Advanced LoRA Training

PixAI Advanced Training is for users who want finer control over their LoRA training results. You can review and edit the description of each image one by one, and set the Training Steps, learning rate, detail capacity (Rank), Alpha and Batch size yourself; all parameters default to the same values as Basic Training.

If you're familiar with kohya / sd-scripts or other community LoRA training tools, note three differences first: Advanced Training measures training volume in total steps and doesn't offer epochs or repeats; Alpha can only be set to half, equal to, or double the Rank; and image descriptions are generated automatically by PixAI, after which you edit them one by one.

You can open Advanced Training in another tab and follow along with this guide as you set it up:

▸ Open PixAI Advanced Training

Key Points

• Training volume is measured in total steps: "Training Steps" can be set from 50–800 (default 325); epochs and repeats aren't available.
• There is only one learning rate: The default is 6e-4. It applies to the whole network and can't be set per module.
• Alpha has three levels: It can only be set to half, equal to, or double the Rank, and defaults to half the Rank. Credit cost depends mainly on Training Steps and update smoothing (Batch size); learning rate, detail capacity (Rank) and Alpha don't affect credits.
• Image descriptions must be generated automatically first: Descriptions are in natural language, not Danbooru tags, and can be edited one by one once generated.
• Trigger words are automatically added to the start of every description: Once you fill them in during the first step, "Basic Settings", the system adds the trigger words to every image's description — no need to add them to each image by hand.
• Automatic descriptions and training are billed separately: Images that already have descriptions aren't charged again; if training fails, the training itself isn't charged; deleting images that already have descriptions doesn't get you a refund.

Training Workflow

Advanced Training has three steps: Basic Settings, Descriptions and Advanced Parameters. The overall order is similar to the LoRA training workflow commonly used in the community. The main differences are that image descriptions (captions) are generated by PixAI, and some settings can't be changed once the training is created.

Descriptions are saved automatically; if you leave the page partway through, you can pick up again from Training tasks. You can close the page while automatic descriptions are being generated, while queued, and while training is in progress.

Compared with how kohya is typically used, there are two differences:

• Automatic description is a separately billed stage: Each time you run automatic description on images that haven't been described yet, it costs 100 × the number of images described this time + 500 credits. The cost is shown before you start, and you can edit each description once it's generated. Images that have already been described automatically won't have their credits refunded, even if you delete them later.
• Images that already have descriptions aren't charged again: Automatic description only processes images that don't have a description yet; when you add images later, you're only charged for the new ones.

Preparing Images

Your image set determines what the LoRA can learn. When choosing images, keep the features you want the LoRA to remember consistent across every image, and make everything else as varied as possible, so the model can tell what belongs to the LoRA itself and what just happens to appear. Taking a character LoRA as an example:

• Angles: Include front, side and back views, as well as different perspectives such as high-angle and low-angle shots. An image set with only front views tends to go off-model when generating side or back views.
• Actions and poses: Cover different poses such as standing, sitting, walking and reaching out. If the whole set uses the same pose, the LoRA tends to get stuck in that pose.
• Composition and distance: Mix face close-ups, half-body and full-body shots. With only close-ups, it's harder for the model to learn the outfit and body shape.
• Expressions: If you want to control expressions when generating, include images with different expressions.
• Backgrounds: Keep backgrounds as varied as possible. If most images use the same background, the background gets learned too.
• Image quality: Choose clear images without watermarks or text. Crop multi-panel images or storyboards into single images before adding them.
• Avoid duplicates: Don't include near-identical images, such as different-size versions of the same image — duplicates get overlearned.
• Character design: Use only one version of the character's design; remove watermarks, signatures and speech bubbles where possible; drop blurry or badly drawn images; crop out large empty areas.

For an art style LoRA, the approach is the opposite: make the characters, scenes and subjects as varied as possible, and keep only the art style consistent.

Teaching One LoRA Multiple Characters

If you want one LoRA to learn multiple characters at once — for example, Character A and Character B — your image set needs three kinds of images: images with only A, images with only B, and images with A and B together. With only solo images, the model has trouble drawing the two separately in the same scene, and the two characters' features commonly get mixed together; with only images of them together, it's hard to generate either one alone. The three kinds don't need exactly the same number of images, but each should have enough (at least 10 per kind, with the counts as close as possible).

Each LoRA can currently have only one trigger word, so the characters need to be distinguished by how they're referred to in the descriptions. Choose a fixed name for each character and use that same name throughout the whole image set; likewise, don't describe each character's unchanging features (such as hair color and eye color), so those features get attached to their respective names. In images with both characters, describe each one's position and action separately. When generating, write the corresponding name in your prompt to specify which character to draw.

Parameter Overview: Advanced Training vs. kohya / sd-scripts

Parameter Details

Training Steps (Training Length)

Training Steps determines how many times the model studies your images. The default is 325, adjustable from 50–800 in increments of 5. The UI recommends: unless your image set is unusually large or small, just leave it at 325.

Advanced Training doesn't offer epochs or repeats; every image in the dataset has the same chance of being sampled. If you need to convert to kohya-style settings, you can estimate the average number of times each image is trained like this:

Average times each image is trained ≈ Training Steps × conversion factor ÷ number of images

• Turn it up: The model learns the visual features more deeply, but may also start copying the backgrounds and poses of the training images. Training time and credit cost go up as well.
• Turn it down: Training is shorter, but may not pick up the features you want.
• Above 500 steps, the UI warns that training takes longer; going above that is generally not recommended.

Learning Rate

The learning rate determines how large an adjustment the model makes each time it learns. The default is 6e-4 (0.0006), with a range of 5e-5 (0.00005) to 1e-3 (0.001). Advanced Training has only one learning rate, applied to the whole network; it can't be set per module as in kohya. The learning rate doesn't affect credit cost.

• Turn it up: Learns faster, but may overshoot and make images unstable.
• Turn it down: Learns more cautiously, but needs more steps.
• UI recommendation: If results look messy or burnt, lower the learning rate.

Detail Capacity (Rank)

Detail capacity corresponds to kohya's network_dim and represents how much detail the LoRA can hold. Options are 8 / 16 / 32 / 64, with a default of 64. In the UI, they're labeled "Very light", "Light", "Balanced" and "Recommended" respectively. Rank mainly affects file size and how much detail can be retained; it has very little effect on training time and doesn't affect credit cost.

• Turn it up: Holds more detail, but also memorizes the training images' backgrounds sooner.
• Turn it down: Smaller file, but may fail to learn complex patterns and accessories.

Alpha

Alpha corresponds to kohya's network_alpha and determines how much each training update affects the LoRA. To stay consistent with the Rank, Alpha can only be set to half, equal to, or double the Rank, labeled "Smoother", "Balanced" and "Sharper" in the UI respectively; the default is half the Rank. When you change the Rank, Alpha keeps its current level (Smoother, Balanced or Sharper) and its value adjusts automatically. Alpha doesn't affect credit cost.

• Turn it up: Learns faster, but also picks up flaws in the images just as fast.
• Turn it down: Learns more gently; suited to image sets with fewer images or mixed styles.

The actual scaling is determined by the ratio of Alpha to Rank. Taking Rank 64 as an example, the three levels set Alpha to 32, 64 and 128, giving ratios of 0.5, 1 and 2; the common kohya setting of alpha = 1 gives a ratio of only 1/64 at Rank 64.

Update Smoothing (Batch Size)

Update smoothing corresponds to kohya's batch size. The larger the value, the more images each update looks at. Options are 1 / 2 / 4, with a default of 2. Larger values aren't currently available due to training cluster resource limits.

• Turn it up: Steadier training, and individual odd images have less influence. Each step processes twice as many images, so training time and credit cost also rise noticeably.
• Turn it down: Sharper updates, but more sensitive to image quality.
• UI recommendation: Unless your images have noticeably inconsistent styles, just leave it at 2.

Batch size also affects how many times each image is trained. At the same number of steps, changing Batch size from 2 to 4 doubles the average number of times each image is trained, and training time increases accordingly; to keep the original training volume, halve the steps.

Trigger Words

Trigger words are filled in under "Basic Settings" and are automatically added to the start of every description during training, so there's no need to write them into folder names or each caption file as in kohya. When generating, add the trigger words to your prompt to activate the LoRA. Trigger words can't be changed after you submit the training. There's no minimum length, and the maximum is 256 characters; no consecutive spaces are allowed, and they can't start or end with a space.

Image Descriptions (Caption)

Image descriptions correspond to kohya's caption files. PixAI uses a description model to generate a natural-language description for each image, not Danbooru tags. Descriptions must be generated automatically before they can be edited one by one; your edits are saved separately and can be reverted to the automatic description when needed.

Descriptions tell the model what's in each image. During training, the trigger words are added to the start of every description: whatever is written in the description, the model treats as something that can be changed with the prompt; whatever isn't mentioned in the description gets attached to the trigger words and becomes part of the LoRA itself. So don't write down features you want the LoRA to lock in, and do write down things you want to be able to swap out when generating. The writing guidelines provided in the UI are:

1. Describe only what you can see — not personality, backstory or anything outside the frame.
2. Write down elements you may want to swap out later, such as outfits, and include images with different outfits.
3. When there are multiple people in the image, make it clear who is doing what; avoid phrasing like "one of them".
4. Don't describe features that never change, such as hair color, eye color and signature accessories, so they become part of the character itself.
5. Use the same word for the same thing across the whole image set — for example, once you pick "sitting", always use "sitting".
6. You don't need to write the trigger words yourself; the system automatically adds them to the start of every description.

Every image needs a description before you can move on to the next step; if any images haven't been described yet, the UI will prompt you to finish the descriptions first.

System-Determined Settings

The following are determined by the system and can't be adjusted:

Training resolution: Determined by the model and not adjustable. Training images must have a total pixel count of at least 512×512 (about 262,000 pixels) and an aspect ratio between 1:3 and 3:1; images that don't meet these requirements can't be added to the dataset.

Because the check is based on total pixels, an image can still qualify even if one side is shorter than 512: for example, a 400×700 image has 280,000 pixels and can be added; a 1200×300 image has enough pixels, but its aspect ratio is 4:1, so it can't be added.

Images can be uploaded from your device, selected from your generation history or imported from a past training, in PNG, JPG and WebP formats. Each training needs at least 10 and at most 100 images.

Network type: Fixed to standard LoRA; LyCORIS variants such as LoCon and LoHa aren't available.

Cost and Training Time

Automatic descriptions and training are billed separately, in credits, and the required credits for both are shown before you start.

Advanced Training deducts credits in two separate charges:

1. Description fee: Paid when you click "Describe Automatically" in the "Descriptions" step.
2. Training fee: Paid according to the quote shown in the UI when you click "Start training" in the "Advanced Parameters" step.

Examples:

• First training: After adding 30 images, you click "Describe Automatically" and pay a description fee based on the number of images, plus an extra 500 credits; the exact amount is shown before you start. After reviewing the descriptions and setting the parameters, you click "Start training" and pay the training fee according to the UI quote (about 70,000 credits with the default settings).
• Adding images: After the descriptions are done, you add 5 more. Automatic description only processes the 5 new images and charges based on that number, plus another extra 500 credits; the original 30 aren't charged again.
• Some descriptions fail: If 2 images can't be described, you're only charged per image for the 28 that were described successfully, and the extra 500 credits still apply. You can run automatic description again on the 2 that failed, but running it again charges another extra 500 credits.
• Training fails: The training fee isn't charged, but description fees already paid aren't refunded; you can choose "Try again with the same settings" in Training tasks.
• Canceling training: If you cancel while queued, the training fee is refunded in full, but description fees already paid aren't refunded; the task turns back into a draft, and you can keep editing and submit again.

Training time and credit cost depend mainly on Training Steps and update smoothing (Batch size); learning rate, Rank and Alpha don't affect credits and have very little effect on time. After you submit the training, the task has to queue first, and the actual completion time depends on the queue.

Advanced Training takes about 2 to 6 hours (not including queue time). The image set is locked while training is in progress; if training fails, or you cancel it while queued, the image set is unlocked.

If training is paused, the system resumes it automatically at no extra charge.

After Training: Publishing and Management

When training is complete, click "Publish" in "Training tasks". After publishing, you can adjust the following settings at any time from "Edit Model" on the model page.

Visibility

• Private: Not publicly visible; you can switch it to public at any time later. Only members can set a LoRA to private, and the number of private LoRAs is limited — you can check your remaining slots in the visibility settings under "Edit Model".
• Public: Available to all users. Once public, it can't be switched back to private.

Rebates

In "Edit Model", under "Rebate settings", you can choose "Opt in to rebate" or "Opt out of rebate". Once you opt in, you earn credits based on how many times other users use this LoRA; your own use doesn't count toward rebates, so if the LoRA is set to private, you won't earn any credits even if you opt in.

For details, see "LoRA Rebate Upgrade: No More Cap, Keep Earning Over Time".

Downloads and Commercial Use

• Tsubaki.2 and Tsubaki.3 are both DiT-architecture models (shown as DiT.2 and DiT.3 on the model page). LoRAs trained on them can't be downloaded — they can be used directly on PixAI, and only on PixAI.
• Under "Permissions" in "Edit Model", you can choose whether to allow others to use images generated with this LoRA for commercial purposes.

Model Information

The model name, tags, description and cover can be changed at any time in "Edit Model"; the LoRA category and base model can't be changed. Your training settings and final image descriptions are saved with the model, so you can check them later, and they also help other users find your LoRA.

FAQ

Why are there no epochs or repeats?

Advanced Training is measured in total steps (Training Steps), and every image in the dataset has the same chance of being sampled. For how to convert, see the "Training Steps" section.

Why can't Alpha be set to 1?

Alpha can only be set to half, equal to, or double the Rank. The actual scaling is determined by the ratio of Alpha to Rank.

Why is Batch size limited to 1 / 2 / 4?

Because of training cluster resource limits. A larger Batch size makes updates steadier, but each step has to process more images, so training time and credit cost increase accordingly; in most cases, the default of 2 is fine.

Can I skip automatic description and use my own captions?

Not currently. Every image must have a description generated automatically before it can be edited, and importing existing caption files isn't supported yet either.

Do I need to write the trigger words into every description?

No. Trigger words are automatically added to the start of every description; you only need to fill them in once in Basic Settings. When generating, add the trigger words to your prompt to activate the LoRA.

Does setting Rank to 64 make training much slower?

The effect is very small, and it doesn't increase credit cost either. Training time and cost depend mainly on Training Steps and Batch size; Rank mainly affects file size and how much detail can be retained.

Can I change the base model after training starts?

No. Once training starts, the basic settings are fixed; to change them, create a new training. You can still make changes any time before training starts: before submitting, just go back and edit; after submitting, if it's still in the queue, cancel the training first — the task turns back into a draft — then go back and edit.

Can I cancel after submitting the training?

Yes, before training starts (while queued); once training starts, it can't be canceled. After you cancel from Training tasks, the training fee is refunded in full, but description fees already paid aren't refunded. The task turns back into a draft, and you can keep editing the images, descriptions and basic settings (including the base model) before submitting again.

Can I download my trained LoRA?

Not currently. Tsubaki.2 and Tsubaki.3 are both DiT-architecture models, and LoRAs trained on them can only be used on PixAI.

Pre-Training Checklist

Before you start training, check the following:

• Images: Each has a total pixel count of at least 512×512, an aspect ratio between 1:3 and 3:1, and is in PNG, JPG or WebP format; 10 to 100 images in total.
• Image content: Enough variety in angles, actions and composition, with no multi-panel images or near-duplicates; if one LoRA is learning multiple characters, the set includes images of each character alone and together.
• Basic Settings: LoRA name, category, base model and trigger words are all confirmed — they can't be changed after you submit the training.
• Image descriptions: Reviewed one by one, describing only what's visible and what changes, using the same word for the same thing. Every image has a description ("Not described yet" is 0); otherwise you can't move on to the next step.
• Parameters: If you have no specific needs, use the defaults (325 steps, learning rate 6e-4, Rank 64, Alpha 32, Batch size 2).
• Cost and time: You've confirmed the quote and estimated time shown in the UI. After submitting, you can only cancel while queued (training fee refunded, description fees not refunded); once training starts, it can't be canceled.

After training, for how to apply the LoRA and set its weight when generating, see How to Use LoRA Models in AI Art: A Complete PixAI Guide and LoRA Weight Settings: Complete Guide to Mastering LoRA Strength.

▸ Start Advanced Training Now