How to Build a LoRA Dataset From Scratch: Where to Get Your Training Images

⚡ Quick Answer

You don't need a photoshoot to build a LoRA dataset. Use your own photos when you have 15 or more of the subject, stock or web images for a style LoRA, and AI-generated reference images — through a chat tool like ChatGPT, Google's Nano Banana, or a ComfyUI character-consistency workflow — when you only have one or two photos to start from. Most real datasets end up mixing two of these methods.

Most LoRA training guides jump straight into settings and assume you already have a folder of images ready to go. The real question that stops people before they get there is simpler: where do those images actually come from? This guide covers the four practical ways to build a LoRA dataset— your own photos, stock or web images, AI-generated reference variations, and ComfyUI's own character-consistency workflows — so you know exactly which one fits your situation before you collect a single image.

What You Need Before You Start

You're gathering images here, not training anything yet, so there's no special hardware required. Depending on which method below fits your situation, you'll need one of the following:

Your own photos — any phone camera works
A free stock photo account, or just a browser for Google Images
A free ChatGPT account, or a paid Google Gemini subscription (for its "Nano Banana" image model) for reference-image generation
ComfyUI already installed, if you want the local character-consistency workflow method — see the ComfyUI installation guide if you haven't set it up yet

Which Method Should You Use?

Start here. Match your situation to a method below, then jump to that section — you don't need to read every method if only one applies to you.

Your Situation	Best Method
You already have 15+ real photos of the subject	Use your own photos
You’re training a style, not a specific subject	Stock or web images
You have 1–3 photos of a person or character	AI-generated reference images
You have zero photos, only a concept or description	AI-generated + ComfyUI consistency workflow

How to Use Your Own Photos for a Character or Product LoRA

If you're training a LoRA on a real person, your own pet, or a physical product, your own photos are almost always the strongest starting point — they're already the exact subject, with no risk of an AI tool drifting away from the real appearance.

Check your phone's camera roll first. Most people already have more usable photos than they think, scattered across old events, casual snapshots, and product listings. If you're short on real photos and can take new ones, try to capture several angles and lighting setups in one sitting rather than repeating the same shot.

Tip: Once you've gathered your photos, the next question is which of them are actually good enough to train on. That's covered in full — resolution, blur, watermarks, and variety — in Fix Your LoRA Training Dataset Before You Hit Train.

Where to Find Stock or Web Images for a Style LoRA

A style LoRA is different from a character LoRA — instead of one consistent subject, you need many different subjects that all share the same visual look. This is where stock and web images work well, since you're not trying to keep one identity consistent across the set.

Free stock sites like Unsplash and Pexels are a good starting point for realistic styles, since their photos are high resolution and licensed for broad reuse. Google Image Search can widen your search further when you need a specific look that stock sites don't cover, but the results come with mixed and often unclear licenses.

Warning: Images from stock sites and web searches come with usage terms attached — check the license before training on anything you plan to share or sell a LoRA from. Free stock sites like Unsplash's license page spell this out clearly. Random web images generally don't, so treat those as personal or experimental use only unless you can confirm the source.

How to Generate a Full Image Set From One Reference Photo Using AI

This is the method to use when you only have one or two photos of your subject — a single clear reference photo can be turned into a full set of pose, angle, and background variations using an AI chat tool, without needing any new real photos at all.

ChatGPT can do this for free: upload one clear reference photo, then ask it to generate the same subject in a different pose, outfit, or background while keeping the face the same. Google's Gemini app, using its "Nano Banana" image model, does the same job as a paid option, and many people find it holds the identity slightly more consistently across generations — see the full Nano Banana workflow guide for the complete setup and prompting approach.

Upload your one clear reference photo to the chat tool. This becomes the identity anchor every later generation is compared against.
Ask for one specific change at a time — a new pose, a new background, or a new outfit — rather than describing a whole new scene from scratch. Small, specific changes hold the face steady better than big ones.
Save each result before generating the next one. This keeps a record of every variation and lets you go back to your original reference if a later generation starts drifting.
Repeat for different angles: front-facing, side profile, and three-quarter view, plus a mix of close-up and full-body shots.
Review the full set before moving on — check that the face still matches your original reference photo in every image, not just the first few.

One reference photo next to a AI-generated pose and background variations of the same subject — One reference photo on the left, A AI-generated variations on the right — same face, different pose and background.

Tip: Generate from your original reference photo each time, not from a photo the AI already generated. Chaining generations off of generations is the fastest way to lose the likeness — see the identity-drift issue in the troubleshooting section below.

Fix Inconsistent AI Characters: A Ready-to-Use 20+ Copy-Paste ChatGPT Prompt Sheet

The method above works, but writing a new prompt from scratch for every pose is slow, and it's easy to accidentally change the wording enough that the face drifts. This section gives you a ready-made prompt sheet built around one consistent character, so you can copy a prompt, paste it into ChatGPT, and move straight to the next one.

Every prompt below assumes you've already generated (or uploaded) one anchor photo and are re-using it as the reference for each new image — the same rule covered in the identity-drift section further down.

Step 1: Generate Your Anchor Photo

Paste this into ChatGPT first. This is the one photo every other prompt on this page builds from — save it once it's generated.

Anchor prompt

Generate a photorealistic image of a young French woman in her mid-20s, with fair skin and short, curly red hair. Give her a healthy, curvy body type — not skinny, not athletic, somewhere between average and fuller-figured. Frame the shot from the waist up, not too close, in landscape orientation, with natural lighting and a simple background. This becomes your anchor photo. Save it, and re-upload it as a reference for every prompt below instead of generating from a previous AI output.

Anchor reference photo of the red-haired French woman used as the base for every prompt below — Your Step 1 result should look something like this — this is the photo you re-upload for every prompt below.

Step 2: Generate Style & Scene Variations

Each group below covers a different setting. Expand a group, hit "Copy all," and paste the whole batch into ChatGPT — it'll work through the numbered prompts one at a time.

Sample grid of Everyday & Travel results generated from the anchor photo — What the "Everyday & Travel" batch looks like once generated — same face and hair across every result.

Sample grid of Glamour & Event results generated from the anchor photo — What the "Glamour & Event" batch looks like once generated — same face and hair across every result.

Sample grid of At-Home & Candid results generated from the anchor photo — What the "At-Home & Candid" batch looks like once generated — same face and hair across every result.

Step 3: Generate Structured Outfit & Background Sets

These prompts spell out the outfit, pose, background, and shot type separately — useful once you want tighter control over a specific look, like a fashion-editorial or travel-style set.

Sample grid of Style Set 1: Travel & Casual Looks results generated from the anchor photo — What the "Style Set 1: Travel & Casual Looks" batch looks like once generated — same face and hair across every result.

Sample grid of Style Set 2: Glamour & Editorial Looks results generated from the anchor photo — What the "Style Set 2: Glamour & Editorial Looks" batch looks like once generated — same face and hair across every result.

Warning: If a result starts looking like a different person, don't keep building on it — go back to your anchor photo from Step 1 and generate the next prompt from that instead. See the identity-drift fix in the troubleshooting section below for the full explanation.

Download a Free Sample LoRA Dataset (No ChatGPT Needed)

Don't want to generate the images yourself? This sample dataset was built using the exact prompts above, so you can see what a finished set looks like, or drop it straight into a LoRA trainer to see how the training process works before building your own.

📦 Free Sample LoRA Dataset

A ready-made set of AI-generated images built from the prompts above, so you can see exactly what a finished training dataset looks like — or use it to test your first LoRA training run without generating anything yourself.

⬇ Download Sample Dataset (.zip, 55 MB)

How to Generate Character-Consistent Training Images With ComfyUI

A chat tool works well to get started, but it charges per generation (in the case of Gemini) or slows down under heavy use (in the case of free ChatGPT). If you're building datasets regularly, a ComfyUI character-consistency workflow— a node-based setup that generates variations of a reference image locally on your own GPU — is a free, faster alternative once it's set up.

Earngenix already has full setup guides for four of these workflows, each suited to a slightly different job:

Character consistency in ComfyUI — the general-purpose starting point for keeping a character consistent across generations
Nano Banana in ComfyUI — run Google's Nano Banana model locally in ComfyUI instead of paying per generation in the Gemini app
Krea 2 character consistency workflow — best when you want strong pose variety while keeping the face steady
Flux Kontext character consistency workflow — best for editing a character directly into new outfits or settings
SCAIL-2 character replacement workflow — best for dropping your reference character into entirely new backgrounds and scenes

Grid of character-consistent images generated locally with a ComfyUI character consistency workflow — A batch of dataset-ready images generated locally from one reference photo using a ComfyUI consistency workflow.

Should You Combine Methods for One Dataset?

Yes — this is the most common real-world setup, not an edge case. A typical mix is 5–10 real photos to anchor the actual identity, plus 10–15 AI-generated variations to fill in angles, poses, or settings the real photos don't cover.

Warning: Don't mix wildly different quality levels without noticing — a set of sharp real photos next to soft, low-resolution AI generations will teach the model inconsistent detail. Match resolution and sharpness across sources as closely as you can before finalizing the set.

Common Mistakes When Sourcing a LoRA Dataset

My AI-generated images look like a different person after a few tries

What causes it: Identity drift — generating a new image from a previous AI generation instead of your original reference photo, so small errors compound with each step.

How to fix it: Go back to your original reference photo for every new generation, and keep the number of variations per reference to around 8–15 before the drift becomes noticeable.

My stock images don't match in style or lighting

What causes it: Pulling images from multiple stock sites or search results without checking that they share a consistent color grade, lighting style, or overall look.

How to fix it: Review your full set side by side before finalizing it, and remove any image that stands out sharply in tone or lighting from the rest.

I'm not sure if I'm allowed to use the images I found

What causes it: Collecting images from Google Image Search or unclear sources without checking the license attached to them.

How to fix it: Stick to stock sites with a clear license (Unsplash, Pexels) for anything beyond personal or experimental use, and check the source page directly if you're unsure.

Frequently Asked Questions

Yes, and it’s actually one of the most common real-world setups. A handful of real photos anchors the identity, and AI-generated variations fill in angles or settings you don’t have real photos for. Just make sure the resolution and quality level don’t vary too wildly between the two sources.

It depends on the image’s license and what you plan to do with the trained LoRA. Personal or experimental use is generally treated differently than sharing or selling a LoRA trained on images you don’t have rights to, so check the source before assuming it’s fine. When in doubt, stock sites with clear licenses are the safer choice.

In practice, somewhere around 8-15 generations before you start noticing the face drifting from the original reference. Past that point, feed the AI tool your original reference photo again instead of one of its own outputs, since generating from a generation compounds the drift.

A chat tool is enough to get started and costs nothing to try. ComfyUI character-consistency workflows are worth setting up once you’re generating datasets regularly, since they run locally, don’t charge per image, and give you more control over pose and background.

Stock photos work best for style LoRAs because you need many different subjects in one consistent look. For a character LoRA you need the same person or subject across every image, which stock libraries generally can’t provide unless you’re training on a recognizable public figure.

This is identity drift, and it happens when you generate too many variations in a row or generate from an already-generated image instead of your original reference. Go back to your original reference photo, regenerate from there, and keep each new dataset image within a few generations of that original.

No. A free ChatGPT account can generate every prompt on the sheet, though free accounts generate more slowly and may hit a temporary cooldown after several images in a row. A paid plan just removes that wait.

What to Do Next

Once your images are collected, check them against the dataset quality checklist before you train.

Resolution, blur, variety, and what makes an image usable are all covered in the follow-up guide.

Check Your Dataset Quality →ComfyUI Roadmap →

Published: 2026-08-17 · Last updated: 2026-08-19

Discussion

Join the discussion

💬

No comments yet

Be the first to share your thoughts!