Skip to content

Custom Model Training

Train AI Models That Match Your Art Style

Off-the-shelf models don't know your game. They produce generic output that your art director rejects, voices that sound like nobody in your cast, and animation that belongs to someone else's character. The gap between "AI-generated" and "on-brand" is the single biggest blocker to putting AI into a real production pipeline.

Custom Model Training closes that gap. Layer trains on your own material — your artwork, your voice recordings, your reference footage — so every generation starts from what your game already is. Teams at 400+ game studios and entertainment brands use trained styles to make AI output that actually belongs in their game.

Key Capabilities

LoRA Training on Your Artwork Build a Reference Set from your existing art, pick a base model, and train. You can train multiple LoRAs per set, so one collection of artwork can power several styles tuned for different jobs.

Edit-LoRA Training Teach a model how your team edits, not just how it draws. Edit-LoRA training learns from paired before/after images, capturing transformations like your studio's cleanup pass or a signature recolor treatment.

Instant Style Clone Recraft V4 derives a style straight from reference images, raster or vector, with no training run to wait on. It's the fastest route from a handful of references to a style the whole team can generate against.

Voice Cloning and Voice Design Clone a voice from a sample of real audio, or design one from a written description when nobody has recorded a line yet. Either way you get a reusable speaker that narrates any script — barks, dialogue, or store copy — in a consistent voice.

Motion Training Train a motion from a text prompt or lift one out of a reference video, and retarget it onto any rigged character. A run cycle trained once becomes animation every character in the game can use.

Auto-Train When a Reference Set qualifies, one click starts a training run with settings Layer chooses for you. Artists get a production-ready style without touching a single hyperparameter.

Caption Review Before Every Run Layer generates captions for your training images automatically — and lets you review and edit them before training starts, so the model learns exactly what you intend.

Full Run Visibility Every run shows an up-front price estimate in Creative Units, live training progress, and the option to cancel mid-run. Run history keeps a record of what was trained, when, and from what.

How It Works

Curate a Reference Set Gather what defines the thing you're training — artwork for a style, voice samples for a speaker, a reference video for a motion — into a Reference Set. This is the ground truth your model learns from.

Train or Auto-Train Choose a base model and configure the run yourself, or let Auto-train handle the settings. Review the AI-generated captions, check the price estimate, and start.

Generate With It Your trained style, voice, or motion lands in the styles library, organized and ready for the whole team. Anyone in the workspace can generate with it immediately.

Built for Game Production

Consistent Game Art at Scale A trained style keeps hundreds of icons, characters, and environments visually coherent, so live-ops content shipped in month twelve matches the art you launched with.

Characters That Look and Sound the Same A character trained as a style, a voice, and a set of motions stays itself across every surface — key art, in-game dialogue, and the animation that carries them.

On-Brand UA Creative Marketing teams generate ad variations in the exact style of the game, keeping store pages and playable ads aligned with what players actually see.

One Library for Every Style Trained styles, cloned voices, trained motion, imported LoRAs, uploaded checkpoints, and model URLs all live in a shared styles library. Your whole team pulls from the same source of truth instead of passing files around.

Fast Starts, Enterprise Guardrails 40+ pre-built models let you generate on day one while your custom styles train. Your artwork stays yours — customer data is never used to train foundation models, and Layer is SOC 2 Type II compliant with role-based access.

Your art style is your competitive advantage. Custom Model Training makes it the starting point of every generation instead of something you fix in post.

Custom Model Training — FAQ

What do I need to start training a custom model?+

A Reference Set built from your own material — artwork for a style, audio for a voice, a reference video for a motion. Layer trains against a base model you choose, and you can train multiple models per set.

What kinds of custom models can I train?+

Four. LoRAs and Edit-LoRAs on image models, style clones on Recraft V4, voices on text-to-speech models, and motion on Uthana Animate. Each one starts from a Reference Set and lands in the same styles library.

How does voice cloning work on Layer?+

Upload a sample of the voice you want and Layer clones it into a reusable speaker you can point at any script. ElevenLabs TTS V3, Qwen 3 TTS, and Seed Audio 1.0 all support cloning from reference audio.

Can I create a voice without a recording?+

Yes. Voice design builds a speaker from a written description — "gravelly old sailor, warm and unhurried" — on ElevenLabs TTS V3 and Qwen 3 TTS, so you can cast a character before anyone records a line.

Can I train animations?+

Yes. Uthana Animate trains motion from a text prompt or from a reference video, and the trained motion retargets onto any rigged character you animate afterwards.

What is Auto-train?+

When a Reference Set qualifies, Auto-train starts a run in one click with settings Layer chooses for you — no hyperparameter tuning required.

Can I see the cost before a training run starts?+

Yes. Layer shows an up-front price estimate in Creative Units before every run, and you can watch live progress or cancel a run in flight.

Can I bring models I trained elsewhere?+

Yes. Upload a LoRA, a checkpoint, or a voice embedding, or paste a model URL, and it becomes a style in your workspace alongside models trained on Layer.

Is my artwork used to train foundation models?+

No. Your data is used only for the styles you train in your workspace, and customer data is never used to train foundation models.

Try Custom Model Training today

Train your first custom model on your own artwork today. Layer is free to start — no credit card required.