SeaArt LoRA Training — Practical Notes: Getting It to Generate as Intended

Overview

In the previous part I covered the basics of training a LoRA with SeaArt. This post is a practical record of actually creating one.

Long story short, the overall results were so-so. Here are the causes and what I learned.

---

Preparation

Preparing the training data

• Tool: Antigravity
• Processing:
  • Auto crop & resize into face close-up, bust-up, and full body
  • Started from 1024-square sources → cropped so subjects fit within a 512×512 rectangle
  • Auto-generate captions (tags) at the same time
  • Then manually revise

File layout

dataset/
├── image001.png  (512x512)
├── image001.txt  (caption)
├── image002.png
├── image002.txt
└── ...

Sample training data

Examples of the data I actually used.

Caption examples

The trigger word is kanachan (a coined token). Captions were auto-generated by Claude, then lightly edited by hand.

Full body:

kanachan, 1girl, hoodie, t-shirt, sweatpants, v sign, smile, looking at viewer, standing, monochrome, greyscale, manga, cowboy shot, white background

Upper body:

kanachan, 1girl, hoodie, t-shirt, v sign, smile, looking at viewer, monochrome, greyscale, manga, upper body, white background

Face close-up:

kanachan, 1girl, hoodie, smile, looking at viewer, monochrome, greyscale, manga, portrait, close up, white background

Final training dataset

• 104 images
• Monochrome materials
• Variations: face close-up, bust-up, full body, various situations

---

Plan selection

Plans considered

Initial choice: Standard

• LoRA training entitlement (fast when the queue is empty)
• With 700 stamina/day, training 104 images looked feasible
• Can train one LoRA task at a time

---

Attempt 1 — Failed with Illustrious

Settings

Item	Value
Base model	Illustrious v0.1
Trigger word	kanachan
Repeat	4
Epoch	10
Total steps	4,160

FLUX was initially selected, but that was tough on an RTX 4060 8 GB, so I switched to Illustrious v0.1. v1.0 exists, but v0.1 had more reports of being stable, so I chose that.

Preview prompt

For checking during training:

kanachan, 1girl, cardigan, shirt, necktie, smile, upper body, monochrome, greyscale

Tuning stamina consumption

• Repeat 5, Epoch 10: estimated 786.47 → with stamina 758 it didn’t fit
• Repeat 4, Epoch 10: estimated ~600 → fit within the daily stamina

Result: failed

• The side-ponytail hairstyle didn’t appear at all
• LoRA 1–5: facial features were far from the source
• LoRA 7–8: it finally started tying the hair (too late)
• LoRA 9–10: a bit better but still insufficient
• Despite specifying monochrome, outputs looked slightly colorized/greyish
• The body build didn’t resemble the source

Why it failed

1. Repeat 4 was too shallow — 104 images × Repeat 4 = only ~40 passes per image
2. Illustrious v0.1 pairs poorly with monochrome manga
3. Actual stamina usage exceeded the estimate (estimated in the 600s → actually ~700)

---

Understanding loss and Epoch/Repeat

How to read the loss

loss value	State
0.2–0.3 or higher	Barely learning yet
Around 0.1	Mid-training
0.05 or lower	Well learned (too low risks overfitting)

During the first run, loss went from 0.106 → 0.112 at the halfway point, meaning it ended while still mid-learning.

What Epoch and Repeat mean

• Epoch: Snapshot count at which LoRA files are saved. If Epoch is 10, you get LoRA 1–10.
• Repeat: How many times each image is repeated within an epoch.

Typical pattern:

• LoRA 1–3 → too shallow, weak characteristics
• LoRA 5–7 → just right
• LoRA 8–10 → likely to overfit

---

Upgrading to Pro

Problems

• Standard plan’s 700 stamina/day wasn’t enough
• Increasing Repeat drives stamina usage up further
• Tough to iterate many times
• Training priority is available (not sure how different it is)

Decision: Pro plan (¥4,300/month)

Cheaper than buying a PC. An RTX 4090 costs ¥250–300k, needs 24 GB VRAM, and adds power costs. Compared to that, paying ¥4,300 to iterate many times is cost‑effective.

---

Attempt 2 — Improved with Anything

Settings

Item	Value
Base model	Anything (万象熔炉), SD1.5 family
Repeat	10 (up from 4)
Epoch	15
Total steps	15,600 (about 4× the first run)
Estimated stamina	603

Why choose SD1.5

• Works well with monochrome manga
• Lighter to train
• Tends to capture characteristics easily
• Easy local generation on an RTX 4060

Result: much better

• The side ponytail shows up from LoRA 1
• With Illustrious it only started around LoRA 9–10, but Anything captured it from the start
• Effect of raising Repeat to 10 + better pairing with SD1.5

Generation tests

Prompt:

kanachan, かなちゃん, 1girl, カーディガン, シャツ, ネクタイ, スマイル, アッパーボディ, モノクローム, グレースケール

Remaining issues

• Tends to get too dark toward the later LoRAs (13–15)
• Clothing geometry is a bit off (e.g., boundary between cardigan and shirt)

---

Attempt 3 — Tweaked

Settings

Item	Value
Base model	Anything (万象熔炉)
Repeat	10
Epoch	12 (reduced from 15 to avoid overfitting)
Estimated stamina	492.48

Result: stable

• Outputs are stable overall
• LoRA 9–10 feel well-balanced
• The side ponytail appears correctly
• Facial features are consistent
• LoRA 11–12 look slightly overfit (tends to lock the outfit into a hoodie)

---

Online generation tests

Upper body: good

Prompt:

kanachan, 1girl, cardigan, shirt, necktie, smile, upper body, monochrome, greyscale, white background, simple background

Negative:

dark, high contrast, black background, shadow, colorful

Result:

• Side ponytail ✓
• Ahoge ✓
• Cardigan + shirt + necktie ✓
• White background ✓
• Good brightness

One concern: the face looks a bit different from the source (eye shape, outline, etc.).

Full body: not good

Tested prompt (running late scene with toast in mouth):

kanachan, 1girl, cardigan, shirt, necktie, running, bread in mouth, toast, late for school, street, motion blur, monochrome, greyscale, full body

Result: Turned into a completely different character

• Neither the face nor hairstyle look like kanachan
• The LoRA effect weakens at full body
• Because the dataset had more bust-up images, full-body outputs dilute the features

Maid outfit: good

Prompt:

kanachan, 1girl, maid, maid headdress, maid apron, smile, upper body, monochrome, greyscale, white background, simple background

Negative:

dark, high contrast, shadow, colorful

---

Troubleshooting

Upgrade-ad modal problem

When I clicked “Train now”, an upgrade-to-Pro modal appeared and I couldn’t proceed.

Tried:

• Clicking the × in the top-right → no effect
• Clicking the dark backdrop → no effect
• Esc key → no effect

Solution: Reload the browser

Side effect: After reloading, as many training tasks were registered as the number of clicks I’d made (five). I deleted the extras from the training history page.

LoRA selection mistake

During generation I got an error, “The LoRA and base model don’t match.” I meant to use the LoRA trained on Anything, but I had mistakenly selected an older one trained on Illustrious.

Although the UI looked correctly updated, internally it behaved as if the previous settings were still in effect.

Lessons:

• Give LoRAs clear names (e.g., kanachan_anything_mono)
• LoRA trained on Anything (SD1.5) → use with SD1.5 models
• LoRA trained on Illustrious (SDXL) → use with SDXL models
• Don’t trust the UI state or carry over previous settings; starting fresh reduces mistakes

---

Wrap-up

Iteration summary

Conclusion: overall so-so

The cause seems to be resolution, not the number of images.

• Recommended 512-square, but the dataset actually mixed in 1024/2048-square images
• I just threw the data at SeaArt and let it ingest them → I should have unified the resolution up front
• I also left captions (tags) entirely to the AI and didn’t curate them enough
• Curation gets tough as the numbers grow
• Since generated images do capture the characteristics, the issue doesn’t seem to be the count

Next

I’ll try again when I have time, but as it stands I don’t have the energy to produce a color version.

---

Bonus

Edited with Gemini 3.0 Pro on Antigravity.

Getting it to this point takes quite a bit of work…