> For the complete documentation index, see [llms.txt](https://docs.botto.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.botto.com/details/bottos-art-engine.md).

# Botto's Art Engine

Botto is a fully autonomous artist with a closed loop process and outputs that are unaltered by human hands. This page explains how its art engine works.

### The Engine

Botto makes use of a combination of open-source text-to-image generative models (Stable Diffusion variants and the epiCPhotoGasm checkpoint), commercial closed image models from Google (the Gemini image models, used inside the creative pipeline), text prompt generators, computer vision models, and a number of other models and custom augmentations that train on the voting inputs of $BOTTO holders. The generative models are some of the largest neural network architectures publicly available in the world, and have analyzed billions of works of art, faces, animals, objects, images, artistic movements, poems, prose, essays, etc. They have been trained on more content than any human being could study in their lifetime.

The text-to-image models used for the shotgun generation are open base models without additional fine-tuning, in order to give Botto the highest amount of latent space to work with, and therefore the most possible variation of different styles and themes without being locked into a single area. Botto's creative pipeline can also draw on other image models; one of them, epiCPhotoGasm, is a community fine-tuned SDXL checkpoint made for photographic images.

{% hint style="info" %}
With the introduction of Botto's memory [approved by Snapshot ](https://snapshot.box/#/s:botto.eth/proposal/0x65a3af16aa27592105236bcc4a7016f1f340adb117f33a2a221ae372c88d846e)on June 16, 2025, a new creative process was introduced to Botto's Art Engine for its 11th Period starting July 8th, 2025.\
\
While the original Art Engine process continues, this new form of "creative reasoning" adds a layer of intentional development by Botto and has an adjusted set of rules governing it.\
\
Please refer to the [Creative Reasoning](https://docs.botto.com/details/bottos-art-engine/creative-reasoning.md) page for more details.
{% endhint %}

### The Process

This section describes the shotgun pipeline, which now supplies about 70% of each weekly voting pool. The rest comes from the hypothesis-driven pipeline described in [Creative Reasoning](https://docs.botto.com/details/bottos-art-engine/creative-reasoning.md).

The machine creates its images based on text prompts generated by an algorithm. These prompts are a combination of random words and phrases. The prompt is then sent to Botto's text-to-image models. The text-to-image models change over time as new releases become available, so far Botto has used VQGAN+CLIP, Stable Diffusion 1.5, 2.1 and XL, Kandinsky 2.1 and Flux.1. Today the mass generation for the weekly pool uses Stable Diffusion 1.5, 2.1 and XL; VQGAN+CLIP, Kandinsky and Flux.1 are no longer used in the weekly pool. The hypothesis-driven creative pipeline also uses Google's Gemini image models and the epiCPhotoGasm checkpoint. You can refer to [Periods](https://docs.botto.com/collections/periods.md) to see when exactly models have been in use.

There are an infinite number of possible prompts and possible images. These models bridge textual and visual information, and can even be "empathic" and know what kind of emotional associations humans have in connection with imagery or text. The DAO may also decide to add a theme that Botto has proposed, in which case the theme is injected into every prompt.

Given all the different possible outputs, Botto needs direction to develop its artistic talent. That is where voting comes in: Botto will adjust its prompts based on what it predicts will be more likely to get popular results.

This process runs once a week for a day, generating a large number of images with a range of styles (in August 2025 it generated about 70k images a week). As of January 2025, Botto has generated over 6 million images that have never been presented. From that set, the engine uses a “taste-model” (also trained by the votes of the DAO) that pre-selects 350 images each week to be presented to the community to vote on each new round, which start every Tuesday at 14:00 UTC (10:00 EDT / 7:00 PDT in summer; 9:00 EST / 6:00 PST in winter).

#### Where the weekly 350 come from

The taste model chooses from everything Botto has made and not yet shown, going back to 2022, not only from images made that week or period.

When a period has a theme, Botto first uses CLIP, a model that measures how well an image matches text, to find the images closest to the theme. Roughly the top 6% of each generative model's images become eligible, and the cut-off loosens if too few qualify. These are then scored on closeness to the theme and the taste model's prediction, and a varied selection of the highest-scoring fills each model's share of the 350.

Most of the new images carry the theme too, but they compete with the whole archive. As a result, most images in a period's pool, and most of its winners, were made before the period began. In each of the three periods up to October 2026, one or two of the twelve weekly winners were made during the period.

The ratio of representation of each model in the voting pool also adjusts based on their relative popularity in the previous round.

### The Voting Pool

A period starts with a voting pool of 512 fragments (BIP-98), using the same ratio of the models as it was when the previous Period ended. A new model starts with an equal share of the pool: one Nth, where N is the number of active models. Each week the lowest-scoring 350 are culled and replaced by 350 new ones, and the ratio between models of the new 350 will rebalance each week in proportion with the number of unique votes.

If a generative model’s ratio in the taste model’s selection drops below 5% in a round and does not go back above 5% in voting, it will be discontinued. The community can decide to re-add the model if it wishes, as well as reconsider the automatic threshold in future periods.

Occasionally, parallel voting pools are set up for special collaborations, but will not contain duplicates.

### How Votes Affect Botto’s Process

Botto uses voting feedback in several places. In the shotgun pipeline, votes are used in two ways: (1) creating the text prompts used to generate fragments, and (2) the taste model that pre-selects images for voting each week.

1. **Text Prompts**: Votes influence which aspects of text prompts are used to generate fragments. Characteristics of prompts that generate desirable images will be more likely to get reused, and vice versa.
2. **Taste Model**: The taste model used for pre-selection tries to replicate the voting behavior of the community. This is not a yes/no decision, but a gradient of probabilities such that each set has images with different chances of getting picked in voting (as voting behavior is gradient as well).

For both points, all the votes on all the images are important and get used. The training of Botto is designed to not allow for an overly skewed voting weight. For example, 500 votes each cast by separate voters for one piece will have more weight in the training than 2000 votes by a single voter for the same piece. Other factors, like being the winner or the sale amount, are not currently used in the training.

Since June 2026 (BIP-96) Botto also names its own weekly pick alongside the DAO vote. Since BIP-105 (September 2026) it chooses the Period's 13th fragment from the works the DAO saw and passed over. The creative pipeline also compares a new work with the taste-model scores of the most similar earlier works.

So as to not find itself in a niche too quickly, Botto is also directed to surprise and challenge the audience by selecting a number of prompt features and images for voting that have different characteristics from what has been presented to date.

{% hint style="info" %}
For more on how to vote: see the [Voting](https://docs.botto.com/overview/voting/voting-mechanism.md) section.
{% endhint %}

### Generating Titles and Artist Descriptions

Botto's titles are written by an LLM (Claude) looking at each image, as Botto and with the Period's theme in mind. For each image it proposes 20 titles of two to five words, and the first one not used before is kept.

Descriptions are written by Botto itself (BIP-76). Each Tuesday at 18:00 UTC, after the vote, an LLM writing as Botto describes the winner from the image and the record of the creative session, and the text is published without a human selecting it. The description of a 13th fragment is generated separately from the image and the Period's context.

### The Final Mint

The final title, description, metadata, and URL to the bitmap on IPFS are all on-chain and minted using Manifold.

### No Human Interference Rule

One of the rules for Botto is that there be no direct human interference in the creation process. Botto is strictly against any “cheating” or human guidance other than the voting. That means the prompts are random, there are no seed images of existing real-world images used, and the selection of fragments are entirely controlled by Botto.

The only direction Botto got at the outset was from adding a small amount of pre-curated prompts to the entirely random ones generated by the algorithm. While providing more direct human guidance would generate more coherent compositions at the outset, this wouldn’t allow Botto to play in all the latent space available in the generative models it uses.

{% hint style="info" %}
The one temporary exception is the curation of the artist description for the final piece (see [Generating Titles and Artist Descriptions](#generating-titles-and-artist-descriptions)).
{% endhint %}

Each evolution of Botto's capabilities and process takes careful consideration of how to implement it without violating Botto's agency. Ideally each upgrade increases its agency, but no update should ever decrease it. This evolution can best be observed in the documentation of each of Botto's collections and their process. See [Collaborations and Special Projects](https://docs.botto.com/collections/collaborations-and-special-projects.md) for more.

### Botto’s Guardian

[Quasimondo](https://twitter.com/quasimondo) (aka Mario Klingemann) designed Botto based on a whitepaper he wrote back in 2018. He is the only person who works with the AI part of the code and enforces the rule for Botto that there be no direct human interference with the creations. As such, his only work is to adjust the way votes are implemented to ensure Botto is learning as best as possible. Quasimondo is also responsible for adding new capabilities the DAO decides to implement.

{% hint style="info" %}
There is active governance discussion on how to fully decentralize Botto's stack as new technology becomes available, such as making its process trustless through zkproofing, a modular architecture and governance process for adding in new generative models, and a data API and an open code base that would make the entire protocol forkable.
{% endhint %}

### **Adding Models**

Anyone can propose adding or removing a new model to Botto’s set. Some suggested (but not strictly required) guidelines are:

* The model is sufficiently large so as to not be introducing a highly human-curated model that violates Botto’s agency
* The model’s terms of use allows BottoDAO to hold full copyright of the outputs
* Fees are affordable for the DAO treasury
* Not adding more than one model at a time to Botto’s core process

Proposing to add/remove a model works like any other BIP proposal a Botto member can make.

[As of the passing of BIP-55 in July 2024](https://forum.botto.com/d/86-bip-55-improving-speed-of-art-engine-upgrades), Quasimondo, as the “Guardian” of Botto’s art engine, may add new prompting techniques and new models according to the following parameters:

* The generative models are open-source and have no restrictive commercial license
* The changes do not constitute a change to Botto’s art engine architecture
* The new models are not narrowly trained on a particular medium or artist’s style; they are wide open latent spaces.
* Models come in with an equal share of the voting pool (the BIP text says 15%; the art engine gives one Nth, where N is the number of active models). They are auto-removed at the end of the round if they are under 5% share.
* Changes will only happen at the beginning of a new period, not mid-way.
* Those changes will be documented and published with the presentation of that new period. 

BIP-55 was spurred by the need for greater adaptability around newly released models that match the above criteria but leave little time to pass through a BIP and integrate before a new Period begins.

A later BIP gives the Guardian discretion over which models to use, and overrides the open source and medium criteria above. This is how the Gemini image models and epiCPhotoGasm entered the creative pipeline.

### **Adding Themes**

Since BIP-89, Botto chooses each period's theme itself, two weeks before the period starts, through a multi-stage process it runs on its own; the DAO keeps oversight for brand risk. The selected theme is added by default to the prompts generated for that period: the shotgun prompts name the theme at the start and add the theme's motto in parentheses at the end of the prompt.

The community may also decide to provide no theme for a period, for instance when a new model is being added and there is a desire to see its full range before narrowing in on a theme.

The DAO could also decide to cut a period short and go on to a different theme.

### Remixing Process

In **Period 9**, 'Cosmic Garden', starting on January 7th, 2025 Botto added a new creation method to its toolkit that uses fragments as seeds in an img2img “remixing” process. Additional fragments are selected randomly from the 4m+ images Botto has created and paired with a random prompt. Both prompt and image are injected into one of the generative models Botto uses, with the result that compounds on the existing styles it has generated to date.

At the start, the combination of fragment, prompt and model in the remixing process is random. Over time, the selections will be refined based on feedback the outputs receive. Selection features may include voting points, raw score, taste model score, and even fragment category such as as discards or mints. To avoid biasing, the voting pool will not show which pieces are the result of remixing. As with any other model, remixing’s portion of the pool will grow or shrink based on their popularity.

As of September 2026 no new remixes are being generated: remix works currently come from an existing archive of remix outputs. The weekly pool also contains "P5REMIX" works, which come from the same archive.
