> For the complete documentation index, see [llms.txt](https://docs.seaart.ai/guide-1/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.seaart.ai/guide-1/advanced-guide/lora-training/edit-model-training.md).

# Edit Model Training

Train an editing LoRA with examples of an image before and after an edit.

This guide explains how to organize those examples, describe the change you want the model to learn, set up training, and test the saved LoRA.

### Edit Training Type Recommendation <a href="#edit-model-training--edit-training-type-recommendation" id="edit-model-training--edit-training-type-recommendation"></a>

<table><thead><tr><th width="212">Category</th><th>Models</th></tr></thead><tbody><tr><td>Single-image edit</td><td>qwen edit 2511 / qwen edit 2509 / qwen edit / kontext</td></tr><tr><td>Multi-image edit</td><td>qwen edit 2511 / qwen edit 2509 / qwen edit / kontext</td></tr></tbody></table>

***

### Differences Between Edit Model Training and Conventional LoRA Training <a href="#edit-model-training--differences-between-edit-model-training-and-conventional-lora-training" id="edit-model-training--differences-between-edit-model-training-and-conventional-lora-training"></a>

<table><thead><tr><th width="180">Item comparison</th><th>Training image count</th><th>Applicable scenarios</th></tr></thead><tbody><tr><td>Lora</td><td>Generally requires 50-100 images or more</td><td>Style transfer, character/object feature learning</td></tr><tr><td>Edit Lora</td><td>Usually 10-50 sets is enough for good results</td><td>Creative scenarios requiring precise semantic control, image modification, partial transfer</td></tr></tbody></table>

***

### Dataset Preparation <a href="#edit-model-training--dataset-preparation" id="edit-model-training--dataset-preparation"></a>

#### Image Screening <a href="#edit-model-training--image-screening" id="edit-model-training--image-screening"></a>

* **The images in the training dataset must be grouped as original image and result image:**&#x20;

Each group contains the original image, the result image, and the annotation.

* **Number of image sets:** 10-50 sets recommended, maximum upload limit is 100 sets.
* **Image selection:** multiple types of original images; images with watermarks need watermark removal.
* **For example:** to convert into ink style, original images can cover realistic style, line art style, people, objects, and landscapes, to increase generalization.
* **Dataset image range:**

| Name                 | kontext | qwen edit | qwen edit 2509 | qwen edit 2511 |
| -------------------- | ------- | --------- | -------------- | -------------- |
| Original image count | 1       | 1         | 1\~3           | 1\~3           |
| Result image count   | 1       | 1         | 1              | 1              |

#### Image Annotation <a href="#edit-model-training--image-annotation" id="edit-model-training--image-annotation"></a>

* **Annotation language:** English annotation is recommended to avoid translation errors.
* **Default annotation:** Can be filled in when there is no annotation. If all groups are annotated, it does not need to be filled in.
* **Annotation content:** Correct annotation: describes the difference between the original image and the result image, only a simple description is needed.

**✅️Correct examples:** Convert the image into ink style; convert into a fisheye lens; turn the character into a cat.

**❌️Wrong annotation:** Like ordinary image model training annotation that describes the image content.

**❌️Wrong example: A**slim young Asian woman with long black hair, wearing a pink tank top and blue denim shorts, standing on a green railing by the river, with a bridge in the background.

#### Upload Dataset <a href="#edit-model-training--upload-dataset" id="edit-model-training--upload-dataset"></a>

* <mark style="background-color:yellow;">Click to upload images, a maximum of 50 images can be uploaded at once; the part beyond 50 images will not be uploaded successfully.</mark>
* <mark style="background-color:yellow;">Select all images to upload. If there is no fixed naming format, they will be uniformly placed into the unmatched images directly below the matching group. Unmatched images can accumulate up to 50, and you need to manually move the images into the corresponding group.</mark>
* Finish manual annotation for each group of images. Or, under the default annotation, enter a unified annotation prompt to complete all annotations.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FcRLPteMKtZp5q9qSveMl%2F%E4%BC%81%E4%B8%9A%E5%BE%AE%E4%BF%A1%E6%88%AA%E5%9B%BE_20261027143007.png?alt=media&amp;token=1d1195ca-97da-444a-8f98-e5fda12f43b4" alt=""><figcaption></figcaption></figure>

***

### Parameter Settings <a href="#edit-model-training--parameter-settings" id="edit-model-training--parameter-settings"></a>

#### Training Steps <a href="#edit-model-training--training-steps" id="edit-model-training--training-steps"></a>

* Training steps optimization: the model needs to show its proper effect on every image and maintain better generalization. This requires more dataset images and more steps.
* For only a small, specific use, choose fewer dataset images and reduce the steps.
* Recommended steps: default is 1500 steps. Within a 20-50 set dataset, 2000-3000 steps is recommended.

#### Training Epochs <a href="#edit-model-training--training-epochs" id="edit-model-training--training-epochs"></a>

* Times per Image Repeat: actually calculated by the number of groups, the number of training times per group.
* Cycles Epoch: when all groups have completed training according to the number of Times per Image Repeat, Cycles Epoch is 1. It is consistent with the Epoch and Repeat concepts in image training.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FDffYQYlSwQe6p3JLlX4f%2F7d08ea02-7c1e-480b-a518-b55ba6da10f7.png?alt=media&amp;token=6b7a6bc8-9977-45bc-a170-fe27196a90c1" alt=""><figcaption></figcaption></figure>

* In most cases, use the default parameters for training. Generally, when there are more groups, you can choose fewer Times per Image Repeat and Cycles Epoch.

#### Learning Rate <a href="#edit-model-training--learning-rate" id="edit-model-training--learning-rate"></a>

* You can directly use the default learning rate, no need to modify.

#### Default Annotation <a href="#edit-model-training--default-annotation" id="edit-model-training--default-annotation"></a>

* Default annotation means that when all image groups or individual image groups have no annotation, this default annotation can be used to fill in, to avoid missing annotations.

#### Model Effect Preview Original Image <a href="#edit-model-training--model-effect-preview-original-image" id="edit-model-training--model-effect-preview-original-image"></a>

* The preview image uses the initial image (upload as many initial images as you have).

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FYVGT47kb7e4AHhUoCWKa%2Fd84c9b34-2753-41b5-9ddf-dc64c27f18e0.png?alt=media&amp;token=c82b2e4b-d106-45bc-8363-56b6bbaf9c93" alt=""><figcaption></figcaption></figure>

#### Model Effect Preview Annotation <a href="#edit-model-training--model-effect-preview-annotation" id="edit-model-training--model-effect-preview-annotation"></a>

* The preview annotation can be the same as the dataset annotation.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FT8dTdQU3GnyAn4Uqz0oc%2F140aece3-1965-414c-81de-93536a0a0bc5.png?alt=media&amp;token=db12f734-11ff-4440-9f0d-b73c28947bda" alt=""><figcaption></figcaption></figure>

***

### General Training Workflow <a href="#edit-model-training--general-training-workflow" id="edit-model-training--general-training-workflow"></a>

* The only difference between multi-image training and single-image training is the number of initial images uploaded. Everything else is the same.

#### Single-Image Training <a href="#edit-model-training--single-image-training" id="edit-model-training--single-image-training"></a>

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F4cSlavNR5AD4DijL0DOF%2F%E9%80%9A%E7%94%A8.gif?alt=media&amp;token=73f67b53-a4d0-4d34-aa6a-e9a098438331" alt=""><figcaption></figcaption></figure>

**Training flow:**&#x20;

Upload images - match the initial image with the result image - finish annotation - fill in the default annotation - upload a sample image - fill in the sample image prompt - adjust epochs - click Train.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F3pZnN2DCJA7GOA0Y6tjI%2F1c39a9c1-3744-4327-8ea2-4e033be97c4a.gif?alt=media&amp;token=ed311498-449f-42ec-94d2-bb4f58ed599f" alt=""><figcaption></figcaption></figure>

**Testing flow:**&#x20;

Training progress page - save model - go to personal profile - model - model detail page - generate - workflow - upload a test image - enter trigger word and prompt - click generate.

#### Multi-Image <a href="#edit-model-training--multi-image" id="edit-model-training--multi-image"></a>

Multi-image training.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FYPw5NT7nSkFtv8hw1ify%2F%E5%A4%9A%E5%9B%BE%E8%AE%AD%E7%BB%83.gif?alt=media&amp;token=cba0b178-fca2-4676-abe6-e49bb7a04e12" alt=""><figcaption></figcaption></figure>

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FcQBdXQavzq37ZWMGCspK%2F%E5%A4%9A%E5%9B%BE%E8%AE%AD%E7%BB%832.gif?alt=media&amp;token=74322456-c761-4464-aec4-e8ede78493ef" alt=""><figcaption></figcaption></figure>
