Integrations
Providers are routes, not the system.
Swap one and the definition does not change.
What it talks to, and what that costs you.
The rule
A provider is infrastructure. It is not the product.
Models and services change. Prices change, licences change, quality changes, and the one that is best this year will not be next. A team might use a local model for privacy, a rented GPU for power, a hosted service for speed, or a recorded human performance because it is the right creative choice.
Those are routes through the system, not the system itself. The definition, the candidates, the review, the target, the history and the replacement path survive a provider being swapped — which is the ninth principle of the manifesto and the reason this page can grow without anything on it being rewritten.
Models you run yourself
Models you run yourself
On your own GPU, or on one rented by the hour when the job outgrows the card. Nothing leaves the machine, nothing is metered, and a seeded definition regenerates the same result — which is what makes a definition a recipe rather than a receipt.
NVIDIA Audio2Face-3D
A sound wave to facial animation, emitting ARKit blendshapes natively. Takes emotion as an explicit input as well as reading it from the audio.
Your machine or rented · Free. Runs in Docker.
via FaceForge ACE
NVIDIA Audio2Emotion
Reads emotion out of a performance, so a face solve is directed rather than guessed.
Your machine · Free.
via FaceForge ACE
StableMaterials and MatFuse
Material candidates generated on your own GPU, reviewed before one is graduated into a PBR map set.
Your machine · Free.
via SurfaceForge Local
Unreal Engine audio-to-face
The engine’s own audio-driven solver: a sound wave in, 81 MetaHuman face-board curves out, with no MetaHuman asset required.
Your machine · Free, local, offline. Ships with the engine.
via FaceForge MetaHuman
Hosted providers
Hosted providers
Your own account and your own key, billed by them and never by us. Every run is priced before it commits, and the key lives in the operating system’s credential vault where no tool can read it back.
PATINA, through fal.ai
Hosted material generation at a higher resolution than the local runner reaches — the route you choose for the final map set, not a fallback for a weak card.
Hosted · Metered, on your own account.
via SurfaceForge PATINA
Drawing the reference
Drawing the reference
A mesh is generated from a picture, so the picture decides most of what you get. This is the single largest quality lever in the pipeline: the same prompt through a local SDXL and through nano-banana produced a smooth white box and, respectively, a machine with legible plate lettering, hazard decals, gauge faces and copper piping. There are two routes, and the difference between them is an account, not a pipeline.
FLUX.2 klein 4B
The default local image model, drawing the reference on your own card. Free, private, nothing leaves the machine, and a seed makes it repeatable. Its weakness is long descriptions — it draws material and wear convincingly but drops clauses, so describe one subject plainly.
Your machine · Free. Apache-2.0, no use restrictions. 8.1GB peak VRAM; about 15 seconds at 8 steps on a 16GB card.
via MeshForge Trellis
SDXL, FLUX.1 schnell, Qwen-Image
The other models the same local runner can load, for a different look or a smaller download.
Your machine · Free. SDXL is OpenRAIL++-M and carries use restrictions, which is part of why klein is the default.
via MeshForge Trellis
Frameworks it plugs into
Frameworks it plugs into
The generation sets know nothing about any game framework, deliberately. Binding them to one is an add-on, which is why the same plugin serves a project built on something else entirely.
MetaHuman
Faces solve onto the MetaHuman face board, and garments fit MetaHuman bodies. One module links the MetaHuman plugin — remove it and everything still loads with one fewer solver.
Free with the engine.
via FaceForge MetaHuman, MeshForge Garment
What it is built on
What it is built on
The seams the platform already gives you, used rather than replaced. Every one of these is somebody else’s standard, and that is the point: a shared standard beats a better private one.
Standards it speaks
Standards it speaks
A vocabulary somebody else defined is worth more than a better one nobody shares. These are the shapes results arrive in, and the reason a clip or a solve can move between characters at all.
ARKit blendshapes
The 52-shape facial vocabulary Audio2Face emits natively, and what a non-MetaHuman character is most likely to already speak.
Free
via FaceForge
MetaHuman face board
81 control curves. The face solve is character-agnostic — only the control names are MetaHuman-shaped, and curves are retargeted into whatever vocabulary your character actually speaks.
Free
via FaceForge
SOMA skeleton
The rig Kimodo generates onto. A clip is built on the SOMA skeleton and moved across by Unreal’s own IK Retargeter, which is what lets one generated motion serve characters with different proportions.
Free
via MotionForge Kimodo
glTF and FBX
What meshes and motion arrive as before import. Interchange already converts to centimetres; the traps are documented rather than hidden.
Free
via MeshForge, MotionForge
Your keys
One place, and nothing can read them back.
A Runpod key or a Hugging Face token is an account you have, not a thing one plugin owns. So they are entered once, on one page, into the operating system’s credential vault — and every set that needs one finds it there. No tool can read a key back out.
That page also says the thing everybody learns the hard way: a token alone is not access. Some models need their terms accepted or their access requested, and a token without that grant installs perfectly and fails much later, somewhere unrelated.
Where a set runs a model locally or on a rented card, the editor and the hub can start it, stop it, and say what a rented one is costing right now — so nothing bills overnight because a window was closed. The hub is where that lives.
Names
Every product name and mark on this page belongs to its owner, and appears here to say what Automation Forge connects to. None of it implies endorsement. Where a mark is not shown it is because we do not have one we are entitled to use, not because the integration is any less real.