Your style. In every frame.
Redraft makes AI work for you.
Redraft trains custom models that actually help scale your creative work instead of giving you generic, unusable output.
Click the image to see how Redraft restyles this scene.
Objects stay where
you placed them.
Custom models that preserve artistic intent
To build your custom model, Redraft will use between 10 and 50 images in your desired style (usually provided by you). The end result will be a model that looks on the surface like any other image model. You can input images and prompts and expect an image as an output. The only difference? It'll actually follow your instructions, because it'll be custom tuned to you.
Our model is the leader in style preservation
Each model received the same base image and style reference. The task was to redraw the base image in the reference style.
Thin, overlapping lines
Keeps the facial expression
Keeps thin, layered lines
Softens the expression
Loses fine line detail
Thicker, darker outlines
More saturated color
Each model recreated the starting image five times, using its previous output as the next input. These are the results after the fifth iteration.
Keeps the face and linework
The jeans stay blue
Loses line definition and complexity
Curves become flat shapes
Changes the facial structure
Jeans turn purple; saturation increases
50× more data.
From your references.
Our synthetic data pipeline turns a small set of references into hundreds of examples for training and evaluation. We generate the data, grade it, and train your custom model.
Measured on our benchmark dataset
More examples of subjects, scenes and style errors, without collecting each one by hand.
What that could mean
for your model.
Projection10–50 references could become 500–2,500 synthetic images at 50× expansion. This is a projection; the final dataset depends on the references and quality checks.
Style match,
before and after.
49 images · same grader
JEV 1.14.0 · 0–4 scale
Average style-match score against the reference, before and after Redraft processed the same images.
Improved style match
Every output scored higher than its input in this test.
Strong style matches
Outputs scored at least 3 out of 4 on the style-match rubric.
Grader agreement with human reviewers
Across 200 comparisons.
Internal 49-image evaluation of rendering style.
Provisional patent
applications filed.
Covering synthetic data generation, training-example construction, and dataset sampling.
We handle the data pipeline, training and workflow integration. Your subscription is priced for your model and requirements.
Discuss your modelWhat these figures include
Data expansion. Three original images plus 152 synthetic images made a corpus of 155 unique images. New synthetic images divided by source images gives 50.7×, rounded down to 50× here. The corpus included 95 training inputs and 44 inputs reserved for evaluation. The reference-count calculator extrapolates the rounded multiplier; it is not a measured result for each dataset size.
Style evaluation. The chart uses 49 matched input/output pairs from the 5,000-update checkpoint, assessed with the same JEV 1.14.0 style-match rubric and reference. This rubric rates rendering style, not subject identity or composition. Scores are ordinal ratings, not percentages of accuracy. This is an input-to-output comparison, not a comparison against another model.
Grader agreement. The grader agreed with human reviewers in 98% of 200 comparisons.
Pricing. $50/month is the starting subscription price. The proposal depends on your references, intended use, training needs and delivery requirements.
How our model works
Read the images
Turns the input and reference images into numbers the model can process.
Keep the scene
Uses information about the input’s subjects and layout to guide the restyle.
Apply the style
Trained transformerThe model uses diffusion to adjust the image towards the reference style without changing the underlying entities.
- Style
- Structure
- Palette
Build the image
Turns the processed numbers back into visible pixels.
Frequently asked questions.
01What does Redraft actually make?
We build small custom image models trained around a client’s desired visual style. You bring the aesthetic and reference material; we work with you to train and refine a model to help scale up your creative work.
02Who is it for?
Creative teams, studios, brands, and artists with a distinct visual direction that they want preserved. A useful starting point is a style you want to explore across more subjects, scenes, or projects.
03What should we bring to the first conversation?
A few representative images, a description of what you want to create, and anything that helps explain your taste. Rough references are welcome. We can discuss what a suitable training collection would look like together.
04How many reference images do we need?
There isn’t one number that fits every visual style, but expect somewhere between 10 on the lower end and 50 for robust models. The range of subjects, the consistency of the references, and the intended use all matter. We’ll assess the material with you before agreeing on the scope.
05Can we review the model during the process?
Of course! You can review the model at any stage, and if you don't like what it outputs at the end, we'll work together to correct it.
06What do we receive at the end?
You receive a custom model that we'll help directly integrate into your workflow. We won't just give you an API key and let you figure it out! We'll design a system that fits right in to what you're currently doing.
07Will the model work with any image tool?
Yes, the model works with any image tool. Our underlying base models that we train off may differ, but you can always use the outputs of other image models as inputs for ours.
08How much does a custom model cost?
Redraft is a subscription based service, and pricing is scoped by model. Your references, intended use, training needs, and delivery requirements inform the proposal. A sales conversation is the first step toward a quote, but costs can be as low as 50 dollars a month.
09How long does a project take?
For most models, we can deliver in under 24 hours. But even for large, robust datasets, we can guarantee delivery in less than a week.
10What about commercial use and ownership?
You have the full commercial rights to everything the model outputs, and usage of the model itself! They all belong to you, and we claim no ownership of it. You can request a data deletion at any time.
