What we're working on

Photoroom's goal is to deliver the AI image editing solution that offers the strongest possible level of product fidelity.

To achieve this goal, we have a dedicated in-house Machine Learning team that actively conducts research to determine what type of AI models are able to edit images with minimal risk of hallucinations.

Right now this team is exploring a new idea: training a model that operates directly in pixel space.

What does pixel space mean?

Current state-of-the-art AI image editing models don't operate directly on the pixels of the image.

Instead they operate on a learned representation of the image that lives in what is called a latent space.

It's a fairly technical topic but here's the gist of it:

First, a separate model compresses your image into a much smaller grid of numbers (called tokens). Think of it like a zip file — smaller, but with all the important information preserved.

The editing model then uses that compressed version as a reference when performing the editing.

Finally, another model decompresses it back into a full image.

But here's the important tradeoff: that compressed format is an interpretation of your image, not a perfect copy.

When the model edits it and decompresses it back, it's essentially reconstructing the image from an approximation, which means it can sometimes fill in details that weren't there originally.

This is where hallucinations come from: the model confuses what was in the image with what it expects to be there.

To avoid this issue, the team at Machine Learning team at Photoroom is currently working on training a model that operates directly in the pixel space, without using any intermediary representation of the image.

If you're interested to learn more about this process, here's an article that explains how the training works:

PRX Part 3 — Training a Text-to-Image Model in 24h!huggingface

You can also try an early version of the model here and even download its weights here.