Skip to content

Datasets#

A dataset is a set of images, each with a caption, plus a trigger word that names the subject. Good datasets are small and clean: for a character, 10 to 25 sharp frames with a clearly visible face beat a hundred mixed ones.

The project list

From a character#

The fastest path. On a character card, Train character LoRA creates the project and a first dataset from the base reference and the angles. In the dataset dialog you can:

  • pick which angles to include and whether to add the base image;
  • add extra frames from the gallery (well-lit stills of the same person); photos you uploaded to the Assets tab are listed next to generated images, and the board menu counts both;
  • choose the crop mode: auto centres a square crop on the detected face at the training resolution; wide takes the largest square with the face in its upper third, so medium and full-body photos keep their framing at native resolution; none takes a centre crop without detection; full frame (Krea 2 only) keeps the whole photo in its aspect ratio, so a tall full-body shot keeps its head and feet (the frame area equals the resolution squared: at 1024 the training takes about twice as long as at 768);
  • set the resolution (768 square is the default for video LoRAs), the trigger word, the description and the caption style.

The build report lists every frame with its crop and warnings: no face found, several faces, face too small (upscaled), face too small for identity (identity_weak: the frame teaches framing, not the face) and, for the whole dataset, framing_uniform when nearly every frame is a close-up: add medium and wide shots, or the LoRA learns the framing along with the face. Images added later report their own batch and the dataset total; the framing check and the recommended frame count always look at the whole dataset. Fix or drop flagged frames before training.

Create a dataset in the project and add images from the gallery. The picker shows generated images and uploaded photos alike, with a board filter; the default view spans every board. The same crop and caption path applies. Use this for styles, objects and for image LoRAs.

Captions and the trigger word#

Captions are generated from a template: a still-frame note, the trigger word, the description, and a style line. Edit any caption in the sample grid. The trigger word (chr_<name> for characters) appears once per caption. For video LoRAs it is what you write in prompts to call the subject; Krea 2 and FLUX.2 image LoRAs are prompted without it, because those models print unknown words as text (see Train an image LoRA).

Keep captions honest. Describe what is in the frame, not what you wish the model would learn; the LoRA learns the difference between caption and image.

Caption detail. By default a character caption carries only the class phrase of the description (Short; Krea 2 projects preselect the per-frame VL caption below): A still frame of chr_rbv2, a woman, in a close-up portrait…. A full physical description in every caption lets the model explain the frames with words instead of binding them to the trigger word, and that identity then does not carry without a reference image. Choose Full only when the words are the point (a style, an object).

Per-frame VL caption (recommended for photo sets). Choose Per-frame VL caption and Maystro writes each crop's caption with the Krea 2 Qwen3-VL encoder (already installed with Krea 2, nothing to download): framing, head pose, expression, clothing, background, lighting, camera and medium (phone selfie), image quality (slightly blurry) and people or objects at the edges, for example chr_rbv2, a woman, close-up selfie, slight tilt, smiling, yellow sweater, string lights, indoor warm light, phone selfie, sharp, partial figure on left edge. It never names the face, age, hair color or ethnicity, so only the identity binds to the trigger word; everything the caption names stays controllable from the prompt. Use it whenever the photos share a look you do not want in the results (phone selfies with the same warm indoor light): with short captions that look binds to the trigger and leaks into every image. Captioning takes about a second per frame; the captions are shown on the samples and can be edited. Frames the captioner could not describe fall back to the frame note and are listed in the report as caption_failed.

Style tail. If the photos share a look (phone selfies, one studio light), name it in the Style field, for example in a casual phone snapshot. It is appended to every caption, so the look binds to that phrase instead of the trigger; leave the phrase out of your prompts and the character renders in the scene's own look.