Quick Training Guide

Training Entry

From the Create dropdown menu, select Model -> Training to enter the dataset screen.

Click Train Now to enter the training screen.

You can also click the training icon in the lower left corner of the creation flow to enter.

Basic Training Process

Image Material Selection

Upload image resolution: 256*256 ≤ image size ≤ 2048*2048

Supported image formats: .png .jpg .webp

Note: The quality of image materials determines the final effect of the model.

Number of training images: 10-30 images are recommended; the more images uploaded, the better the effect.

Selection Key Points

The main points for selecting images are displayed at the bottom of the screen.

Clear Character

This means the character's facial features are clear without obstruction, the scene does not cause visual errors for the character, and the character occupies the main part of the image.

Rich Angles

This refers to different views of the character, such as front, side, back, from below, and from above. It also includes various poses, such as standing, squatting, and lying down, as well as different expressions.

Rich Backdrop

This refers to images of the character in various scenes, including different decorations (clothing, hair accessories, expressions), different lighting, as well as full-body shots, half-body shots, actions in different scenes, close-ups of parts, such as facial features, hands, feet, etc. These images should account for about 30% of the total.

Varied Styles

When training realistic characters, you cannot select anime-style images or images in other styles, and vice versa.

Blurry Image

Image materials should be high-definition and of good quality.

Watermark

Images cannot have watermarks. Otherwise, the final model will likely also generate images with watermarks, affecting the model's effect.

Training Steps

Click the pencil to modify the training set's name.

Training Type: Currently, only character is available (other types will be enabled later).

Training Theme: Divided into Realistic Character and Anime Character (including furry).

Training Model: You can select various types of models.

Click Upload Images on the right to upload images from your local device. The box on the left is for uploading personal works (you can generate character images through the creation flow for uploading).

After uploading some images, you can also select Upload Image from above to upload more.

If there are upload errors, you can click Delete All on the left to delete images.

The Training Strength in the lower left is determined based on the number of images, and you can select the strength according to the recommendation.

The icon in the upper right is used to save the uploaded images.

Use the Training Object Name field to give your trained character a name, making it more convenient for you to generate this character after completing training.

Inconsistent upload image resolutions will not affect the model effect.

Click Train Now to start training.

View Training Test

You can quickly switch to the task list by clicking the icon next to the Train Now button in Quick Training.

In the dataset screen, click Task List to enter the training queue, then click View Training to check the training status.

Here you can check the training progress or stop the training.

Select models that perform well in sample images (usually the last 2-3 rounds of models) for saving.

Select and test them in the Model section of your personal center.

Select your own LoRA in the creation flow for image generation testing.