> 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/image-training.md).

# Image Training

Train an image LoRA to learn a character, an object, or a visual style from your images.

This guide takes you through choosing a base model, preparing and tagging your dataset, running training, and testing the saved model.

### Image Training Type Recommendation <a href="#image-training--image-training-type-recommendation" id="image-training--image-training-type-recommendation"></a>

Start with what you want the LoRA to learn:&#x20;

a realistic subject, an anime subject, or a visual style. The table below lists the recommended base models for each category.

<table><thead><tr><th width="170">Category</th><th>Base models</th></tr></thead><tbody><tr><td>Realistic</td><td>Krea 2 / Z-image-turbo / Klein-9b / Wan2.2 / Qwen 2511</td></tr><tr><td>Anime</td><td>Anima / Krea 2 / Illustrious / Pony / Z-image-turbo</td></tr><tr><td>Style</td><td>Krea 2 / Klein-9b / Wan2.2 / FLUX / Z-image-turbo</td></tr></tbody></table>

***

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

#### Images <a href="#image-training--images" id="image-training--images"></a>

**Quantity:** 20 to 40 images. More images improve generalization. Adding low-quality material instead lowers model quality.

> <mark style="color:$warning;">**📝**</mark><mark style="color:$warning;background-color:yellow;">**Image selection rules:**</mark>
>
> 1. **Low-quality image material**
>
> Too dark to identify the scene, unreasonable composition, unclear image, details hard to replicate, unclear primary/secondary subject, details hard to stack repeatedly, appearance of irrelevant elements, inconsistent subject with stitched images, incorrect or misaligned limbs.
>
> 2. **High-quality image material**
>
> Unified style, clear image, unified features, no watermark text, varied content, rich angles.
>
> 3. **Training base model selection:**&#x20;
>
> For style, character, and object, all need a base model that matches the style of the image material.

#### Trigger Words and Tagging <a href="#image-training--trigger-words-and-tagging" id="image-training--trigger-words-and-tagging"></a>

* **Trigger word:** a trigger word must be used to make the LoRA work at its best. It is a very important part of the prompt stage. (A uniquely identified word is required.)
* **Tagging:** use auto-tagging first, then screen out wrong tags that do not match the image.
* **For example:** if character training has fixed hair features, there is no need to tag them, so the features merge into the trigger word. For style training, tags containing style words need to be deleted.

#### Tagging Algorithms <a href="#image-training--tagging-algorithms" id="image-training--tagging-algorithms"></a>

<table><thead><tr><th width="122">Name (natural language)</th><th width="123">deepseek</th><th width="129">cogflorence</th><th width="116">qwen</th><th width="125">qwen vl 4b</th><th width="125">blip</th></tr></thead><tbody><tr><td>Training type</td><td colspan="5">flux, wan2.2, qwen, qwen2512, flux kre, flux2, 3.5m, 3.5l</td></tr><tr><td>Max images per tagging run</td><td>350</td><td>370</td><td>120</td><td>120</td><td>500</td></tr></tbody></table>

<table><thead><tr><th>Name (word)</th><th>deepbooru</th><th>wd 1.4 tagger</th><th>wd-swinv2-tagger-v3</th><th>wd-eva02-large-tagger-v3</th></tr></thead><tbody><tr><td>Training type</td><td colspan="4">sd1.5, xl, il, pony, anima</td></tr><tr><td>Max images per tagging run</td><td>500</td><td>500</td><td>500</td><td>500</td></tr></tbody></table>

* **Word tagging threshold:** the smaller the value, the finer the description.
* **Natural language recommendation:** qwen > qwen vl 4b > deepseek > cogflorence.
* **Word tagging recommendation:** wd-eva02-large-tagger-v3 > wd 1.4 tagger > wd-swinv2-tagger-v3.

#### Image Preprocessing <a href="#image-training--image-preprocessing" id="image-training--image-preprocessing"></a>

* All training can use no cropping (currently, all training is configured with an ARB bucket by default).

<table><thead><tr><th width="149"></th><th width="221">Focus crop</th><th>No crop</th></tr></thead><tbody><tr><td>Model type</td><td>sd1.5, xl</td><td>pony, il, flux, krea, wan2.2, z image, klein, qwen, qwen2512, anima</td></tr><tr><td>Resolution</td><td>512*512, 1024*1024</td><td>/</td></tr></tbody></table>

***

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

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FHiLekuJsGhFommlrwCgD%2F1.gif?alt=media&amp;token=d67e0ccd-6156-4fa7-ba20-26a814c360ad" alt=""><figcaption></figcaption></figure>

**Training flow:**&#x20;

Upload images - enter trigger word - finish tagging - adjust epochs - click train.

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FTwazbpzUfpFTBgTAdvXv%2F2.gif?alt=media&amp;token=eeb8f42d-7819-45c7-8ef3-299022e47478" alt=""><figcaption></figcaption></figure>

**Testing flow:**&#x20;

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

***

### Krea 2 Training <a href="#image-training--krea-2-training" id="image-training--krea-2-training"></a>

* **Tagging recommendation:** qwen (natural language tagging).

Both word tagging and natural language tagging can train good results (official recommendation is natural language tagging).

* **Training base model:** krea 2 raw.
* **Generation base model:** [Krea-2 - SeaArt AI Model](https://www.seaart.ai/models/detail/d8u8iate878c73dk82bg?v=b9364d8153cb921fac0e1cfec0600a80) (turbo\_bf16).

<table><thead><tr><th width="94"></th><th width="84">Tagging</th><th width="127">Image count</th><th width="113">Steps</th><th>Scheduler</th><th width="119">Scheduler</th><th>Dim/Alpha</th></tr></thead><tbody><tr><td>Realistic</td><td>qwen</td><td>more than 10</td><td>500~1000</td><td>cosine_with_restarts</td><td>AdamW8bit</td><td>default</td></tr><tr><td>Style</td><td>qwen</td><td>more than 40</td><td>500~1500</td><td>constant</td><td>Prodigy</td><td>default</td></tr><tr><td>Anime</td><td>qwen</td><td>more than 10</td><td>600~1500</td><td>constant</td><td>AdamW8bit</td><td>default</td></tr></tbody></table>

**Compare natural-language and word-based training**

The following images show the dataset and the results from the two tagging methods.

{% columns %}
{% column width="50%" %}

<p align="center"></p>

<p align="center"></p>

<p align="center"></p>

<p align="center">Dataset original images</p>
{% endcolumn %}

{% column width="50%" %}

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FKsxGVPHJPkGFePkxc0XX%2Fimage.png?alt=media&amp;token=53de817d-fed1-464b-b069-1f59d4e08ddd" alt="" width="188"><figcaption></figcaption></figure>
{% endcolumn %}
{% endcolumns %}

<table><thead><tr><th width="193">Name</th><th>Natural language</th><th>Word</th></tr></thead><tbody><tr><td>krea 2</td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FRWFEFD22tKC8vEzxWo9t%2Funknown.webp?alt=media&amp;token=4d658f81-5c17-4a41-88a3-79a5f9511b18" alt="" data-size="original"></td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2Frx1mwWXQYr2BF6CtBM5h%2Funknown.webp?alt=media&amp;token=33c21732-fd02-40f2-9222-4b7cbacd34a3" alt="" data-size="original"></td></tr></tbody></table>

> <mark style="color:$warning;background-color:yellow;">**Conclusion:**</mark>&#x20;
>
> the final effect and style performance show little difference. Whether training or generating images, natural language describes composition and element positions more accurately, and this is more obvious in complex environments. Natural language training is more recommended.
>
> Although krea 2 is a realistic model, its output quality in anime is very close to anima, so training anime styles or characters with krea 2 will also give good results.

***

### Anima Training <a href="#image-training--anima-training" id="image-training--anima-training"></a>

* **Tagging recommendation:** wd-eva02-large-tagger-v3 (word tagging), qwen (natural language tagging).

Both word tagging and natural language tagging can train good results (official recommendation is natural language tagging).

* **Training base model:** anima base.
* **Generation base model:** Anima BASE - SeaArt AI Model.

<table><thead><tr><th width="108"></th><th width="101">Tagging</th><th>Image count</th><th>Steps</th><th>Scheduler</th><th>Scheduler</th><th>Dim/Alpha</th></tr></thead><tbody><tr><td>Anime character</td><td>qwen</td><td>more than 10</td><td>1500~3000</td><td>cosine_with_restarts</td><td>AdamW8bit</td><td>default</td></tr><tr><td>Style</td><td>qwen</td><td>more than 40</td><td>2000~3000</td><td>constant</td><td>Prodigy</td><td>default</td></tr></tbody></table>

<mark style="background-color:violet;">Anima natural language training and word training effect comparison.</mark>

{% columns %}
{% column %}

<p align="center"></p>

<p align="center"></p>

<p align="center"></p>

<p align="center">Dataset original images</p>
{% endcolumn %}

{% column %}

<figure><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FKsxGVPHJPkGFePkxc0XX%2Fimage.png?alt=media&amp;token=53de817d-fed1-464b-b069-1f59d4e08ddd" alt="" width="188"><figcaption></figcaption></figure>
{% endcolumn %}
{% endcolumns %}

| Name | Natural language                                                                                                                                                                                                                                             | Word                                                                                                                                                                                                                                                         |
| ---- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| base | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FBLzY9OCVVSE66A0LUndh%2Funknown.webp?alt=media&amp;token=bbcbc983-0103-4c5a-b890-6b9445ea2c4f" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FFexXSIN5MhqdVFtZwFNP%2Funknown.webp?alt=media&amp;token=3adb99af-b812-49fc-aef9-198e1ae05c78" alt="" data-size="original"> |

> <mark style="color:$warning;background-color:yellow;">**Conclusion:**</mark>&#x20;
>
> the final effect and style performance show little difference, but natural language gives better background details.
>
> Natural language tagging is currently recommended for training. From the tagging perspective, natural language tagging is also more accurate.

<mark style="background-color:violet;">Comparison of different anima base model versions with the same dataset.</mark>

| Name             | Anima preview                                                                                                                                                                                                                                                | Anima preview2                                                                                                                                                                                                                                               | Anima preview3                                                                                                                                                                                                                                               | Anima base                                                                                                                                                                                                                                                   |
| ---------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| Natural language | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FfrgyqavjskjMwVLA1Veo%2Funknown.webp?alt=media&amp;token=d9d5d8f0-6433-4095-b06e-6da2e8c2ad24" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2Fq9oNBsiKYd6KittgeGqF%2Funknown.webp?alt=media&amp;token=c07bcf45-0f57-443d-8b15-062906275eaa" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FYPQQ5r5vdH5DBevt84Ua%2Funknown.webp?alt=media&amp;token=aff05b12-cdd7-48cf-9326-964b4bec5d6c" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2Fy2kYtT8u37nx9MEbF6Se%2Funknown.webp?alt=media&amp;token=bec4a177-eb40-4603-9cb3-f20ce0ca887b" alt="" data-size="original"> |

***

### Z Image Training <a href="#image-training--z-image-training" id="image-training--z-image-training"></a>

* Tagging recommendation: qwen (natural language tagging).
* Training base model: z image turbo, z image base.
* Generation base model: [Z Image Turbo - SeaArt AI Model](https://www.seaart.ai/models/detail/d4kssode878c7387fae0).

| Dataset original images | <img src="https://wdcdn.qpic.cn/MTY4ODg1NTg3MzY3ODk0MQ_967757_v_ULciQk05zhtndD_1786186583?w=1536&#x26;h=2048" alt="https://wdcdn.qpic.cn/MTY4ODg1NTg3MzY3ODk0MQ_967757_v_ULciQk05zhtndD_1786186583?w=1536&#x26;h=2048" data-size="original"> | <img src="https://wdcdn.qpic.cn/MTY4ODg1NTg3MzY3ODk0MQ_923378_6bl1fnbGVAdgp17H_1786186613?w=1568&#x26;h=2016" alt="https://wdcdn.qpic.cn/MTY4ODg1NTg3MzY3ODk0MQ_923378_6bl1fnbGVAdgp17H_1786186613?w=1568&#x26;h=2016" data-size="original"> |
| ----------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |

| Name               | turbo                                                                                                                                                                                                                                                        | base                                                                                                                                                                                                                                                         | turbo                                                                                                                                                                                                                                                        | base                                                                                                                                                                                                                                                         |
| ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| turbo (base model) | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FQBuNimogCidEkvZqCn9P%2Funknown.webp?alt=media&amp;token=80aa13a1-6d27-4ff7-9fc1-f2655db3d794" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FYRAfA6tpxbuUVWtiHR0d%2Funknown.webp?alt=media&amp;token=63023a28-4d46-4f77-bd6d-c8b02e1b8098" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FRRO5mJP4cP6OlzwLdZWJ%2Funknown.webp?alt=media&amp;token=6cd3370a-e643-49aa-b43b-def07df3bebf" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FHIGPlz6BIrhswgoHogRG%2Funknown.webp?alt=media&amp;token=ccef7dfa-8797-478a-856f-e5f8089a113f" alt="" data-size="original"> |
| base (base model)  | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F8zclLmhPl1aX4wHkaTn5%2Funknown.webp?alt=media&amp;token=3e5c370d-78c5-41d3-ba09-5228bc7377a6" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FhpFQCQghTXPddF70Cp1W%2Funknown.webp?alt=media&amp;token=760a2b5e-8ec9-455a-85ba-75ebef10e944" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FA0Pa9faFbV7t37oWqs6p%2Funknown.webp?alt=media&amp;token=bb694329-0727-4cf4-92e3-2d77e5e022eb" alt="" data-size="original"> | <img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FB4LBs24rF1jYwzD5fHSR%2Funknown.webp?alt=media&amp;token=e134982a-cda3-4fad-ada3-9a98ae382f5f" alt="" data-size="original"> |

> <mark style="color:$warning;background-color:yellow;">**Conclusion:**</mark>&#x20;
>
> when training style, some datasets can produce effects normally, but some datasets cannot learn the style normally.> \
> For style training, z image is not recommended.

<table data-header-hidden><thead><tr><th></th><th width="117"></th><th></th><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td><br>Tagging</td><td>Image count</td><td>Steps</td><td>Scheduler</td><td>Scheduler</td><td>Dim/Alpha</td><td></td></tr><tr><td>Realistic character (Z-turbo)</td><td>qwen</td><td>more than 20</td><td>1500~3000</td><td>default</td><td>default</td><td>default</td></tr><tr><td>Style (Z-turbo)</td><td>qwen</td><td>more than 40</td><td>2000~6000</td><td>default</td><td>default</td><td>default</td></tr><tr><td>Anime character (Z-turbo)</td><td>qwen</td><td>more than 20</td><td>1500~3000</td><td>default</td><td>default</td><td>default</td></tr><tr><td>Realistic character (Z-base)</td><td>qwen</td><td>more than 20</td><td>3000-4000</td><td>default</td><td>default</td><td>default</td></tr><tr><td>Style (Z-base)</td><td>qwen</td><td>more than 40</td><td>4000-6000</td><td>default</td><td>default</td><td>default</td></tr></tbody></table>

***

### Wan2.2 Text-to-Image Training <a href="#image-training--wan2-2-text-to-image-training" id="image-training--wan2-2-text-to-image-training"></a>

* Tagging recommendation: qwen (natural language tagging), deepseek (natural language tagging).
* Training base model: wan2.2 t2v-low.
* Generation base model: SeaArt Infinity V2.0 - SeaArt AI Model.

|                 | Tagging | Image count  | Steps      | Scheduler              | Scheduler | Dim/Alpha |
| --------------- | ------- | ------------ | ---------- | ---------------------- | --------- | --------- |
| Anime character | qwen    | more than 10 | 1500\~3000 | cosine\_with\_restarts | AdamW8bit | default   |
| Style           | qwen    | more than 40 | 2000\~3000 | constant               | Prodigy   | default   |

<mark style="background-color:violet;">The table shows the impact of high/low noise LoRA on the final generated image.</mark>

<table><thead><tr><th width="115">Name</th><th>Low noise</th><th>High noise</th><th>High and low noise</th><th>Original image (no LoRA)</th></tr></thead><tbody><tr><td>Character</td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FM4O6cjAPLpILHZsgAMrJ%2Funknown.webp?alt=media&amp;token=37dd701b-d7f2-4d4a-b9ca-8d05c32e659d" alt="" data-size="original"></td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F1zZ2sAg4NUYXN5xhWZeq%2Funknown.webp?alt=media&amp;token=c891d826-7651-4b21-83a8-e45fee26c609" alt="" data-size="original"></td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2Fo6b8UiVvZmFxnfPGxRze%2Funknown.webp?alt=media&amp;token=ee1b6862-7c64-4643-9793-c1bd7fce1128" alt="" data-size="original"></td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FATyqnwznKwvWzCvjrJsN%2Funknown.webp?alt=media&amp;token=ac244415-7989-439e-bff7-ebd333cf6a10" alt="" data-size="original"></td></tr><tr><td>Style</td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FIxMqraIy3rInP7lutD7E%2Funknown.webp?alt=media&amp;token=804aa799-99f9-4549-befa-fe9d724196c6" alt="" data-size="original"></td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FmBJZx4ZDh9TVUfAcM5hr%2Funknown.webp?alt=media&amp;token=5d5b174c-25ba-4886-8341-aa600e784a5a" alt="" data-size="original"></td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2FIOUAxoUZ8kclVbRCgKSw%2Funknown.webp?alt=media&amp;token=7c21b391-c3f9-42ad-b55e-88cc022e41a3" alt="" data-size="original"></td><td><img src="https://2219884424-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FigAtVLBrlaI8jVruJfC8%2Fuploads%2F9ipU97q38IEgOLrQid9Y%2Funknown.webp?alt=media&amp;token=7ad6bd2f-9ec4-42a3-83f2-2f7266158edd" alt="" data-size="original"></td></tr></tbody></table>

> <mark style="color:$warning;background-color:yellow;">**Conclusion:**</mark>&#x20;
>
> It can be seen from the table that whether training characters or styles, a low-noise model must be trained.
>
> The factors affected by the low-noise model are very important. The low-noise model affects the character's facial features and the expression of the style.
>
> The high-noise model affects composition and lighting/shadow.
>
> There are only two training combinations: low noise, and low noise + high noise.

***

### Qwen-Image, Qwen-2511 Training <a href="#image-training--qwen-image-qwen-2511-training" id="image-training--qwen-image-qwen-2511-training"></a>

* Trigger word: for Qwen tagging, do not use meaningless words. Use common descriptive words, such as "A photo of a person named \<trigger word>".
* Tagging recommendation: qwen (natural language tagging), qwen vl 4b (natural language tagging), deepseek (natural language tagging).
