Multi-Version LoRA System & Dataset Reuse on PixAI: Save 50% on LoRA Training
Manage multiple versions of one LoRA on PixAI and cut training costs by 50% with dataset reuse. Full guide to versioning, archiving, and discount rules.
PixAI’s multi-version LoRA system lets you manage multiple versions of a single LoRA under one model card — no more cluttered uploads, no more guessing which file is the latest. Paired with the dataset reuse feature, you can also retrain on a previously uploaded dataset and get an automatic 50% discount on LoRA training costs. This guide covers exactly how both features work, how to switch between LoRA versions during generation, when the 50% discount actually applies (and when it doesn’t), and how to edit or archive specific versions.
Table of Contents
What Is the Multi-Version LoRA System on PixAI?
The multi-version LoRA system lets you stack multiple versions of the same LoRA — different training runs, different base models, different hyperparameter tweaks — under a single model card. Instead of uploading a brand-new LoRA every time you iterate, your followers see one consistent model page and switch between versions with a click.
Why this matters in practice:
- Cleaner pages, easier discovery — followers find your work in one place rather than scattered across half a dozen near-identical uploads.
- Cross-base-model support — the same LoRA card can host versions trained against different DiT base models. For example, you can have a Tsubaki version and a Tsubaki.2 version of the same character LoRA living together. (For which DiT bases are currently trainable, see our DiT LoRA Training Guide.)
- Cleaner version history — older runs stay accessible (until you archive them), so users can pin a specific version they prefer.
How to Add a New Version to an Existing LoRA
Adding a new version to a multi-version LoRA takes three steps:
Go to the model’s detail page by clicking on the LoRA card.
Open the version manager via the icon in the top-right corner of the page.
Click Add version.
You’ll be redirected to the LoRA training page.
After training completes, the new version is automatically attached to the existing model card — no manual upload required.
This is the workflow you’ll use most often when:
- A new base model launches and you want to migrate (e.g., Tsubaki → Tsubaki.2)
- You’ve collected better training data and want a refined v2
- You want to publish a “fast” variant alongside a “high-quality” variant
How to Switch Between LoRA Versions When Generating
Switching multi-version LoRA versions is built into both the model page and the generation panel.
From the model detail page
Select the version you want, then click Use this LoRA.
The compatible base model loads automatically — no need to remember which base each version was trained against.
From the generation panel
After selecting a LoRA, you’ll see a version selector. You can switch between supported base model versions on the fly without leaving the generation flow.
The second flow matters more than it sounds. If you’re testing a character LoRA across base models — say, comparing how it renders on Tsubaki vs. Tsubaki.2 — you can flip versions inline without re-selecting the LoRA each time.
Dataset Reuse: Cut LoRA Training Costs by 50%
Most LoRA iteration involves retraining on the same images. You might be tweaking trigger words, moving to a different base model, or pushing a refined version after better captioning. Re-uploading the same dataset every time wastes time and credits.
PixAI’s dataset reuse feature solves both problems: it loads your previous dataset automatically and applies an automatic 50% discount on training cost. There are two ways to trigger it.
Method 1 — Quick Dataset Reuse from Model Details
This is the fastest path and works for most cases.
Go to your personal LoRA list and select the model you want to update.
On the model detail page, click the three-dot menu (⋮) in the top-right corner.
Open the LoRA model’s detail page and click Edit Model.
The system loads the dataset from your latest version automatically, and you’ll see the 50% discount applied to the training cost before you submit.
Method 2 — Reuse a Specific LoRA Version’s Dataset
If you want to reuse the dataset from an older version — not just the latest one — use this longer path:
- Open the LoRA model’s detail page and click Edit Model.
- Select the version whose dataset you want to reuse.
- You’ll be redirected to the training page with that version’s dataset already loaded.
The Edit Model page is also where you go to manage versions individually — for example, editing a specific version’s details or archiving one (more on archiving below).
When the 50% Dataset Reuse Discount Applies (and When It Doesn’t)
This is where most users get tripped up. The 50% discount is conditional on dataset integrity:
| Action | 50% discount? |
|---|---|
| Reusing the original dataset without changes | ✅ Yes |
| Reusing the dataset and adding new images | ❌ No |
| Reusing the dataset and removing images | ❌ No |
| Creating a new version with a completely new dataset | ❌ No |
The rule is simple: the dataset must be reused unchanged. The moment you add, remove, or replace any image, the system treats it as a new dataset and bills it at the standard rate.
If you do need to change the dataset, expect the full price — but the multi-version system still saves you the upload friction and keeps everything organized under one model card.
Editing and Archiving LoRA Versions
The Edit Model page lets you manage versions one at a time. You can:
- Edit a version’s name, description, and other display details
- Archive a version you no longer want available
A note on archiving — this is a one-way action.
⚠️ Archiving is permanent Once a version is archived, it cannot be used and cannot be restored to its previous state. Archive only when you’re certain you no longer want a version available.
If you just want to discourage use of an old version without removing it, leave it in place — most users will gravitate to your latest version by default.
Tracking Training Tasks and Reusing Datasets from the Queue
The Training Tasks section gives you a live view of any LoRA training currently in progress, including the estimated training time. The Actions column lets you reuse a dataset directly from there — handy when you want to queue a follow-up training right after the current one finishes.
A reminder on timing: the estimated wait shown in the queue is for reference only. Actual training duration depends on system load and dataset size, and the initial estimate is the least accurate part of the run. Submit the job, walk away, and check back later rather than refreshing.
Multi-Version LoRA FAQ
What is the multi-version LoRA system on PixAI?
The multi-version LoRA system lets you publish multiple versions of the same LoRA under a single model card. Each version can target a different base model or use different training settings, but users see them all in one place and can switch between them when generating.
How do I add a new version to an existing LoRA?
Open the LoRA’s model detail page, click the icon in the top-right corner to open the version manager, and click Add version. You’ll be sent to the training page; once training completes, the new version is automatically attached to the same model card.
How does the 50% LoRA training discount work?
PixAI applies an automatic 50% discount on LoRA training cost when you retrain using a previous dataset without modifying it. The discount is visible before you submit. If you change the dataset (add or remove images, or upload a brand-new one), the discount does not apply.
Why didn’t I get the 50% dataset reuse discount?
The discount only applies when the dataset is reused unchanged. If you added images, removed images, or uploaded a completely new dataset for the new version, the system bills at the standard rate. To get the discount, leave the dataset exactly as it was on the source version.
Can I reuse a dataset from an older LoRA version, not just the latest?
Yes. Use Edit Model on the LoRA detail page, select the specific version whose dataset you want, and you’ll land on the training page with that version’s dataset preloaded. The quick-reuse three-dot menu only loads the latest version.
What happens when I archive a LoRA version?
Archiving permanently removes a version from active use. Archived versions cannot be used and cannot be restored to their previous state. Only archive when you are sure you no longer want a version available.
Can I have a Tsubaki and a Tsubaki.2 version under the same LoRA card?
Yes. The multi-version LoRA system supports versions trained against different base models, including across DiT generations. This is the recommended workflow for migrating an existing LoRA to a newer base model. See our DiT LoRA Training Guide for which DiT bases are currently trainable.
Start Using Multi-Version LoRAs Today
Both the multi-version LoRA system and dataset reuse are live on PixAI right now. They cost nothing extra to use, and dataset reuse actively saves you 50% on retraining whenever you keep the dataset unchanged.
