Standalone Local Orchestration Platform - 3D Asset (Beta) Edition
Who is behind it
I am Imre Tőke, an IT engineer and Lead/Senior FX TD. I have spent the last 16 years on effects for feature films and game cinematics across studios like WetaFX, MPC, and Digic Pictures. Over the last two years, I’ve been experimenting with Hybrid local generative AI pipeline solutions as well, and I would like to introduce my latest development.
Before you read further — will it run?
This is a workstation tool, not a web service. It drives models on your GPU.
You need:
-
Windows 10 or 11, 64-bit
An NVIDIA GPU. 8 GB VRAM is the realistic floor. 16 GB is comfortable. 24 GB is
roomy. The platform sizes what it offers you to the card you have, rather than
offering you something that will not load.A recent NVIDIA driver — one new enough for CUDA 12.8. The installer checks yours and tells you before it downloads anything.
-
Free SSD space. It depends on what you tick, and the plan screen shows the
exact figure for your selection — itemised, component by component — before a
single byte moves.
-
What you choose Download
3D only, no image model, no prompt writer ~17 GB
The default 3D setup ~37 GB
Default + light prompt writer (8 GB cards) ~41 GB
Default + strong prompt writer (16 GB cards) ~54 GB
Default + heavy prompt writer (24 GB cards) ~63 GB
Everything on offer ~174 GB
Plan for 80 GB free and you will not have to think about it again.
You do not need:
-
ComfyUI. The platform installs it, configures it and registers it with itself. If you already have one, it will find and use it instead.
-
Python, Git, Visual Studio, a CUDA Toolkit, or any knowledge of any of them.
-
A subscription, an account, or an internet connection once setup has finished.
-
A licence key. During the beta there is nothing to activate and nothing to type in.
-
Not suitable for Importing to Unreal Crowds and large character batches. Single characters, yes. A crowd of hundreds through to Unreal is being built and is not in this beta.
If you are on macOS, on an AMD card, or on a laptop with 6 GB of VRAM, this beta is not for you yet. I would rather tell you now than take your evening.
Features:
-
★ A picture becomes a 3D asset, on your own machine. Describe an object or drop in an image; it generates a concept plate, turns it into a textured mesh, and files the whole thing in an organised folder. Nothing is uploaded, nothing is queued on somebody's server, and there is no per-image cost.
-
What arrives is a finished asset, not a blob. Every 3D run produces the mesh with its texture, a base-colour map, a metallic/roughness map, a tangent-space detail normal map, a render of the geometry's own normals, four orthographic views, a labelled contact sheet and a turntable GIF — named, versioned and filed under *Pack / Asset / *.
-
Batch production, two ways. The 3D Asset Factory takes a list of objects and a style and works through it line by line — concept, mesh, filed folder — one asset at a time so a single failure never costs the run. Or upload your own images and each one becomes an asset, with no concept image generated at all.
-
★ Judge a mesh without opening any Software. The built-in viewer shows the model textured, as untextured geometry, as wireframe and as normals — one click apart, with the light under your control, and the arrow keys to step through a whole pack. A texture hides a bad surface; this is how you see it in ten seconds instead of ten minutes.
-
★ A rigged character straight into your level. A companion script for Houdini exports your animated character and hands it to the platform, which imports it into Unreal as a skeletal mesh — skeleton, animation and material — and puts the clip on the actor, playing, in the level. One press. Finished assets also go into a running Houdini session as a built node graph, rather than as a folder you wire up by hand.
The bridges are optional: everything else works with neither Houdini nor Unreal installed, and your assets land in your output folder either way.
-
The models are current, and named. Image generation on SDXL, ZImage Turbo and Krea 2; multi-angle camera work and background removal on Qwen Image Edit; geometry on Trellis 2, including a four-view projection path for better texturing. You choose which of them exist on your disk, and you are told what each one is licensed for before it downloads.
-
A local prompt writer, sized to your card. Three tiers on your own Ollama, and "none" is a supported answer that installs nothing. The vision tier can read an image back into a prompt.
-
★ A setup that tells you everything before it does anything. The plan screen lists every component it would install — name, version, size, licence, whether that licence permits commercial use, the exact URL it comes from and the folder it lands in — and a Simulate switch walks the entire installation without touching your disk. Every download is checksum-verified on arrival, and anything that fails leaves nothing behind.
-
Nothing leaves your machine. No telemetry, no analytics, no crash reporting. It binds to localhost. Your prompts, your files and your file paths stay where they are. The only outbound connections are the downloads you approved on the plan screen, each from the source named there.
AI transparency, and which duty is whose.
-
Every output image carries a machine-readable AI marker embedded in the file
itself — the IPTC DigitalSourceType term for content created by a generative
model, in the metadata, in XMP and in the PNG text chunks — written automatically on
every run, with no setting that turns it off. That is the provider's duty under
Article 50(2) of the EU AI Act, and it is mine.
-
Mesh files carry the provenance sidecars but no embedded marker, because no established machine-readable marker for a mesh exists; Article 50(2) names audio, image, video and text.
The renders of a mesh — turntable, contact sheet, every orthographic view — are images, and are marked like any other.
-
Disclosing a deepfake when you publish it is Article 50(4), and that duty is yours: I never see what you publish. The application states all of this in plain language on its Legal tab, and again in the workspace where the work happens.
Honest limitations
-
It is a beta. Things will break, and the version you get will not be the version that ships.
-
The installer is the newest part of it. It has been run from nothing — no ComfyUI, no Python, no CUDA — on a clean machine, end to end, and it works. What it has not met is the hundred variations of Windows, drive layout and antivirus that a hundred testers will bring. If it stops, it is built to stop loudly and tell you which step and why — send me that and I will get you moving. This is the single thing I most want tested.
-
It will not quietly change a ComfyUI you already have. If you point it at an existing installation running Python 3.13, it stops, explains why the 3D components cannot load there, and offers to build a separate backend in an empty folder instead. Your working setup is left alone. Anything it does install is named on the plan screen first.
-
The environment doctor reads your setup and tells you what is wrong; it does not yet repair everything for you.
-
Windows will warn you the first time you run it — "Windows protected your PC", unknown publisher. The build is not code-signed yet; a certificate comes before the paid release, not before a free beta. More info → Run anyway. The SHA-256 fingerprint of every build I send is in the email that sends it, so you can check that what arrived is what I published.
-
3D reconstruction is very good at props and very variable at characters. Set your expectations by what you see in the examples, not by the best result on the internet.
-
The models are not mine, and I cannot keep them available.
This platform installs and drives components published by other people — model weights, node packages, the render backend itself — from their own servers, under their own licences. Any of them can be withdrawn, put behind a login, re-licensed or replaced with an incompatible version by its publisher, overnight and without telling anyone. If that happens it is not a fault in this software and I cannot undo it.
What I do instead:
every component is named — version, licence, exact source URL — before it downloads, so what you installed is always a matter of record; and where something I relied on disappears, I will say so and, where I reasonably can, name an alternative in a later update.
What you have already downloaded is unaffected. It stays on your disk and keeps working — so keep your own backup of the model files you depend on. That is the only real protection against somebody else's decision, and it is one you can take today.
-
This is one person. I write it, I test it, and I answer the email — which is why you get direct access to the person who wrote the thing rather than a ticket queue. It is also the limit: replies come in days, not hours, and there is no support rota behind me. The licence inside the build says the same in clause 6A, so you are not finding it out later.
What you get in the beta
The full PRO build (3D Asset Factory, Image Edit, Houdini & Unreal bridges). Local prompt writers sized to your card (8/16/24 GB). Direct access to me. Anything you generate is yours.
A local prompt writer, sized to your card. Three tiers, and "none" is a supported answer that installs nothing:
-
Light — 8 GB cards. Qwen 3 · 4B, 2.5 GB. Small enough to sit beside the renderer.
-
Strong — 16 GB cards. Qwen 3 · 14B, 9.3 GB, plus Qwen 2.5-VL · 7B (6 GB) so it can read an image back into a prompt.
-
Heavy — 24 GB cards. Qwen 3 · 30B-A3B, 19 GB — a mixture-of-experts model, so it answers roughly at the speed of a small one — plus the same vision model.
All Apache-2.0, all pulled through your own Ollama, none of them mirrored by me. One model is loaded at a time, so the tier's download total is larger than the VRAM it needs.
How you get it. When your seat opens I send you a download link and a short setup guide, by email. Everything the platform needs it installs itself, into a folder you choose — nothing goes into Program Files, nothing touches your system Python, nothing is added to your PATH.
Nothing to activate. There is no beta key. Your build carries its own signed stamp with the cohort's end date inside it, and everything is unlocked until then the moment you run it — no key, no account, no activation step. If a screen ever asks you for a licence key during the beta, something is wrong and I want to hear about it that day.
Price
Not set yet — deliberately.
The full licence will be a one-off payment, perpetual, with twelve months of updates included. The Starter tier will stay free. What the licence costs is one of the things this beta is for: I would rather hear what it is worth to the people using it than pick a number in advance and then defend it.
The beta is given, not sold — and that is not only a promise on a web page. Clause 0 of the licence agreement inside the build says it in writing: nothing was bought, nothing was paid, no price is quoted anywhere in the software, and the clauses covering purchase, refunds and the right of withdrawal do not apply to you until you actually buy a licence.
Beta participants get the best terms I offer anyone, and you will hear from me before they are public. If you want to be told when it launches but do not want a beta seat, there is a box for that on the form.
What I am asking for
This is a closed beta with a small number of places, and I am picking for fit rather than speed. In exchange for early access I am asking for one thing: A written note or a short call at the halfway point — what you tried to make, what worked, what got in your way, and what you would not pay for. That is the whole deal. No forum to moderate, no bug tracker to learn. Just send the feedback to my email.
What happens when the beta ends
The Factory, Image Edit and the Houdini and Unreal bridges pause. Create, the Prompt Studio, the Library and System keep working with no time limit. Everything you generated stays in your output folder, with its provenance sidecars, and it is yours — commercial use included. Nothing is removed, locked or watermarked. The application tells you the same thing itself, in the same words, on the day.
Build game assets without leaving your machine
A desktop orchestrator that turns local generative models into a real asset pipeline — written idea to textured 3D mesh, filed with full provenance, straight into Houdini and Unreal.
⚠️ Important: If you are browsing on an Apple device, please use Chrome, Edge, or Firefox to submit this form. Safari's strict privacy settings currently prevent the form from saving your application.
Closed beta · [SEATS] places left · applications close [COHORT CLOSES]
I read every application myself and reply to everyone. Your details are never sold, shared or used for anything other than what you ticked above. See the privacy notice below.
Privacy notice — beta applications
Who is responsible. Imre Tőke, timrefx@gmail.com.
I am the data controller for the information you submit through this form.
What I collect.
Your name, email address, occupation, GPU and system details, your answers about your intended use and your pipeline tools, and which consent boxes you ticked. Why, and on what basis. To assess and administer your beta application and to contact you about the beta programme — on the basis of your consent.
If you ticked the second box, also to tell you when the product launches, again on the basis of your consent. Those are two separate consents and you can withdraw either one independently. What I do not do. I do not sell, rent or share your details with anyone.
There is no advertising, no profiling and no automated decision-making. The information goes no further than my own inbox and the form provider below.
Who processes it for me.
This site runs on Wix, which stores form submissions on my behalf as a data processor.
How long I keep it.
Beta application details: until the beta programme ends, and up to twelve months afterwards so I can honour the beta participants' terms and invite you to a later cohort. Launch-notice consent: until you withdraw it. After that everything is deleted.
Your rights. You may ask me for a copy of what I hold, ask me to correct it, ask me to delete it, or withdraw your consent at any time — email timrefx@gmail.com and I will act on it.
You also have the right to complain to your national data protection authority; in Hungary that is the NAIH.
The software itself collects nothing. Once installed, the application sends no usage data, no analytics and no crash reports. It contacts the network only when you activate a licence — and during the beta it does not do that either. Your prompts, your generated files and your file paths never leave your machine. The only outbound connections it makes during setup are the downloads you approved on the plan screen, each from the source named there.