MifBridge lets an AI assistant drive Unreal Engine and Blender through typed editor commands, and the engines build every asset. This page shows where the line with generative AI falls.

Used as described here, the assistant works like a fast pair of hands in the editor: it chooses commands, and the engine builds every asset from parameters you control. The sections below explain how that works, which features cross the line, how to keep authorship of your work, and what to tell clients and stores. It covers MifBridge 1.1, the version on sale, with Unreal Engine 5.3 to 5.8 and Blender 3.6 to 5.2. Anything marked "from 1.2" is in the next release and not in 1.1. Sources were checked on October 1, 2026. This page is not legal advice.

How MifBridge works

MifBridge is two programs that sit between an AI assistant and your editors: a plugin inside Unreal Engine and an add-on inside Blender, both reached through one MCP server (the standard way assistants call tools). The assistant never draws, models or paints anything itself. Each request goes through the same steps:

You describe the job. "Add a point light 300 cm above the table, warm white, intensity 8."
The assistant picks a typed command and fills in its parameters. Each command has a fixed list of named, typed parameters. A parameter name MifBridge does not recognize is refused, with the accepted names listed.
MifBridge checks the call before it runs. Wrong names, wrong types, impossible values and edits to protected engine files are refused with a message saying why.
Unreal or Blender carries it out. MifBridge calls the engine's own editor and scripting functions, the same systems you use when you work by hand. The engine builds the result.
MifBridge reads the result back. For most changes it reports what actually changed in the editor, so you see the outcome as well as the request.
Calls are journaled. Unreal and Blender each keep a journal of the calls MifBridge runs, on by default: what was asked, with what values, and whether it worked. Together with your version control history, that is your record of who did what.

The assistant itself is an AI language model. What it produces is commands and code. What ends up in your project is built by Unreal and Blender. That difference is what this page is about.

Where the line is

What decides it is what made the content. Which program was open does not matter.

Editor automation: the engine builds it

The assistant chooses a command and values, for example "place a cube at 0, 0, 0 sized 100 × 100 × 10" or "set this material's roughness to 0.4". The engine builds the result with its own code, using the same systems you would use by hand. Nothing in the result comes from a trained model.

Generative AI: a trained model produces it

A model trained on large collections of other people's images, meshes, audio or text outputs the content itself: the pixels of a texture, the shape of a mesh, the words of a line of dialogue. Image generators, text-to-3D tools and AI voice tools work this way.

Nearly all of MifBridge is editor automation. The next section lists the one feature group that is generative, and one that needs care.

Which features are which

Unreal Engine tools: levels, actors, Blueprints, materials, UI widgets, Niagara VFX, audio, landscape, animation, data tables. Not generative. They run Unreal's own editor functions with the values given, then read the result back.
Blender tools: modeling, UVs, rigging, materials, physics, rendering, compositing. Not generative. They run Blender's operators and Python API with the values given.
Procedural builders: bl_build_tree; from 1.2 also bl_build_rock, bl_build_plant and bl_build_kit_piece. Not generative. Hand-written algorithms: seeded noise and fixed rules. The same seed and settings give the same shape on a given Blender version. No trained model is involved.
VFX recipes: campfire, sparks, smoke, explosion and others. Not generative. A recipe turns a named effect into a fixed list of Niagara build steps with measured values. It is a lookup table.
Blender to Unreal hand-off: mif_send_to_unreal. Not generative. It exports your Blender mesh and imports it into Unreal as a static mesh with its collision, pivot and material slots, then reads the asset back to check it.
Run Python: run_python. AI-written code. It runs Python the assistant writes inside the editor. The code is AI-written; anything it builds is still made by the engine. Treat it like any AI-assisted code.
Generated assets (optional): bl_gen_image, bl_gen_mesh, bl_gen_texture, bl_gen_asset, bl_gen_status. Generative. They send requests to a ComfyUI install you run yourself, which uses the FLUX image model and the Hunyuan3D-2 mesh model to create images, meshes and textures.

The generative tools cannot run on their own. MifBridge does not ship ComfyUI or any AI model. The five bl_gen_* tools only work if you install ComfyUI and the models yourself and keep it running. Without it, those tools stop with a "cannot reach ComfyUI" error and nothing is generated.

From MifBridge 1.2, each project chooses a mode. The first time MifBridge is used in a project, the assistant asks you: generative (the bl_gen_* tools may create images, meshes and textures with AI models) or non-generative (MifBridge only drives the editors). Your answer is saved with the project (the mif_project_mode tool records it), in MifBridgeProject.json beside the Unreal project file and inside the .blend for Blender. Choose non-generative for any project that must contain no AI-generated content. Also from 1.2, you can switch generation off for everything with the server setting MIF_GENERATIVE=off or the "Allow generative tools (ComfyUI)" preference in the Blender add-on. While it is off, the bl_gen_* tools are hidden from the assistant, any attempt is refused before ComfyUI is contacted, and status reports and call journals record the mode.

Who makes the creative choices

Even with no generative feature in use, a second question matters: did a person decide what the result looks like, or did the assistant? In the United States, copyright law cares about this. The US Copyright Office's report Copyright and Artificial Intelligence, Part 2: Copyrightability (released January 29, 2025) concludes that using AI to assist in creating a work does not bar copyright, that the mere provision of prompts is not enough, and that copyright can cover a person's creative arrangements or modifications of AI output. It also says that choosing one AI output from several does not by itself make you its author. Other countries have their own rules.

You make the creative choices: "Place these 50 building pieces at the positions in this layout sheet, rotate the corner pieces 90°, apply the brick material we made." The creative decisions came from you. The assistant carried them out.
The assistant is choosing: "Build me a medieval village." The assistant decides which houses go where, how tall and what color. Under the Copyright Office's view, those raw choices may not be protected as your work unless you rework or arrange them creatively yourself.

Both requests use only non-generative tools. The difference is who decided. A port of an existing game is the easy case: the layouts, art and design already exist and belong to the studio, and MifBridge rebuilds their choices in a new engine.

Good and risky requests

Request	Rating	Why
"Import the crate from crate_v3.blend, set its pivot to the bottom center, add one convex collision hull, and send it to /Game/Props."	Good	Every value is yours.
"Make some props for a warehouse."	Risky	The assistant picks what props exist and what they look like.
"Wire this door Blueprint so it opens 90° over 0.6 s when the player overlaps the trigger and holds the key item."	Good	The behavior is specified in full.
"Design a fun puzzle for this room."	Risky	The design is the assistant's.
"Recreate the lighting from the Unity scene: sun at 42° elevation, color 255, 236, 210, intensity 6 lux."	Good	The values come from the original project.
"Generate a brick texture."	Avoid	If ComfyUI is installed this would use a generative model. Use a texture you made, bought or licensed.
"Build a young tree with bl_build_tree, 4 m tall, seed 3, then send it to Unreal."	Good	A procedural algorithm with values you chose.
"Write the quest dialogue for the bartender."	Avoid	AI-written text that players read is AI content. Write it yourself, or rewrite any draft fully.

If you use the generative tools

The bl_gen_* tools run two third-party AI models through your own ComfyUI. MifBridge does not ship them; their terms are between you and each model's publisher. Read the license before you use either, and check it again later, because terms change.

Hunyuan3D-2: restricted

Meshes and textures. License: Tencent Hunyuan 3D 2.0 Community License.

Not licensed for use in the European Union, the United Kingdom or South Korea.
Products with more than 1 million monthly active users need a separate license from Tencent.
Outputs used publicly must be "expressly and conspicuously identified" as machine-generated.
Outputs may not be used to improve other AI models.

FLUX.1 [schnell]: commercial use allowed

Images, and MifBridge's default. License: Apache 2.0. Licensed for personal, scientific and commercial use.

FLUX.1 [dev]: non-commercial model

Images, only when you choose variant "dev". License: FLUX.1 [dev] Non-Commercial License. The model itself may only be used for non-commercial purposes without a license from Black Forest Labs. Its outputs may be used commercially, must carry an AI-generated disclosure where the law requires it, and may not be used to train a competing model.

Anything these models produce is generative AI content. Disclose it where your store or publisher asks, and keep it out of any project that promises to contain none.

The AI lawsuits and what they mean here

The major US lawsuits over AI (against OpenAI, Anthropic, Meta, Stability AI and others) are about companies training models on copyrighted material without permission. Courts have split so far. In Bartz v. Anthropic, training on lawfully bought books was held to be fair use while keeping pirated copies was not, and the case settled. In Thomson Reuters v. Ross, the Third Circuit held in September 2026 that the AI training at issue was not fair use.

MifBridge trains no model and ships none, so these cases are about a different kind of tool. The open question for users is AI output, which only arises when a generative model creates content. A project that uses no generative tool avoids that question.

A safe workflow

Decide first, then ask. Bring concrete specifications: sizes, positions, names, colors, references you made or own, values from the original project.
Work in small steps. One system, one asset or one area at a time, so each result gets reviewed before the next. Coming in 1.2: any call that would change more than 25 items returns a preview of the changes instead of writing them.
Review every result in the editor. MifBridge reports what it changed, and you approve what stays. Move, change or rebuild anything that is not what you meant.
Make art in Blender, by hand or with the procedural builders, and hand it to Unreal with mif_send_to_unreal. Do not use the bl_gen_* tools, and from 1.2, choose non-generative when the assistant asks about the project.
Keep the record. Leave MifBridge's call journals on and keep your version control history. Together they show who decided what, in what order.
Tell people. If you work for someone else, tell them in writing which tools you use and how.

Do and avoid

Do

Use MifBridge for setup, conversion, porting, wiring, placement and repetitive editor work.
Give exact values and references you own.
Review and edit what the assistant builds.
Use the procedural builders with values you chose.
Write player-facing text yourself, or fully rewrite any draft.
Keep call journals and commit history.

Avoid

Installing ComfyUI or calling any bl_gen_* tool.
Open-ended creative requests where the assistant designs the content.
Shipping AI-written dialogue, item text, lore or store copy unedited.
Textures, sounds, music, voices or images from any AI generator, inside or outside MifBridge.
Sending a client's source to a cloud service their contract does not allow.

Your data and clients

MifBridge itself runs on your machine. The editor commands, checks and logs stay local. The assistant you connect is a separate choice with its own terms.

With a cloud assistant

Your requests and anything the assistant reads, including source code, pass through the provider. On Anthropic's consumer plans (Claude Pro and Max) with model training turned off, conversations are kept for 30 days and are not used for training. Feedback you send is kept longer. Check the provider's privacy center for the current terms.

With a local model

Nothing leaves your computer. MifBridge works with any assistant that supports MCP, so a locally run model is an option when a contract forbids sending code to third parties.

Before client work

Read the client's contract first. Look for clauses on AI tools, third-party services and confidentiality.
Tell the client in writing which assistant you use and that generative features are not used.
Keep client source out of feedback reports you send to any AI provider.

Store disclosure

Store rules focus on content players see or hear. Steam's content survey, for example, defines pre-generated AI content as "any kind of content that ships with your game and is consumed by players that is created with the help of AI tools during development", and live-generated AI content as content created with AI tools while the game is running. About AI development tools it says: "Efficiency gains through the use of these tools is not the focus of this section."

Following this page, no shipped art, audio or text comes from a generative model.
AI-assisted code (scripts or Blueprints the assistant wrote) is usually viewed as development work rather than content players consume, but Steam's survey does not name code either way.
The final decision on what to disclose belongs to the developer or publisher, with their own advice.

Copy this into a contract note, a studio handbook or an email to a client. Edit it to match what you actually do, and only send it if every sentence is true for your project. The bracketed sentence applies from MifBridge 1.2; drop it on 1.1.

AI tool use on this project

We use MifBridge as editor automation for Unreal Engine and Blender. An AI assistant selects typed editor commands; Unreal Engine and Blender build every asset with their own code from parameters set by a person on our team. Its call journals record every command it runs.

MifBridge's optional generative features (the bl_gen_* tools, which require a separately installed ComfyUI with image and 3D models) are not installed and are not used. [The project's MifBridge mode is set to non-generative.] No art, audio, music, voice or player-facing text in this project is produced by a generative AI model.

A person on our team specifies and reviews every change. AI-assisted code is reviewed before it is kept. Client source is handled according to the agreement, and we can use a locally run model if third-party services are not permitted.

Questions

Isn't the assistant itself generative AI?

Yes, the assistant is a language model. In this setup its output is commands and code, and the game content is built by Unreal and Blender. That is why the line on this page is drawn at what made the content.

Are the procedural trees generative?

No. They come from hand-written algorithms using seeded noise and fixed rules, the same kind of technique games have used for decades. The same settings give the same shape every time, and no trained model is involved. The rocks, plants and kit pieces that arrive in 1.2 work the same way.

What about Unreal's own procedural tools, like PCG?

Unreal's Procedural Content Generation framework and similar tools are rule-based systems built into the engine, with no trained model in them. Using them through MifBridge is the same as using them by hand.

Is AI-written code a problem?

Code is usually viewed as development work, separate from the content players consume, though store rules do not settle this explicitly. Review it like any code you did not write yourself, keep what you understand and change what you need. Whether to mention it to a client or store is your decision.

Can the generative tools be switched on by accident?

In MifBridge 1.1 they need ComfyUI and its models installed and running on your machine or network; without them the bl_gen_* tools stop with an error and nothing is created. From 1.2, choose non-generative when the assistant asks about the project (or set MIF_GENERATIVE=off): the tools are then hidden from the assistant and refused before anything is contacted.

How can I prove generation was off?

In 1.1, the Blender call journal records every operation the add-on runs, the generation ones included, so a journal with none of them is your record. From 1.2, the project's recorded mode is a file you can keep in version control (MifBridgeProject.json), MifBridge's status check (self_audit and bl_status) reports generativeAllowed, the Blender add-on's sidebar and Unreal's MifBridge setup tab show the mode, and every Blender call journal entry records it. Keep those with your project.

How do I show a client what was done?

Share the MifBridge call journals and your version control history. The journals list each command, its values and its result, in order. They are on by default; leave them on for client work.

Does any of this replace legal advice?

No. It describes how the tool works and common practice. For a specific contract, publisher or store submission, ask them or a lawyer.

Terms

Term	Meaning
Generative AI	A model trained on large datasets that outputs new content such as images, meshes, audio or text.
AI assistant	The language model you talk to, such as Claude. With MifBridge it chooses commands and writes code.
MCP	Model Context Protocol, the standard way an assistant calls external tools like MifBridge.
Typed command	A tool call with a fixed list of named parameters and value types, checked before it runs.
Procedural	Content built by hand-written rules and seeded noise. Deterministic, with no training involved.
Call journal	MifBridge's record of each command it runs in Unreal or Blender: what was asked, the values, and whether it worked. On by default.

Checklist before you start

ComfyUI is not installed or not running on this machine (from 1.2: the project's mode is set to non-generative).
The assistant has been told not to use any bl_gen_* tool.
I have concrete specifications or the original project to work from.
I review each result in the editor before moving on.
Textures, audio and images come from me, a purchase or a license.
Player-facing text is written or fully rewritten by a person.
The client's contract allows the assistant I use, and they know I use it.
Call journals are on, and version control history is kept.

Sources and limits

Checked on October 1, 2026. Rules change; check the source before relying on any of these.

US Copyright Office, Part 2 release notice and the Copyright and Artificial Intelligence reports.
Valve, Steamworks content survey documentation, AI-generated content section.
Anthropic, Claude Code data usage and the Privacy Center on data retention.
Model licenses: Hunyuan3D-2, FLUX.1 [schnell], FLUX.1 [dev].
Case status: BakerHostetler AI case tracker; Thomson Reuters v. Ross, Third Circuit, September 29, 2026.
MifBridge's README and documentation, for what each tool does.

What this page is not. It explains how MifBridge works and summarizes public guidance. It is not legal advice, it does not cover every country, and it cannot tell you what a particular contract, publisher or store requires. For those, ask them or a lawyer.

Using MifBridge Without Generative AI