AI Disclosure

This AI Disclosure explains how Veyraflow uses artificial intelligence (AI): which models create your assets, what happens to your request before it reaches a model, what AI can and cannot do, how we label AI-generated files and what we expect from you when you use them. Veyraflow is operated by Jan Nastały (see the Impressum). This document complements our Terms and Conditions and our Privacy Policy. Capitalized terms used here, such as "Services" and "Generated Assets", have the meaning given to them in the Terms and Conditions.

KEY POINTS

• Everything Veyraflow generates is made by AI. Every 3D model, sound effect, music track and voice line that you create with the Services is generated by AI models. No person designs, edits or reviews it before you receive it.
• We use models from other companies. Single assets are generated by models developed by Hyper3D, ElevenLabs and Google, which run on fal.ai. Packs uses open-source models developed by Alibaba's Qwen team, Black Forest Labs, Microsoft and others, which we run ourselves on machines rented from RunPod. Before that, models from OpenAI check your request against our content rules and restate it for the chosen model.
• AI makes mistakes. Results vary and may not match your description. Check every asset before you use it.
• Generated files are labeled. Each generated file carries a machine-readable "AI-generated" marker. Sound effects, voice lines and the color textures and previews of 3D models also carry an invisible watermark in the content itself, and music carries Google's SynthID watermark.
• We do not train AI on your content. We do not use your prompts, uploads or Generated Assets to train AI models ourselves.
• Be open about AI when it matters. If you publish an asset that looks or sounds like a real person, place or event and could be mistaken for the real thing, you must make clear that it was generated by AI.

1. WHERE WE USE AI

We use AI for three tasks:

• Generating assets: AI models create the 3D models, textures, sound effects, music and voice lines that you request (Section 2).
• Checking requests: an AI-based filter checks the text of every request against our content rules before anything is generated (Section 3).
• Preparing requests: an AI model restates your request in the form that the chosen generation model understands best; in Packs, a language model also writes the prompt for each item and, if you describe the pack instead of listing its items, plans the list (Section 3).

We do not use AI to chat with you or to answer your messages: emails sent to our contact address are read and answered by a person. We do not use AI to recognize emotions or to categorize people on the basis of biometric data.

2. THE MODELS WE USE

As of 7 November 2026, single generations use the following models, all of which we access through fal.ai:

• 3D models and characters: Rodin Gen-2.5 by Hyper3D (from a description, from up to three photos, or both);
• Sound effects: Sound Effects v2 by ElevenLabs;
• Music: Lyria 3 Pro by Google;
• Voice lines: Text to Speech Turbo v2.5 by ElevenLabs, with one of 21 preset synthetic voices. We do not offer voice cloning: you cannot upload a recording to copy the voice of a real person.

Packs uses open-source models that we run ourselves on machines rented from RunPod:

• Planning and prompts: Qwen3.8-27B by Alibaba's Qwen team;
• Reference images: FLUX.2 [klein] 4B by Black Forest Labs;
• Background removal: BiRefNet, an open-source model by Peng Zheng and co-authors;
• 3D models: TRELLIS.2-4B by Microsoft, which uses the DINOv3 image encoder by Meta.

After that, our own software and Blender clean up the mesh and export it for your engine; this step does not use AI.

Requests are checked, and requests for single generations are prepared, by language models from OpenAI.

The identifier of the model that generated a file is recorded in the file's AI marker (Section 5). We may replace models and providers over time (see Section 15 of the Terms and Conditions). When we do, we update this list and the "Last updated" date.

3. WHAT HAPPENS TO YOUR REQUEST

When you start a generation, your request goes through the following steps:

• Content check. The text of your request, such as the prompt, the text and instructions for a voice line or the items of a pack, and the photos you attach, are checked automatically by an AI-based filter against the content rules in Section 8 of the Terms and Conditions. A blocked request is not sent to any generation model, and no credits are charged. The consequences of repeated blocks are described in Section 12 of the Terms and Conditions.
• Preparation. An AI model restates your request in the form that the chosen model understands best: it translates the request into English if needed, adjusts its wording and order, and combines it with your project's style guide. It is instructed not to add anything that you did not ask for, and if it cannot carry over part of your request, the generation stops with an error instead of guessing. The same step also chooses technical settings that are not available in the form, for example which preset voice best matches your instructions or whether a character is generated in a T-pose. Settings that you choose in the form always take precedence. For voice lines, the text to be spoken is never changed; only the delivery instructions are prepared. The prepared version of your request is not shown to you, but it is stored with the generation, and our administrators may view it to fix problems.
• Generation. fal.ai runs the selected model. Depending on the model, fal.ai may pass your request to the developer of that model.
• Processing. Our software converts the result into the formats and engine presets that you can download (for example, it trims silence, normalizes loudness and packs files) and adds the AI marker and, for the assets listed in Section 5, the invisible watermark.

A pack goes through the same content check, but after that it does not use fal.ai: on a machine rented from RunPod, a language model writes a prompt for each item (and plans the list of items if you described the pack), an image model draws a reference image, which is cut out from its background, a 3D model is made from it, and our software cleans up the mesh, exports it for your engine and adds the AI marker and the watermark. Every item goes through these steps once; nothing is retried automatically, but you can regenerate any item.

We do not review or filter the generated results ourselves. The model providers may apply their own safety filters and refuse a request. In that case, the reserved credits are returned to you.

4. WHAT AI CAN AND CANNOT DO

AI models produce results automatically, based on patterns learned from large amounts of data. Please keep the following in mind:

• Results vary. The same request can give a different result each time, and a result may not match your description, for example in its details, proportions, colors, pronunciation or timing.
• Results can contain errors. 3D models may have holes, uneven geometry or texture seams, and models created from photos may differ from the original object. Models made by Packs can have thin parts with small holes or loose edges, and the number of models a pack makes per hour is only an estimate. Audio may contain unwanted sounds or artifacts, and the length of music is only approximate.
• Results are not guaranteed to be unique. A Generated Asset may be similar to material generated for other users or to existing works, and it may reflect stereotypes or biases found in the data that the model was trained on.
• Legal protection may be limited. In some countries, AI-generated material is protected by copyright only partly or not at all (see Section 3 of the Terms and Conditions).

Check every Generated Asset before you use it in a product, and fix or replace it if needed. You decide whether it is suitable for your purpose (see Section 19 of the Terms and Conditions).

5. HOW WE LABEL AI-GENERATED CONTENT

Visible labels. The note "Assets are generated with AI" appears in the footer of every page of the Services, and pages that show assets shared through a link carry the label "Generated with AI". The README files included in engine export packages state that the files were generated with AI, and the MANIFEST.json file in multi-asset downloads contains the field "ai_generated": true. Pack downloads also contain a file named AI_GENERATED.txt.

Machine-readable marker. When a file is generated, we add a marker to it so that software can recognize it as AI-generated. The marker uses the IPTC digital source type "trainedAlgorithmicMedia", a standard value for content created by a generative AI model, and contains AI_GENERATED=true, the name of the generator (Veyraflow), the identifier of the model, the time of generation and a link to this page. It does not contain your prompt, your name or any information about your account. The marker is stored as follows:

• GLB and glTF: in the asset.extras field of the file's JSON data (the .bin buffer files of glTF have no room for metadata, so the marker is in the .gltf file that uses them);
• OBJ and MTL: in a comment on the first line;
• PNG (textures and previews): in text chunks (AI_GENERATED, Software and Comment);
• WAV: in the LIST/INFO chunk (comment and software fields);
• MP3: in ID3v2 tags (TXXX:AI_GENERATED and a comment);
• OGG: in Vorbis comments;
• FBX: in the custom properties of each object;
• ZIP: every file inside is marked according to its format, and the archive comment contains the marker.

You can view the marker with common tools, such as ExifTool or ffprobe for images and audio, or a glTF viewer for 3D models. The marker is ordinary metadata: editing, converting or optimizing a file can remove it.

Invisible watermark. Sound effects and voice lines, as well as the color textures (inside GLB models and as separate albedo maps) and the preview images of 3D models, also carry an invisible watermark: a faint pattern hidden in the sound or the image itself that our detection software can recognize. It is designed not to change how an asset looks or sounds in any noticeable way, and it does not limit how you can use the asset. Unlike the marker, it usually survives format conversion, compression (for example to MP3, OGG, JPEG or WebP), resizing, turning or flipping an image by 90 degrees, trimming and changes in volume or pitch. Heavy editing, strong compression, cropping an image, mixing a sound with other audio, and very short or pure-tone sounds can weaken it or make it undetectable; for voice lines it is kept especially faint so that it cannot be heard, and is therefore less resistant. Maps that store technical data rather than color, such as normal, roughness or metallic maps, are left unchanged. The watermark contains no information about you, your account or your request; it only shows that the content was generated with Veyraflow. To carry the watermark, voice lines are delivered as WAV files. Music generated with Lyria carries SynthID instead, a watermark that Google embeds in the audio itself and that, according to Google, is designed to survive common editing. Assets generated before 2 October 2026 do not carry our watermark.

6. YOUR RESPONSIBILITIES

• Do not deceive. As stated in Section 8 of the Terms and Conditions, you must not use Generated Assets to deceive people, for example by presenting AI-generated voices or media as genuine recordings of real people.
• Disclose realistic imitations. If you publish AI-generated images, audio or video that resemble real people, objects, places, entities or events and could falsely appear authentic (a "deepfake"), the EU Artificial Intelligence Act (Regulation (EU) 2024/1689, Article 50) requires you, as the person publishing them, to disclose that they were generated or manipulated by AI. For evidently artistic, creative, satirical or fictional works, such as most games, this can be done in a way that does not spoil the work, for example in the credits.
• Use only material that you are allowed to use. Upload only reference photos that you have the right to use, and do not create assets that imitate real people without their permission.
• The marker and the watermark do not replace your disclosure. If the AI marker or the watermark is removed from a file or weakened, for example when you optimize or edit the file for your game, your obligations described in this section still apply.
• Platform rules. Some game stores and platforms ask developers to disclose AI-generated content. The information on this page can help you answer such questions.

7. YOUR DATA

• What we send to AI providers. The content of your generation requests (prompts, voice-line text and instructions, style guide notes and reference images) is sent to OpenAI and fal.ai, and through fal.ai possibly to the developer of the selected model. fal.ai keeps the files of a generation — the reference images you attach and the files the model returns — for no longer than 24 hours and then deletes them automatically; during that time each of these files can be opened by anyone who has its unguessable link. The links are known only to us, fal.ai and, where applicable, the developer of the model, and we do not publish them. fal.ai does not keep the content of your requests for its request history. We do not send them your email address, password or payment details. Do not include personal data about other people in your requests unless you are allowed to.
• Packs. The texts and photos of a pack are processed on machines rented from RunPod, which may be located in the United States. A machine receives only the items it works on, deletes its working files after each item and does not receive your email address or payment details.
• No training by us. We do not use your prompts, uploads or Generated Assets to train AI models ourselves. The AI providers process this data as our service providers, under their agreements with us and their own terms.
• Processing outside the EU. Some AI providers process data outside the European Economic Area, in particular in the United States, with the safeguards described in Section 7 of the Privacy Policy.
• More information. How we handle personal data when you use AI features, including your rights, is described in Sections 4, 6 and 7 of our Privacy Policy. How long we keep Generated Assets is described in Section 22 of the Terms and Conditions.

8. HUMAN OVERSIGHT

No person reviews your request or the Generated Asset before you receive it. Decisions of the content filter, and the warnings and suspensions that may follow them, are made automatically. You can ask for any such decision to be reviewed by a person by writing to janeknastaly50@gmail.com. Our administrators can also review activity and suspend or restore accounts manually.

9. LICENSE FOR GENERATED ASSETS

You may use Generated Assets freely, including in commercial games and in other products that you sell, under the license described in Section 3 of the Terms and Conditions. That license is subject to applicable law, the rights of third parties and any restrictions in the terms of the AI model providers. We do not claim ownership of Generated Assets.

10. CHANGES TO THIS AI DISCLOSURE

We update this AI Disclosure when we change the models or providers that we use, the way requests are processed or the way we label AI-generated content. The date at the top of this page shows when it was last updated.

11. CONTACT

If you have questions about how we use AI, contact us at janeknastaly50@gmail.com or by phone at +48 792 466 019. Our full contact details are listed in the Impressum.