Published October 3, 2025by ktiseos_nyx

Dataset Preparation Services (Commission Details)

216 views0 reactions1 comments on CivitAI5 collected

data guidedataset guidedata preplora trainingdataset editinge621dataset prep

Dataset Preparation Services (Commission Details)

An article detailing into how dataset preparation USUALLY for us works and how you could in theory commission us for it and what we can and can't do. First of all our LORA training services are available HERE.

Preparation:

Data Gathering

For data gathering you'll need to pinpoint what you're going to be training, for us thats "Concept, Character, Style". For ease of the rest of this article we'll just mention it for Style & Character as World Morphs for us are not far off of "STYLE".

To gather for a style OR Character you'll figure out your subject
Services like Booru boards, Pinterest, Screenshot Repositories etc are a great starting point, we have an article HERE on where good places are for that.
For characters when we used to train on Civitai we'd have 100-300 images, when we use our offsite trainer we use 50 images up to about 100 at the VERY MOST. We've trained on less than 25 so it just depends on how good the data is. Styles can vary, we've usually stuck to 100+ images, ignoring the whole "CONTEXT GOES BYE BYE" theory. (That's just us).

Once you've ascertained HOW and where you're getting your data you move to the next part:

Data Organization

This is KINDA the same step almost but it's easier to break it down.

Make SURE all of it's in the same folder.
If you DO have an app for duplicate photo finders -- you can use that but we prefer to do it by eye sight as Dupes only go by the file metadata and can lead to incorrect results.
If you've got 100+ images make sure you're looking in your set for the BEST of the best that match what you're doing.
Remove Duplicates.
Mass re-name for ease of use.
If you don't feel you have enough re-peat some of the process above.

Once that's all organized you go to the next stage..

Data Editing

You'll have images that need upscaling, this is critical as not every way of training is the same , there are offsite upscalers but you'll likely want to use an img2img software or Civitai's image gen use.
You'll need to remove logos from corners, signatures, and any dating effects from tht images themselves.
Cropping CAN enhance this, so if you want an easy way of just 1:1 cropping use this site HERE.
We use for SDXL anything 2048x2048 maximum size ven fi ti's a 1024 base for images, it gives the best quality output for pre existing images.
If you NEED to edit hands/eyes by hand make sure you're using your mouse OR a wacom tablet.

Once that's done? Make sure everything's where it needs to be an prep for training!

SERVICES

We've been asked "DO WE HAVE A PAGE FOR COMMISSIONS FOR DATASET PREPARATION" - and the answer is well no not really.

If you REALLY need help sourcing and editing and sourcing data for any particular reason here is our service details:

Flat fee for up to 100 images with basic cropping will be 20 USD.

Flat fee for up to 500 images with basic - 40 USD

Flat fee above that? 50 USD.

Add ons:

Advanced editing of pre existing data is going to be charged by the hour, as well as the flat fee structure. We KNOW THIS IS GOING TO RUFFLE FEATHERS BUT HEAR ME OUT:

Duskfall of Ktiseos Nyx has a three year design degree and has done professional graphic design as well as photo editing and digital art.

The FEE per hour will be on top of everything? 5 USD per hour. See? Not so bad!

We feel like if you're already doing a lot for just a lora and you only need a one hour stint help cleaning your content up- you'll likely only be paying 30 USD maximum.

NOW: Flatfee includes sourcing. However there comes some rules with this!

SERVICE RULES:

NO REALISTIC UNDERAGE CONTENT EVEN IF ITS SFW
NO REFUNDS (Except if it's within like 5 hours of payment and something REALLY Screwed up)
NO XXX CONTENT. (We're not going on E621.)
NSFW is a slider scale: Is she in a scantily clad bikini and it's just horny? -- You're Gucci! Is it currently XXX and lots of hentai sex? Yea, no we're not even taking loras for that.
NO CELEBRITIES
No real people.
No using the "BUT CIVITAI HAS DIRTY STUFF AND THEY DONT CARE ARGUMENT" - if you paid ahead of time and you use this y ou're getting no refunds :D
SHOW YOUR CONCEPT FIRST BEFORE PAYING: We reserve the right just like training to NOT process something. Plenttty of talented people, if we're not comfortable, we're not a right fit for your ideas!

Dataset Preparation Services (Commission Details)

Preparation:

Data Gathering

Data Organization

Data Editing

SERVICES

SERVICE RULES:

OTHER USEFUL ARTICLES OF OURS: