grid in front of it and kept publishing duplicates, so visual coverage fell to 0.527 from 0.614 at the default.

One earlier turn stopped most of the repeats and gave the best recall of any memory length tested, 0.087 against 0.083 for two turns. At ten turns, recall fell to 0.079, and archives across several seeds filled with near-photo-real top-down views of soda cans that almost never showed up at other lengths.

The authors suspect the agent falls into loops that reinforce its own picks. The technical blog puts it more simply: too much memory seems to overload the agent's context.

One result carries over to teams that run AI agents on long tasks. In this setup, one step of history kept the agent from repeating itself, and a long history seemed to push it back toward its own habits.

The AI-made skull CPPN holds up better than a gradient-descent one

The skull the agents evolved stayed recognizable when its weights were changed, while a skull trained with gradient descent fell apart. Earlier work by Kumar and colleagues in 2025 argued that training toward a fixed target with stochastic gradient descent (SGD) leaves a network "fractured, entangled." Open-ended search in Picbreeder, they argued, gives more unified internals.

The team nudged each weight of the skull's CPPN by values from -1 to 1 and watched the picture. The agent-made skull changed fairly smoothly and mostly still looked like a skull, while the same test on a skull CPPN trained with SGD gave chaotic, destructive changes.

The human-evolved skull stays ahead, though. Its single weights act as clean controls, such as opening the mouth or winking an eye, and the agent-made skull has nothing that tidy.

The authors are careful about what this shows. Their comparison pits NEAT-style evolution against SGD on a network of fixed shape, and the two skulls started out different, so either factor could explain part of the result. The technical blog adds that it is not yet clear how much of the sturdiness comes from CPPNs and NEAT and how much from the VLMs' judgment and the interaction loop.

Can AI be creative? The gap that human creativity still holds

Not yet, in the full sense. The agents did notice interesting patterns but tended to get stuck in them, while people were better at spotting a lucky accident, refining it, and then making a bigger leap. The researchers say they do not yet know what lets humans do this, or which ingredient today's AI lacks.

For the broader AI creativity question, this is one narrow test, built on one website's rules and one main model. Within it, the difference showed up less in what the agents could notice than in what they did next.

Serendipity: people pursue it, and the AI gets trapped by it

Serendipity means finding something good by chance, and Picbreeder ran on it. The technical blog names the tension in asking agents to join in. Modern agents are trained and tested as goal-followers, yet this experiment asks them to wander with none.

The team's own look at the archives lines up with the numbers. The agents often got pulled into their own attractors, forms they kept drifting back to.

People seemed to seize a lucky break and give it attention until it became something refined, while the AI seemed to take luck in more passively. It matches the stubborn favoritism that held up under forced randomness.

Not every attractor was dull. They differed enough between seeds and settings that the team found running the system again and again fairly entertaining.

The missing ingredient in open-ended learning is still unknown

Sakana AI's team says it does not yet fully understand what lets people search this way, or which ingredient today's AI systems lack. For now, the post says, the results suggest that human creativity holds something important that AI agents have not yet learned to reproduce.

The paper names three next steps: letting agents evolve with no end point, trimming what goes into their context, and fine-tuning VLMs on recorded Picbreeder sessions. The last idea would give the agents something like long-term memory.

The study has clear limits. Each setting ran with 6 seeds, and most tests used one default model, gemini-2.5-pro. The authors say more varied conditions and larger scale are needed to judge how open-ended VLMs really are, and that user studies with people as judges may help calibrate the measures.

Running the study is also expensive, according to an appendix of the paper. One full run with complete in-session memory, sized to match the human dataset at about 10,000 images, costs around 10,000 USD. By the time the paper was written, the team had spent about 150,000 USD on Gemini API calls.

The results are open to browse. The technical blog on pub.sakana.ai links the paper, the code at github.com/smearle/picbreeder-vlm, and the data, and it hosts a homage to the original site filled with images from replays of the AI runs. The same company has also tested agents left to work on their own in a very different job, running a shop, described in CoffeeBench by Sakana AI: How It Works and Results.

The paper and the technical blog are long English pages. The tool below turns a web page into Markdown, so you can keep either one as plain text with its headings in place.