Translated from the original Japanese article
Tech 5 min read
SeaArt LoRA Training Guide
Overview
I was going to set up a local LoRA training environment, but someone said, “Why not just use the cloud?”
It turns out you can upload a training dataset and create a LoRA on SeaArt (I happened to be subscribed). There’s no environment setup, and if you download a safetensors file you can generate locally.
Benefits
- No environment setup (training runs in the cloud)
- Train without owning a GPU
- Download as safetensors → usable locally
- Supports many base models: FLUX / SDXL / Pony / Illustrious
Pricing (as of December 2025)
| Plan | Stamina/day | Notes |
|---|---|---|
| Free | 0 | LoRA must be public |
| Beginner SVIP | 300 | Private LoRA allowed, 3‑day free trial |
| Standard SVIP | 700 | - |
| Professional SVIP | 2,100 | - |
- Training is available on all plans (consumes stamina)
- The free plan requires your LoRA to be public
- Choose a paid plan if you want to keep the LoRA private
Preparation
Training data
- Images: 200 (100 color / 100 monochrome)
- Size: 512x512 (auto crop & resize by the tool)
- Variation: face close‑up, bust‑up, full body
- Captions: auto‑generated and then manually corrected
File layout
dataset/
├── image001.png (512x512)
├── image001.txt (キャプション)
├── image002.png
├── image002.txt
└── ...SeaArt LoRA Training Steps
1. Create an account & log in
You can sign in easily with a social account.
2. Go to the training page
“Create → Training → Create dataset”
3. Choose a base model
Pick a model that matches what you’ll run locally (RTX 4060 8GB in this example). With 8 GB of VRAM, SDXL‑class models run.
| Model | Base | Characteristics | 4060 support |
|---|---|---|---|
| Illustrious-XL | SDXL | Illustration‑focused | Yes |
| AuthakuMix | Pony | Anime‑focused | Yes |
| NoobAI-XL | SDXL | General purpose | Yes |
| FLUX | FLUX | Latest, high quality | No (tough) |
Note: SDXL LoRAs can only be used with SDXL models.
4. Upload the dataset
“Use dataset upload if you already have captioned images.”
- Upload image and txt files in pairs
- If filenames match, they will link automatically
- Skip the tagging algorithm (captions already exist)
5. Check captions
After upload, confirm the following:
✓ トリガーワード（造語）が先頭にある
✓ 英語タグ、カンマ区切り
✓ キャラ固有の特徴（髪色、目色など）は削除済み
✓ 構図タグが正しく入っているExamples of composition tags
| Crop | Tags |
|---|---|
| Face close‑up | portrait, close up, face |
| Bust‑up | upper body |
| Full body | full body |
6. Set training parameters
Recommended settings for 200 images
| Item | Recommended value |
|---|---|
| Repeats | 5 (lower since image count is large) |
| Epochs | 10 |
| Batch size | 2 |
| Learning rate | 0.0001 (1e‑4) |
| Network size (Dim) | 64–128 |
| Network Alpha | Half of Dim (32–64) |
| Scheduler | cosine_with_restarts |
| Optimizer | AdamW or DAdaptation |
Step count guide
100枚: リピート10 × エポック10 = 10,000ステップ
200枚: リピート5 × エポック10 = 10,000ステップ7. Preview prompt settings
For generating sample images after training:
my_chara, 1girl, upper body, smile, simple background8. Start training
“Enter a dataset name → Start training now”
- There may be a queue (2–4 hours when crowded)
- Training continues even if you close the browser
- Paid plans shorten wait time via priority
9. Training complete → Download
“Training history → Confirm the task is complete”
- Sample images are shown per epoch
- Choose the epoch with the best results
- “Download” → Save safetensors locally
- “Save” → Save on SeaArt (can set private)
Separate color/monochrome LoRA
If you create both
| LoRA | Trigger word | Use case |
|---|---|---|
| Color version | my_chara_color | Cover, color pages |
| Monochrome version | my_chara_mono | Main manga |
Keeping separate trigger words makes it easy to switch between them.
Prompt for monochrome generation
プロンプト:
my_chara_mono, 1girl, upper body, smile, monochrome, greyscale, manga, lineart
ネガティブ:
color, colorfulGenerating in monochrome from the start looks the cleanest. Converting color → grayscale degrades quality.
Local usage
Where to place the downloaded safetensors
ComfyUI:
ComfyUI/models/loras/my_chara.safetensors
A1111:
stable-diffusion-webui/models/Lora/my_chara.safetensorsGeneration on an RTX 4060 8GB
- SDXL‑class models → work
- LoRA + ControlNet together → tight on VRAM but works
- ComfyUI tends to be more VRAM‑efficient
Specify composition with ControlNet
Workflow
1. Clip Studioで3Dデッサン人形ポーズ作成
2. PNG出力（512x512、白背景）
3. ControlNet（Depth推奨）に入力
4. LoRA適用して生成ControlNet types
| Type | Compatibility with Clip Studio 3D | Use |
|---|---|---|
| Depth | Excellent | Preserves 3D sense |
| OpenPose | Good | Extracts pose only |
| Canny | Fair | Contours can sometimes get in the way |
Depth is the most stable when starting from 3D.
Security / Visibility
Differences by plan
| Plan | LoRA visibility | Notes |
|---|---|---|
| Free | Required | Cannot use on SeaArt unless public |
| Paid | Optional | Can keep private and still use |
To use safely
- Use a paid plan and set the model to private → only you can use it on SeaArt
- Download as safetensors → stored fully locally; others can’t use it
- Don’t post to a model marketplace
Even on the free plan, if you download the safetensors and use it locally, others won’t be able to use it.
Advanced settings
Learning rate & optimizer
| Optimizer | Characteristics | Learning rate |
|---|---|---|
| AdamW | Standard, stable | 0.0001 |
| AdamW8bit | Lightweight | 0.0001 |
| DAdaptation | Auto‑tuning | 1 |
| DAdaptAdam | Auto‑tuning | 1 |
| Prodigy | Auto‑tuning | 1 (auto‑set) |
Network settings
| Item | Description | Recommended |
|---|---|---|
| Dim | Higher reflects finer features | 64–128 |
| Alpha | Lower applies LoRA more strongly | Half of Dim |
Scheduler
| Type | Characteristics |
|---|---|
| cosine_with_restarts | Stable; supports multiple styles (recommended) |
| cosine | Smooth decay |
| constant_with_warmup | Stabilizes the start |
Troubleshooting
Features don’t appear
- Increase total training steps (more epochs)
- Increase Dim (64 → 128)
- Make sure the trigger word is in the prompt
Too strong / collapses
- Lower repeats
- Raise batch size
- Lower epochs
Stuck on a specific pose
- Not enough variation in the training data
- Check that composition tags (e.g.,
upper body) are included correctly
Wrap‑up
Recommended flow
- Subscribe to a SeaArt paid plan for one month
- Train a color LoRA → download safetensors
- Train a monochrome LoRA → download safetensors
- Generate locally with an RTX 4060 + ComfyUI
- Use ControlNet (Depth) + Clip Studio 3D to specify composition
- Mass‑produce manga panels
Key points
- Base models: SDXL‑class (Pony / Illustrious)
- For 200 images: repeats 5 × epochs 10
- Use different trigger words for color / monochrome
- Generate in monochrome from the start (post conversion degrades quality)
- Keep the LoRA private so others can’t use it
In the next article, I’ll actually upload the dataset, create a LoRA, and walk through tuning while reviewing the outputs.