Summary Flow Images Multiple characters Fixed settings Basic Settings Descriptions Parameters kohya mapping Credits & time Checklist Publishing FAQ
Guide · Advanced Training
PixAI LoRA Advanced Training Guide
Steps in the flow
3
Default Training Steps
325
Images
10–100
Training time
2–6 h
Advanced Training gives you direct control over how PixAI trains your LoRA. You can review and edit the description for every image, and set Training Steps, learning rate, detail capacity (Rank), Alpha and batch size yourself. Every default matches Basic Training, so you only change what you want to test.
Keep Advanced Training open in another tab and follow along with this guide.
Already training LoRAs with kohya or sd-scripts? Three things work differently here
StepsTraining length is set as total steps. There are no epochs or repeats.
AlphaAlpha can only be half, equal to, or double your Rank.
CaptionPixAI writes the image descriptions first. You edit them afterwards.
Quick summary
Steps
Training length is total steps
Training Steps range from 50 to 800, default 325. No epochs or repeats.
Learning rate
One learning rate
Default 6e-4, applied to the whole network. You can't set it per module.
Alpha
Three Alpha options
Half, equal to, or double your Rank, with half as the default. Credit cost depends mainly on Training Steps and batch size. Learning rate, Rank and Alpha don't change it.
Caption
Descriptions are generated first
Written in natural language, not Danbooru tags. You can edit each one afterwards.
Trigger words
Added to every description automatically
Enter them once in Basic Settings and PixAI puts them at the start of each image's description.
Credits
Descriptions and training are billed separately
Described images aren't charged again, a failed training isn't charged, and deleting described images doesn't refund their description cost.
How the training flow works
Advanced Training has three steps: Basic Settings, Descriptions and Advanced Parameters. The order is close to most community LoRA workflows. The main difference is that PixAI writes the captions, and some settings lock once the training is created.
3
Set steps, learning rate, Rank, Alpha and batch size, check the cost, start training
Descriptions save automatically. If you leave the page, pick up again from Training tasks. You can close the page while descriptions are generating, while the training is queued, and while it runs.
Basic Settings
Descriptions
Advanced Parameters
Preparing your images
Your image set decides what the LoRA learns. Keep the features you want the LoRA to remember consistent in every image, and let everything else vary. That contrast is how the model tells the LoRA itself apart from things that just happened to be in the picture. For a character LoRA:
GoodBadThe stick figures in the bad examples are illustrations
Angles
Include front, side and back views, plus high and low angles. A set with only front views tends to fall apart when you generate the character from the side or back.
Poses and actions
Mix standing, sitting, walking, reaching and other poses. If every image shows the same pose, the LoRA tends to lock onto it.
Framing and distance
Combine close-ups, half-body and full-body shots. With close-ups only, the model has a hard time learning the outfit and body shape.
Expressions
If you want to control expressions when you generate, include images with different expressions.
Backgrounds
Vary the backgrounds as much as you can. If most images share one background, the LoRA learns the background too.
Image quality
Pick clear images with no watermarks or text. Crop multi-panel images and comic pages into single images first.
No near-duplicates
Leave out images that are almost identical, such as the same picture at different sizes. Duplicates get overlearned.
Character design
Use one version of the character design. Remove watermarks, signatures and speech bubbles where you can, drop blurry or broken images, and crop away large empty areas.
Style LoRAs work the other way round: vary the characters, scenes and subjects as much as possible, and keep only the art style consistent.
Training several characters in one LoRA
To teach one LoRA two characters, say Character A and Character B, your image set needs three kinds of images: A alone, B alone, and A and B together.
What the image set needs
10+ images per group Groups about the same size
With solo images only, the model struggles to draw both characters in one scene and often blends their features. With group images only, it's hard to generate either character alone. The groups don't need identical counts, but each needs enough images: at least 10 per group, with the groups roughly the same size.
Each LoRA has one trigger word field, so you tell the characters apart by how you name them in the descriptions. Pick a fixed name for each character and use it in every image. Leave out features that never change (hair color, eye color) so they attach to each name. In group images, say where each character is and what each one is doing. When you generate, write the name in your prompt to choose who appears. For the generation side, see our multi-character LoRA generation guide.
Settings PixAI handles for you
These are fixed and can't be changed.
Training resolution
The base model sets the training resolution. Each image needs at least 512×512 pixels in total (about 262,000 pixels) and an aspect ratio between 1:3 and 3:1. Images outside these limits can't be added.
Because the check uses total pixels, an image can pass even if one side is shorter than 512. A 400×700 image has 280,000 pixels, so it's accepted. A 1200×300 image has enough pixels, but its 4:1 ratio is outside the limit, so it's rejected.
Image sources and formats
Upload from your device, pick from your generation history, or import from a past training. PNG, JPG and WebP are supported. Each training needs 10 to 100 images.
Network type
Standard LoRA only. LyCORIS variants such as LoCon and LoHa aren't available.
STEP 1
Basic Settings
Enter a LoRA name, choose the LoRA type (Character, Outfit, Art style or Something else), pick the base model (Tsubaki.2 or Tsubaki.3), and set your trigger words.
You can come back to this step any time before you submit. After you submit, you can still cancel while the training is queued: it turns back into a draft and you can edit this step again. Once training starts, the LoRA type, base model and trigger words are locked, and you'll need a new training to change them. You can rename the LoRA after publishing from Edit Model. Trigger words are added to every description during training, so double-check them before you start.
Once training starts
- LoRA name Editable after publishing
- LoRA type Locked
- Base model Locked
- Trigger words Locked
Trigger words
Enter trigger words once in Basic Settings. PixAI adds them to the start of every description during training, so you don't need to put them in folder names or caption files as with kohya. To use the LoRA, include the trigger words in your prompt. They can't be changed after you submit. Only letters a–z, numbers, spaces and , . _ - are allowed. There's no minimum length, the limit is 256 characters, and consecutive spaces or spaces at the start or end aren't allowed. For tips, read our LoRA trigger words guide.
STEP 2
Descriptions
Upload images from your device, pick them from your generation history, or import them from a past Advanced Training. Imports bring their descriptions along, so there's nothing to regenerate and no description cost. You can still edit them after importing. Datasets from Basic Training can't be imported into Advanced Training. For more on reusing image sets, see our multi-version LoRA guide.
- You can import a dataset as long as its Advanced Training draft still exists. If you delete a LoRA made with Advanced Training, its images and descriptions stay available for import.
- Deleting the draft is different: the dataset goes with it, and you can no longer import from it. If you may want to retrain with the same images later, keep the draft or back up the images. Otherwise the descriptions have to be generated again and charged per image.
Image descriptions (captions)
Descriptions are the equivalent of kohya caption files. PixAI's description model writes a natural-language description for every image instead of Danbooru tags. Each description has to be generated before you can edit it. Your edits are saved separately, and you can restore the generated version at any time.
A description tells the model what's in each image. During training, the trigger words go in front of every description. Anything a description mentions is treated as something you can change with your prompt. Anything it never mentions gets absorbed into the trigger words and becomes part of the LoRA. So leave out the features you want the LoRA to keep, and describe the things you want to swap when you generate.
What a description does (illustration)
trigger words A girl sitting on a park bench, wearing a red coat and smiling.
Dark: trigger words added by PixAI · Blue: things you wrote, which you can swap in your prompt · Not written (hair color, eye color, signature accessories) → absorbed into the trigger words and fixed in the LoRA
The tips shown in the interface:
- Describe only what you can see. Skip personality, backstory and anything off screen.
- Describe things you may want to change later, such as outfits, and include images with different outfits.
- When more than one person is in the image, make it clear who is doing what. Avoid phrases like "one of them".
- Leave out features that never change, like hair color, eye color and signature accessories, so they become part of the character.
- Use the same word for the same thing across the whole image set. If you pick "sitting", keep using "sitting".
- Don't type the trigger words yourself. They are added to the front of every description automatically.
Every image needs a description before you can move on. If any image is still not described, the interface asks you to finish first.
Credits and training time
Descriptions and training are billed separately in credits. Both show their cost before you confirm.
① Description cost · paid when you click Describe Automatically
100 × images + 500
Images that already have a description aren't charged again.
If it fails: only images described successfully are charged. Failed images can be retried.
② Training cost · paid when you click Start training
About 70,000 credits
Depends on your settings; about 70,000 credits with the defaults. Training Steps and batch size have the biggest effect. The quote shown before you start is the final price.
If it fails: training isn't charged, description costs aren't refunded. Use Try again with the same settings in Training tasks.
Two things differ from a kohya setup
Descriptions are a separate paid stage
Each run of automatic descriptions on undescribed images costs 100 × images described + 500 credits, shown before you confirm. Described images aren't refunded even if you delete them later.
Described images are never charged twice
Automatic descriptions only process images without one, so when you add images later you pay only for the new ones.
Advanced Training charges credits twice:
- Description cost: paid when you click Describe Automatically in the Descriptions step.
- Training cost: paid when you click Start training in the Advanced Parameters step, at the quoted price.
Examples
- First trainingYou add 30 images and click Describe Automatically. You pay per image plus 500 credits, and the exact amount appears first. After checking descriptions and setting parameters, you click Start training and pay the quoted cost (about 70,000 credits with the defaults).
- Adding imagesYou add 5 more images. Only the 5 new ones are described and charged per image, plus another 500 credits. The original 30 aren't charged again.
- Some descriptions fail2 images can't be described. You pay for the 28 that succeeded, and the 500-credit fee still applies. Running automatic descriptions again on the 2 failed images adds another 500 credits.
- Training failsThe training isn't charged, but description costs aren't refunded. Use Try again with the same settings in Training tasks.
- CancellingCancel while queued and the training cost is refunded in full. Description costs aren't. The task becomes a draft you can edit and resubmit.
Training time
2–6 h excluding queue
Training time and credit cost depend mainly on Training Steps and batch size. Learning rate, Rank and Alpha don't change the cost and barely affect time. After you submit, the training waits in a queue first, so the finish time depends on how busy it is. See why LoRA training can take longer.
Advanced Training takes about 2 to 6 hours, not counting queue time. Your image set is locked while training runs and unlocks if the training fails or you cancel it in the queue. If training pauses, it resumes automatically at no extra cost.
Checklist before you start training
Before you click Start training, check these:
- Images: at least 512×512 pixels in total, aspect ratio between 1:3 and 3:1, PNG, JPG or WebP, 10 to 100 images.
- Image content: enough variety in angles, poses and framing, with no multi-panel images or near-duplicates. For a multi-character LoRA, each character alone plus images of them together.
- Basic Settings: LoRA name, type, base model and trigger words confirmed. They can't be changed after you submit.
- Descriptions: checked one by one, describing only visible things that vary, with the same word for the same thing. Every image has a description (Not described yet shows 0), or you can't continue.
- Parameters: with no specific goal, keep the defaults (325 steps, learning rate 6e-4, Rank 64, Alpha 32, batch size 2).
- Cost and time: you've checked the quoted cost and estimated time. You can only cancel while queued (training cost refunded, description cost not). Once training starts it can't be cancelled.
After training: publish and manage
When training finishes, click Publish in Training tasks. After that you can change the settings below any time from Edit Model on the model page.
Visibility
- Private Hidden from others, and you can make it public later. Only members can keep LoRAs private, and there's a limit on how many; check your remaining slots under Edit Model. See Rule Update: Deleting Private LoRAs.
- Public Anyone can use it. A public LoRA can't be made private again.
Rebates
Under Rebate settings in Edit Model, choose Opt in to rebate or Opt out of rebate. Once you opt in, you earn credits based on how often other users generate with your LoRA. Your own use doesn't count, so a private LoRA earns nothing even with rebates on.
Opting in can't be undone, and the LoRA can no longer be deleted. We recommend opting in only after the LoRA is public and you're sure you won't need to delete it.
Full details: LoRA Rebate Upgrade.
Downloads and commercial use
Tsubaki.2 and Tsubaki.3 are DiT models (shown as DiT.2 and DiT.3 on model pages). LoRAs trained on them can't be downloaded and can only be used on PixAI. See our DiT LoRA training guide.
Under Permissions in Edit Model, choose whether others may use images generated with this LoRA commercially.
Model details
You can change the model name, tags, description and cover any time in Edit Model. The LoRA type and base model can't be changed. Your training settings and final descriptions are saved with the model, so you can check them later and other users can find your LoRA more easily.
FAQ
Do I need to write the trigger words into every description?
No. They're added to the start of every description automatically. Enter them once in Basic Settings, and include them in your prompt when you generate.
Can I skip automatic descriptions and use my own captions?
Not at the moment. Every image needs a generated description before you can edit it, and importing existing caption files isn't supported.
Why are there no epochs or repeats?
Advanced Training counts total steps (Training Steps), and every image has the same chance of being picked. See Training Steps for how to convert.
Why can't I set Alpha to 1?
Alpha can only be half, equal to, or double your Rank. The actual scaling comes from the ratio of Alpha to Rank.
Why is batch size limited to 1, 2 or 4?
Training cluster limits. A larger batch size gives steadier updates but processes more images per step, which raises training time and credit cost. The default of 2 works in most cases.
Does Rank 64 slow training down a lot?
Barely, and it doesn't add to the credit cost. Time and cost depend mainly on Training Steps and batch size. Rank mostly affects file size and how much detail is kept.
Can I change the base model after training starts?
No. Once training starts, Basic Settings are locked and you'll need a new training. Before that you can still change them: go back before you submit, or cancel while queued to turn the task back into a draft.
Can I cancel after submitting?
Yes, while the training is queued. Once it starts, it can't be cancelled. Cancelling from Training tasks refunds the training cost in full but not description costs. The task becomes a draft, and you can edit the images, descriptions and Basic Settings (including the base model) before resubmitting.
Can I download my trained LoRA?
Not right now. Tsubaki.2 and Tsubaki.3 are DiT models, and LoRAs trained on them can only be used on PixAI.
To apply your LoRA and set its weight when generating, see how to use LoRA models and our LoRA weight settings guide.
Images ready?
Run through the checklist, then start your first Advanced Training on PixAI.