Fix Your LoRA Training Dataset Before You Hit Train

⚡ Quick Answer

A bad LoRA is almost always a bad LoRA training dataset, not bad settings. For a character or product, use 15–30 clean, varied images — close-ups, medium shots, and full-body, from multiple angles — resized to your base model's native resolution (usually 1024×1024). Style LoRAs need more: 50–150 images. More images with low variety produces a worse LoRA than fewer images with real coverage.

Your LoRA looks blurry. Or it nails the face but breaks on hands. Or every output looks like the same three training photos wearing different clothes. Before you touch rank, alpha, or learning rate, check your LoRA training dataset— that's where most bad LoRAs actually go wrong, and it's the one part of training that has nothing to do with your GPU or your trainer's settings.

This guide covers exactly how many images to use, what resolution to prepare them at, and why a handful of varied photos beats a big pile of near-duplicates every time.

How Many Images Do You Actually Need for LoRA Training?

The right number depends on what you're teaching the model. A character or product LoRA — a specific face, mascot, or item — needs far fewer images than a style LoRA, which has to cover a much wider range of subjects while keeping one consistent look.

LoRA Type	Image Count	Why
Character or product	15–30 images	A single consistent subject needs coverage, not volume — past ~50 images you hit diminishing returns
Style	50–150 images	The model needs to see your style applied across many different subjects to learn the style itself, not one specific image

Tip: If you're not sure which category your LoRA falls into, ask: "Am I teaching one specific thing, or a look that applies to many things?" One specific thing (a face, a product, a mascot) is a character LoRA. A look (an art style, a color grade, a rendering technique) is a style LoRA.

Going over these numbers isn't harmful by itself — the real risk is what usually comes with a bigger dataset, which is more near-duplicate images. That's covered in the variety section below, and it matters more than the count on its own.

What Resolution Should Your Training Images Be?

Every base model has a native resolution— this is the image size it was originally trained at, and it's the resolution it learns best from. Training on images far below that resolution, or on upscaled images pretending to be high-resolution, teaches the model blur and softness along with whatever you actually wanted it to learn.

Model	Recommended Resolution	Note
Krea 2	1024×1024	Skip 768px — community testing has reported inconsistent results at that size on Krea 2 specifically
Ideogram 4	1024×1024	Matches its native training resolution
Flux 2	1024×1024	Matches its native training resolution

Warning: Don't take a small image and upscale it to hit the target resolution. An upscaled 512px image doesn't contain any more real detail than it started with — it just looks like it does at a glance. Your trainer will treat it as a full-resolution image and learn the soft, upscaled detail as if it were real.

A correctly sized 1024x1024 training image next to a blurry upscaled 512px image — Left: a genuine 1024×1024 source image. Right: a 512px image upscaled to look the same size — noticeably softer up close.

Most trainers, including AI-Toolkit, don't require every image to be the exact same pixel dimensions — they bucket different aspect ratios automatically during training. What matters is that each image is genuinely close to your target resolution, not stretched or upscaled to fake it.

Why Variety Matters More Than Image Count

This is the single biggest reason LoRAs turn out bad: the dataset has enough images, but they're all nearly the same shot. Twenty photos from one phone session, same lighting, same angle, same expression, gives the model far less to learn from than twelve genuinely different images — because the model isn't learning your subject, it's learning the one pose you happened to photograph twenty times.

Coverage means your dataset includes different framing and different angles of the same subject, so the model learns what stays consistent (the actual subject) instead of what happens to repeat (one specific pose or lighting setup). Aim to cover:

Framing: close-up, medium shot, and full-body
Angle: front-facing, side profile, and three-quarter view
Lighting and setting: more than one location or lighting condition, where possible

Grid comparison of 12 varied-angle training photos versus 20 near-duplicate photos from one session — 12 images with real angle and framing coverage on the left outperform 20 near-duplicate shots on the right.

Tip: If you're short on real source photos, this is where generating variation with an existing consistency workflow helps — turning one or two reference photos into a wider set of angles and poses before you start training. That process gets its own full guide, since it deserves more depth than a quick tip here.

What Makes a Training Image "Bad" — Even If the Subject Looks Fine?

Some images quietly wreck a LoRA even though the subject in them looks perfectly good. These issues are easy to miss when you're scrolling through your own photos, because you're looking at the subject, not the technical quality of the shot.

A training image annotated with arrows pointing out blur, a watermark, and compression artifacts — Common issues that don't show up until you look closely: motion blur, a watermark, and JPEG compression artifacts.

Remove any image with:

Blur — motion blur or out-of-focus shots, even slight ones
Heavy compression — visible blocky artifacts, usually from screenshots or re-saved JPEGs
Watermarks or text overlays — the model will learn to reproduce them
Inconsistent lighting extremes — one wildly overexposed or underexposed shot pulls the whole set off balance
Other people or distracting objects in frame — the model can't tell what it's supposed to be learning

Warning: Don't keep a flattering photo just because it's a good picture. If it's blurry or low-resolution, it will hurt your LoRA more than it helps — a technically clean but less exciting photo is the better training image every time.

How to Organize Your Dataset Folder Before Training

Once your images are selected, cleaned up, and resized, your trainer needs to be able to read them correctly. Most trainers, including AI-Toolkit, expect each image to sit next to a matching caption file with the same filename.

my_dataset/
├── image_01.jpg
├── image_01.txt      ← caption for image_01.jpg, same filename
├── image_02.jpg
├── image_02.txt
├── image_03.jpg
└── image_03.txt

A dataset folder showing correctly named image and caption file pairs — Each image and its caption file share the same filename, only the extension differs.

Warning: Mixed file formats in one folder — some .jpg, some .png, some .webp — usually work fine, but mismatched filenames between an image and its caption file will make the trainer either skip that image or pair it with the wrong caption. Double-check filenames match exactly before you start training.

What actually goes inside each caption file — and how to write captions that help rather than confuse your trainer — is covered in a dedicated guide, since captioning has its own set of rules worth going through properly.

Common Dataset Mistakes That Ruin a LoRA

My LoRA only reproduces one pose or angle

What causes it: Low variety in the dataset — most or all of your training images share the same pose, angle, or framing.

How to fix it: Add images covering different angles and framing (see the variety section above), then retrain. There's no settings fix for this — the dataset has to change.

My LoRA looks blurry or low-detail

What causes it: Low-resolution source images, or images upscaled to look higher resolution than they actually are.

How to fix it: Replace any upscaled or sub-1024px images with genuine full-resolution source photos, matched to your base model's native resolution.

My LoRA is "overcooked" — faces look distorted, colors are wrong

What causes it: Too many near-duplicate images, or too few images repeated too many times during training.

How to fix it: Prune duplicate or near-identical images from your dataset and prioritize variety over raw count — a smaller, more varied dataset almost always fixes this.

Frequently Asked Questions

Yes, for a simple concept with a consistent look, 10-15 clean images can work. You lose flexibility across poses and scenes compared to a larger, well-covered dataset, so expect the LoRA to work well only in situations similar to your training images.

No. Most modern trainers, including AI-Toolkit, bucket images of different aspect ratios automatically during training. What matters more is that each image is close to your target resolution and not upscaled from something smaller.

Yes, as long as the base model can already generate the general category your subject belongs to. If you’re training a specific character the model has never seen, mix in images the model can generate around that character — similar poses, expressions, and settings — rather than relying only on images of the character itself.

A large dataset isn’t harmful by itself, but a large dataset with low variety usually is. Fifty near-identical images teach the model to memorize your training photos instead of learning the underlying concept, which shows up as an overcooked or inflexible LoRA.

Most trainers handle cropping and aspect-ratio bucketing automatically, so you don’t have to crop everything by hand. It still helps to remove obvious distractions from the frame yourself — other people, cluttered backgrounds, or text overlays — since automatic cropping won’t fix those for you.

Yes. Video LoRA training uses clips instead of stills, and the count is usually measured in seconds of footage rather than image count, with its own rules around frame consistency. This guide covers image LoRAs only — video LoRA dataset prep is covered in a separate guide.

What to Do Next

Sort your existing photos against this checklist before you gather anything new.

Most people already have more usable images than they think — the real work is cutting the ones that don't belong. Once your dataset is sorted, captioning is the next step.

ComfyUI Roadmap →

Published: 2026-08-15 · Last updated: 2026-08-15

Discussion

Join the discussion

💬

No comments yet

Be the first to share your thoughts!