Rogues Ink — Training Data Disclosure

This page describes the data used to train the Rogues Ink AI image generation model. It is published in compliance with California's Assembly Bill 2013, the Generative Artificial Intelligence: Training Data Transparency Act (effective January 1, 2026), which requires developers of generative AI systems made available to people in California to publish a high-level summary of the data used to train those systems.

The disclosure below covers only the Rogues Ink fine-tuning layer — the custom model trained by Rogues NFT Inc. on top of the FLUX.1 [dev] base model. The base FLUX.1 [dev] model is developed by Black Forest Labs, Inc., and any disclosure obligations relating to the base model's training data are the responsibility of Black Forest Labs, not Rogues NFT Inc. If you are interested in the base model's training data, please refer to Black Forest Labs' own publications.

What is Rogues Ink

Rogues Ink is an AI image generation tool that lets verified holders of Rogues NFT collection assets generate images in the visual style of the Rogues NFT character art. Under the hood, Rogues Ink uses a fine-tuned version of FLUX.1 [dev], with a custom LoRA (Low-Rank Adaptation) layer that teaches the base model the visual conventions of the Rogues NFT collection — the character proportions, the linework, the palette, the way light is handled, and the general aesthetic of the artwork.

The LoRA layer is what makes Rogues Ink generate "Rogues-looking" images rather than generic FLUX.1 [dev] outputs. The disclosure on this page describes the data used to train that LoRA layer.

How the LoRA was trained

The Rogues Ink LoRA was trained on Modal's A100-80GB GPU infrastructure using the Ostris ai-toolkit framework, pinned to a specific software version. Training ran for 1,500 steps with checkpoints saved every 250 steps; the production checkpoint selected from the run is approximately 344 megabytes in size and is hosted as a private model on Replicate, where it is loaded on top of FLUX.1 [dev] at the time of each image generation request.

Training was performed by Rogues NFT Inc. directly, using source artwork that Rogues NFT Inc. owns. No third-party training infrastructure beyond Modal was used. The base FLUX.1 [dev] model was not modified or retrained; only the LoRA adapter layer was trained.

The training data, in twelve parts

The sections below address each of the twelve categories of disclosure that AB 2013 requires.

1. Sources and owners of the training data

The training dataset consists entirely of character artwork from the Rogues NFT collection on the Solana blockchain. All source artwork was created for, and is owned by, Rogues NFT Inc. No images from third-party sources, public datasets, web scrapes, or licensed external collections were used in training the LoRA layer.

2. How the training data furthers the system's purpose

The purpose of the Rogues Ink LoRA is to adapt the general-purpose FLUX.1 [dev] image-generation model to reliably produce images in the specific visual style of the Rogues NFT collection. Training the LoRA on the actual Rogues character artwork is the most direct way to teach the model that style. Without Rogues-specific training data, the base FLUX.1 [dev] model would produce technically competent images that do not look like Rogues characters; with the LoRA, the model can faithfully reproduce the collection's aesthetic when prompted.

3. Number of data points in the dataset

The training dataset consisted of 30 images.

This is a small dataset by general-AI-model standards. LoRA fine-tuning is specifically designed to adapt a large base model to a narrow visual domain with a small, focused training set, rather than requiring the millions of images used to train a base model from scratch.

4. Types of data points

The training data consists of digital images paired with text captions describing the content of each image. The images are bitmap files (PNG or similar). The captions describe what is depicted in each image — the character, the pose, the clothing, the setting — and were used during training to teach the model to associate prompts with visual outputs.

Captions were operator-curated with AI assistance. For each image, the operator worked iteratively with AI assist (Gemini and Claude) to draft a descriptive caption, then reviewed and edited the draft to ensure it accurately described the image, used the correct trigger token (X-RGUE-UNIT entity) for the character identity, and avoided language that could bias the model toward unwanted associations. The current dataset and caption set is the seventh iteration; six earlier iterations were discarded after evaluation. Final captions follow a consistent structural pattern: a style identifier, the trigger token, anatomical features of the character, the character's clothing, the character's pose or action, and the background or setting.

5. Whether the data is protected by copyright, trademark, or patent

The training data consists of original artwork created for the Rogues NFT collection, owned by Rogues NFT Inc. The artwork is protected by copyright as creative works under standard intellectual property law; that copyright is held by Rogues NFT Inc. The training set does not include any data in the public domain, nor any data protected by trademarks or patents held by third parties.

6. Whether the data was purchased or licensed

Neither. The training data was created internally by or for Rogues NFT Inc. as part of the Rogues NFT collection. Rogues NFT Inc. did not purchase the training data from any third party and did not enter into any data-licensing agreement to obtain it.

7. Whether the data includes personal information

No. The training data consists of character artwork — stylized illustrated characters that are not photographs of, and are not modeled on, identifiable real people. The training set does not include personal information of any individual.

8. Whether the data includes aggregate consumer information

No. The training data does not include consumer behavior data, demographic data, transaction data, or any other category of aggregate consumer information.

9. Whether the data was cleaned, processed, or modified

The training data underwent standard pre-training preparation including:

• Resizing of source images to a uniform resolution appropriate for LoRA training on FLUX.1 [dev]
• Pairing of each image with a text caption describing its content (see §4)
• Quality filtering to remove low-resolution or otherwise unsuitable images from the training set
• Standard image-format normalization

The purpose of this preprocessing was to ensure the training set met the technical requirements of the Ostris ai-toolkit framework and the FLUX.1 [dev] base model's expected input format. No content-altering modifications were made beyond what was necessary for training compatibility.

10. Time period during which the data was collected

The source artwork used to train the Rogues Ink LoRA was created during the year 2025.

11. First use of the data in development

LoRA training was first run on this dataset on April 18, 2026. The production checkpoint currently served by Rogues Ink was selected from that training run.

12. Use of synthetic data

No synthetic data was used in training the Rogues Ink LoRA. The training set consists entirely of original artwork created by or for Rogues NFT Inc.; no AI-generated images were included as training inputs.

Updates to this disclosure

If Rogues NFT Inc. trains a new version of the Rogues Ink LoRA, retrains the existing LoRA on new data, or otherwise materially modifies the model in a way that affects the training data described above, this page will be updated to reflect the new state. The "Last updated" date at the top of this page will indicate when the most recent update occurred.

This disclosure does not cover any future model versions until the page is updated.

Questions

Questions about this disclosure can be directed to:

Rogues NFT Inc.
Email: contact@roguesnft.com