ExploreModels
Pricing
Explore

Krea 2 Trainer

All training models

Train a custom LoRA on your own images to teach Krea 2 a new subject, character, or style. Provide a set of training images (and an optional trigger word), and the trainer outputs LoRA weights you can use for inference with the Krea 2 LoRA endpoint.

Training$ Commercial use

About Krea 2 Trainer

Train a custom LoRA on your own images to teach Krea 2 a new subject, character, or style. Provide a set of training images (and an optional trigger word), and the trainer outputs LoRA weights you can use for inference with the Krea 2 LoRA endpoint.

lorapersonalization

Pricing

Training models are coming soon. You can still browse this model's inputs below.

Inputs

images_data_urlfilerequired

URL to a .zip archive of training images (PNG/JPG/JPEG/WebP). Each image may carry a same-stem caption: an image named ROOT.EXT (e.g. 001.jpg) pairs with a caption file named ROOT.txt (e.g. 001.txt). Images with no caption file fall back to `trigger_phrase`, so each image must have either a caption file or a non-empty `trigger_phrase` — otherwise the request is rejected (422).

trigger_phrasetextdefault: ""

Instance phrase used as the caption for any image that ships no .txt caption file (DreamBooth-style); also seeds the default validation prompts. Required when some images have no caption file and auto_captioning is Off.

auto_captioningselectdefault: "Off"

Optional automatic captioning mode. Off preserves the existing behavior: every image must include a same-stem .txt caption or fall back to trigger_phrase. Object/Character, Style, and Custom use dataset-captioner to generate captions with google/gemma-4-31B-it; trigger_phrase is included in the captioning prompt and normalized in the final sidecars using the selected LoRA captioning convention. Cost of captioning will be passed through directly.

Options: Off, Object/Character, Style, Custom

stepsintegerdefault: 100

Number of LoRA training steps.

learning_ratenumberdefault: 0.0005

Learning rate for the LoRA B factor (and for A when lora_a_lr_multiplier=1). Krea recommends 3e-4–7e-4 (constant schedule). For muon this is the match_rms_adamw-scaled (AdamW-equivalent) rate.

resolutionselectdefault: 768

Training resolution (images are cover-resized + center-cropped to a square). 768 trains ~1.7x faster at lower VRAM with near-identical style quality; 1024 maximizes fine detail.

Options: 768

debug_datasetbooleandefault: false

When enabled, returns a downloadable archive of the exact dataset used for training after caption backfill or auto-captioning.

More from Krea 2 Trainer

Train Flux LoRA
Train Flux LoRA

Train styles, people and other subjects at blazing speeds.

lorapersonalization
training
Train Flux LoRAs For Portraits
Train Flux LoRAs For Portraits

FLUX LoRA training optimized for portrait generation, with bright highlights, excellent prompt following and highly detailed results.

lorapersonalization
training
FLUX 3 Klein LoRA Trainer✦ new
FLUX 3 Klein LoRA Trainer

Fine-tune FLUX 3 Image Klein with your own images to teach it a subject, a product or a style.

fluxlora
training
Z Image Trainer
Z Image Trainer

Train LoRAs on Z-Image Turbo, a super fast text-to-image model of 6B parameters developed by Tongyi-MAI.

turboz-image
training
FLUX 2 Trainer
FLUX 2 Trainer

Fine-tune FLUX.2 [dev] from Black Forest Labs with custom datasets. Create specialized LoRA adaptations for specific styles and domains.

training
FLUX 2 Trainer V2
FLUX 2 Trainer V2

Fine-tune FLUX.2 [dev] from Black Forest Labs with custom datasets. Create specialized LoRA adaptations for specific styles and domains.

training
Qwen Image 2512 Trainer
Qwen Image 2512 Trainer

Qwen Image 2512 LoRA training

lorapersonalization
training
Z Image Turbo Trainer V2
Z Image Turbo Trainer V2

Fast LoRA trainer for Z-Image-Turbo, a super fast text-to-image model of 6B parameters developed by Tongyi-MAI.

lorapersonalization
training
FLUX 3 Klein Edit LoRA Trainer✦ new
FLUX 3 Klein Edit LoRA Trainer

Fine-tune FLUX 3 Image Klein on before/after image pairs to teach it a custom edit, with up to 10 input images per pair.

fluxlora
training
FLUX 2 [klein] 9b Base Trainer
FLUX 2 [klein] 9b Base Trainer

Fine-tune FLUX.2 [klein] 9B from Black Forest Labs with custom datasets. Create specialized LoRA adaptations for specific editing tasks.

training
Turbo Flux Trainer
Turbo Flux Trainer

A blazing fast FLUX dev LoRA trainer for subjects and styles.

training
Recraft V4 Styles Create Style
Recraft V4 Styles Create Style

Creates a reusable style from your reference images and returns a style ID you can pass to Recraft V4 Styles image and vector generation.

transformutility
training