Trainable models
Last updated: November 4, 2026
The base models you can train a finetune on, by kind of finetune: what each is good for, which trainers it offers, and the settings it starts from. Use it to pick a model in Customize, or to see what Automatic will choose.
A finetune works only on the model it was trained on, so pick the model you will generate with. The one exception is a voice trained on an ElevenLabs model, which speaks on every ElevenLabs voice model. For what a run costs, see Creative Units and layer.ai/pricing. What each setting does is in Training settings.
Image finetunes
Trained on single images, used for text-to-image generation. Every model trains on fal by default; two offer a second trainer under Trainer. The defaults below are what the form starts from. Character sets start from gentler defaults on some models, listed where they differ.
On fal, FLUX.1 [dev] is the only image model with a LoRA rank setting; on the others, fal's trainer sets the rank itself. Kohya and AI Toolkit always take a rank.
Automatic trains the first of these that your workspace can use: FLUX.1 [dev], then FLUX.2 [klein] 9B, then FLUX.2 [dev], then FLUX.2 [klein] 4B.
FLUX.1 [dev]
An all-round choice for characters, styles and items, and the model Automatic tries first. Trainers: fal (default) and Kohya.
On fal: 200 epochs, learning rate 0.0005, LoRA rank 32. On Kohya: 300 epochs, learning rate 0.00005, LoRA rank 90. Character sets, on either trainer: 90 epochs, learning rate 0.00005, LoRA rank 16.
FLUX.2 [dev]
A larger model with stronger prompt following and detail, for high-quality finals. Trainer: fal. Defaults: 200 epochs, learning rate 0.00005.
FLUX.2 [klein] 9B
The larger klein model, balancing detail and speed for production volume. Trainer: fal. Defaults: 100 epochs, learning rate 0.00005.
FLUX.2 [klein] 4B
The lightest of the family, for fast iteration and large batches. Trainer: fal. Defaults: 100 epochs, learning rate 0.00005.
Qwen-Image
Strong at complex prompts and at legible, stylised text in the image, useful for UI, signage and titles in a house style. Trainers: fal (default) and AI Toolkit, which adds extra settings (see Training settings).
On fal: 100 epochs, learning rate 0.0005. On AI Toolkit: 100 epochs, learning rate 0.0001, LoRA rank 32; character sets use learning rate 0.00005, LoRA rank 16 and caption dropout rate 0.3.
Qwen-Image 2512
A later Qwen-Image release with finer detail and closer prompt following. Trainers: fal (default) and AI Toolkit.
On fal: 200 epochs, learning rate 0.0005. On AI Toolkit: 100 epochs, learning rate 0.0001, LoRA rank 32; character sets use learning rate 0.00005, LoRA rank 16 and caption dropout rate 0.3.
Krea 2 Turbo
A speed-tuned model with a wide aesthetic range, for quick iteration. Trainer: fal. Defaults: 200 epochs, learning rate 0.0005.
Edit finetunes
Trained on before-and-after pairs, used when you edit an image on the same model. Every edit model trains on fal. Qwen-Image Edit 2511 and FLUX.1 Kontext [dev] have a LoRA rank setting; on the others, fal's trainer sets the rank itself.
Automatic trains an edit finetune when the set has enough complete pairs, on the first of these your workspace can use: Qwen-Image Edit 2511, then FLUX.2 [dev] Edit, then FLUX.1 Kontext [dev], then FLUX.2 [klein] 4B Edit, then FLUX.2 [klein] 9B Edit.
Qwen-Image Edit 2511
Prompt-guided edits such as angle changes, restyles and composition changes, and Automatic's first choice for pairs. Defaults: 100 epochs, learning rate 0.0001, LoRA rank 32.
FLUX.2 [dev] Edit
Detailed, high-quality edits, including edits guided by several reference images. Defaults: 200 epochs, learning rate 0.00005.
FLUX.1 Kontext [dev]
Context-aware edits from a text instruction, good for iterating on one image. Defaults: 200 epochs, learning rate 0.0001, LoRA rank 32.
FLUX.2 [klein] 9B Edit
The higher-quality klein edit model. Defaults: 100 epochs, learning rate 0.00005.
FLUX.2 [klein] 4B Edit
The lighter klein edit model, for fast edits at volume. Defaults: 100 epochs, learning rate 0.00005.
Image styles
Trained from reference images into a style held by the model itself, with no LoRA and no weights to download. Each model trains its own style, so a style trained on one is not available on another. Each run takes between 1 and 10 reference images (extra images are left out), each with a weight from 0.1 to 2, and Match offers Precise or Flexible. Automatic doesn't train image styles.
Recraft V4
Raster illustration and design work in your style.
Recraft V4 Pro
The higher-fidelity raster tier, with finer detail.
Recraft V4 Vector
Scalable vector art (SVG) in your style: logos, icons and brand illustration.
Recraft V4 Pro Vector
The higher-fidelity vector tier.
Voices
Trained from audio clips (Clone), from a description (Design) or, on ElevenLabs models, from the ID of a voice that already exists on ElevenLabs (Voice ID), used to speak any script. A flat charge per run.
The three ElevenLabs models below train voices the same way, and a voice trained on any of them speaks on every ElevenLabs voice model. Voice ID brings in a voice from the ElevenLabs voice library, or one made on Layer, by its ID; a voice saved only to your own ElevenLabs account can't be reached. See Train a custom voice.
ElevenLabs TTS V3
Expressive speech with emotion and delivery tags. Clone blends several reference clips. Design takes a voice description of 20 to 1000 characters, an optional reference clip, and offers Loudness (-1 to 1), Guidance scale (0 to 100) and Prompt strength (0 to 1). Voice ID takes the voice's ID.
ElevenLabs TTS V4
Expressive speech shaped with inline audio tags such as [whispering]. Clone, Design and Voice ID work as on ElevenLabs TTS V3, with the same description length and the same ranges.
ElevenLabs TTS V4 Turbo
The faster V4 tier, for drafts and volume. Clone, Design and Voice ID work as on ElevenLabs TTS V3.
Qwen 3 TTS
Multilingual speech. Clone works from one clip of up to 300 seconds, with an optional transcript of what it says; Design works from a description, with a script you write or one Layer supplies (Auto script).
Seed Audio 1.0
Natural, expressive speech. Clone only, from one clip.
Motions
A 3D motion any rigged character can play. A flat charge per run.
Uthana Animate
Text generates a motion from a description of 1 to 1000 characters, on model version text-to-motion-3.0 (default) or text-to-motion. Video extracts the motion from exactly one reference video, on video-to-motion-2.1 (default) or video-to-motion-2.0.