AI Training Datasets
Photo Editing Datasets for AI Training, Edited by Hand
Before-and-after pairs edited by our retouchers, and images labelled by people for computer vision: boxes, polygons, masks and keypoints. Built only from images you own or license.
No minimum orderFree trialPay per imageEdited by hand, not by AI
Priced per pair or per labelled image, quoted after a pilot batch. Turnaround agreed per batch.
Data Privacy and Rights
Customer Photos Never Train a Model
Dataset work and client editing are kept completely apart. Here is exactly what we promise.
Customer photos never go into a dataset
Photos that photographers and brands send us for editing are never used for AI training and never put into any dataset, not ours and not a client's.
Only images with clear rights
We build datasets only from images you supply and own or license, or images with documented rights for training use. We keep a record of where every source image came from.
NDA, deletion and ownership
We sign an NDA on request. Files are permanently deleted 30 days after delivery, or sooner if you ask. The delivered dataset belongs to you.
Every pair edited by a person
No AI auto-editors and no generative fills. Each target image is the work of a trained editor, which is exactly what makes it worth training on.
How we handle every file: Security and data protection
What You Receive
What Is in a Dataset
A model learns an edit from the gap between the original and the result. We make that gap clean, consistent and documented.
Paired originals and edits
Each source image comes with its hand-edited result, matched one to one at the same size and framing.
Masks and segmentation
Optional masks that mark exactly which pixels changed, or which object, person or background region an edit applies to.
Labels and metadata
Each pair tagged by edit type, subject category and target style, in a schema we agree with you before work starts.
Written edit instructions
The instruction behind every edit, in plain language, for models that learn to follow a prompt as well as copy a result.
Edit Types
Edits We Build Datasets For
The same edits our team does every day for photographers and brands, applied to your source images. One edit type per dataset, or a mix.
Object removal
Unwanted objects, people, wires and clutter removed, with the background rebuilt by hand.
Background change
Backgrounds cut out, cleaned or replaced, with clean edges on hair, glass and fine detail.
Colour and exposure correction
White balance, exposure, contrast and colour cast corrected to a neutral, consistent result.
Skin retouching
Blemishes and distractions cleaned while skin texture stays real, at the retouching level you specify.
Enhancement and clean-up
Dust, noise, glare and small defects fixed, and detail sharpened without halos.
Style and look matching
A defined colour grade or brand look applied the same way across every image, for style-transfer models.
Multi-step edits
A chain of instructions applied in order, with each intermediate stage saved, for models that edit in turns.
Product image finishing
Marketplace-style product edits: cut-outs, shadows, colour matching and surface clean-up.
Data Annotation
Image Annotation for Computer Vision, Labelled by People
Alongside edited pairs, we label images for detection, segmentation and classification models. Each label is drawn by a person working to your class list, not by an auto-labelling model.
Bounding boxes
Tight rectangles around every object on your class list, for object detection models.
Polygons and instance segmentation
Outlines traced point by point around each object, so every instance gets its own shape.
Semantic segmentation masks
Every pixel assigned to a class, such as road, sky, person or product, and delivered as masks.
Image classification and tagging
One or more labels per image, from your category list or a set of tags we agree with you.
Keypoints and landmarks
Points placed on joints, facial landmarks or product corners, in the order your skeleton defines.
Review and correction of existing labels
Labels made by an auto-labelling model or another vendor, checked and fixed by hand: missed objects, loose boxes and wrong classes.
Where We Work
In Your Tools, or as Files
We fit into the setup your team already uses, so nothing has to move to a new platform.
Your Roboflow workspace
Invite our team as labelers and we annotate inside your project. The project, its classes and every dataset version stay in your workspace.
CVAT or Label Studio
The same arrangement in CVAT or Label Studio: add our labelers to your project, and the work stays where your team can see it.
Files in standard formats
Prefer files? We deliver annotations matched to your images by file name, in the format your training code reads.
How We Check Labels
Every label is drawn or checked by a person. A second person then reviews a sample from every batch for these points.
- Missed objects
- Loose or clipped boxes and outlines
- Wrong class
- Consistency with your labelling guide
We follow your class list and labelling guide. If you do not have a guide yet, we can help you write one during the pilot.
Roboflow, CVAT and Label Studio are products of their respective owners. We work in them on your projects; we are not affiliated with or endorsed by them.
Our Process
How We Build It
Six steps from your brief to a dataset you can load, with a pilot first so nothing is scaled before you have approved it.
01
Scope and pilot
We agree the edit types, style rules, label schema and file layout, then edit a small pilot batch so you can check it against your training needs.
02
Edit instructions per pair
Every source image gets a written instruction before anyone opens it, so the edit is defined, repeatable and recorded.
03
Editing by hand
Our editors carry out each instruction by hand in professional editing software. No AI auto-editors, no generative fills.
04
Quality control
A second person reviews every pair against the instruction and the style rules before it is accepted.
05
Labels, masks and metadata
Accepted pairs are labelled, masked where required and logged against their source record.
06
Structured delivery
Each batch arrives in the folder structure and naming you specified, ready to load into your training pipeline.
Quality Control
What We Check on Every Pair
A model copies mistakes as faithfully as it copies good work, so a second person reviews every pair before it goes into a batch.
- No visible artefacts: halos, smudges, repeated patterns or soft edges
- Tone and colour consistent across the whole batch
- White balance checked against the agreed reference
- Duplicate and near-duplicate source images flagged
- The edit does what the instruction says, and nothing more
- Pairs, masks and labels matched by file name with no gaps
Delivery Formats
Delivered Ready to Load
Paired image files
Originals and edits with consistent, matching file names, in the format and bit depth you need.
Masks as PNG
Binary or multi-class masks saved as PNG, named to match the pair they belong to.
Labels as JSON or CSV
Edit type, category, style, instruction text and source record for each pair, as JSON or CSV.
Your folder structure
Split into training, validation and test sets, or by edit type, as agreed with you before the pilot.
Who It Is For
Built for Teams Training Editing Models
Since 2019 our team of more than 100 editors has worked for photographers, studios and brands. Dataset clients get the same people and the same standards.
Retouching tool developers
Teams training models for portrait, skin and clean-up edits who need targets a professional would sign off.
Background removal models
Hand-cut masks and replaced backgrounds, with the hard edges (hair, fur, glass) done properly.
Colour grading and correction
Consistent before-and-after colour work for auto-correct, grading and style-transfer models.
E-commerce imaging tools
Product cut-outs, shadows and colour-accurate finishing for tools that prepare marketplace images.
Brands and studios
Your own images edited in your house style, as a dataset for an in-house model that should edit the way you do.
Priced after a pilot
Every project is different, so we price per pair for edited datasets and per image or label set for annotation, once the pilot batch shows how much work each one takes. You get a fixed quote before any volume work starts.
Turnaround per batch
We agree a delivery date for each batch based on its size and edit type, and deliver in batches so your team can start training before the whole set is finished.
Common Questions
AI training dataset questions
Do you use photos your editing clients send you?
No, never. Photos sent to us for editing are never used for AI training and never put into any dataset, not ours and not a client's. They are used only to complete the editing that was ordered, and are deleted 30 days after delivery.
Where do the source images come from?
From you. We work on images you supply and own or license, or on images with documented rights for training use. We keep a record of the source and the rights for every image in the dataset.
Who owns the finished dataset?
You do. The delivered pairs, masks and labels belong to you, and we do not reuse them for anyone else. We sign an NDA on request.
Do you use AI to make the edits?
No. Every pair is edited by hand by a person. We never use AI auto-editors or generative AI on the images, which is the point: a model trained on machine edits learns the machine's mistakes.
Can you label images in our Roboflow workspace?
Yes. Invite our team to your Roboflow workspace as labelers and we annotate there. The project, its classes and its dataset versions stay in your workspace, and you decide what we can access.
Which annotation formats do you deliver?
COCO JSON, YOLO TXT, Pascal VOC XML or CSV. If we work inside your Roboflow, CVAT or Label Studio project, you export the labels from there in whichever format you use.
Do you use auto-labelling?
No. Labels are drawn by people. If your project already has auto-generated labels, we can review them and correct them by hand.
How much does a dataset cost?
We price per pair for edited datasets, and per image or label set for annotation. After a pilot batch we know how long each item takes, and we give you a fixed quote before the main work starts.
How long does it take?
Turnaround is agreed per batch, based on the number of pairs and the edit type. Large datasets are delivered in batches so training can start early.
What formats do you deliver?
Paired image files with consistent naming, masks as PNG, and labels and metadata as JSON or CSV. The folder structure, file format and splits are agreed with you before the pilot.
What happens to the files after delivery?
Files are permanently deleted 30 days after delivery, or sooner if you ask. If you want us to keep them longer for follow-on batches, we keep them until you tell us to delete them.
Get Started
Plan Your Pilot Batch
Tell us the edit or annotation types, rough volume and the rights you hold on the source images, and we will reply within 24 hours.
AI Training Datasets
Dataset Details
Describe the model you are training and the edits or labels you need. Please do not send images until we have agreed the scope and rights.