Aixio-AI / aixio-layer-v2

# Aixio Layer v2 — Image-to-Layer AI Model
 **Flat design in. Editable PowerPoint out.**
 On a 150-design image-to-layer benchmark, Aixio Layer v2 recovered **8,096 valid editable layers**: **1.7×** the next best system and **2.1×** Canva Magic Layers, in about **47 seconds per image**.
 Try in Aixio Studio · Watch the video · Benchmark · Samples · What you get back · Layer v1
---
## See it in action
 Introducing Aixio Layer v2 — click to watch on YouTube 
## What's new in v2
Layer v2 returns every part of a design in the form you would have built it in:
- **Editable text**: copy comes back as real text boxes you can retype, with the matched font embedded in the file.
- **Native shapes and lines**: badges, circles, rules and dividers come back as PowerPoint shapes you can recolour.
- **Clean image layers**: photos, products, illustrations and lettering each come back on their own layer.
- **A clean background**: the page background comes back as its own layer.
- **One native PowerPoint file**, in about 47 seconds per image.
## Benchmark
| Model | Valid image layers | Valid text layers | Valid vector layers | **Total valid layers** | Seconds per image |
| --- | ---: | ---: | ---: | ---: | ---: |
| **Aixio Layer v2** | **2,150** | **3,713** | 2,233 | **8,096** | **47** |
| GPT-6 Astra Ultra | 14 | 2,338 | **2,345** | 4,697 | 283 |
| Canva Magic Layers | 1,884 | 1,987 | 0 | 3,871 | – |
| Qwen-Image-Layer | 1,056 | 0 | 0 | 1,056 | 366 |
- **Most valid editable layers overall:** 8,096, against 4,697 for the next best system.
- **Most valid text and image layers** of any system tested.
- **Best reconstruction score** (LPIPS 0.1382), narrowly ahead of Canva Magic Layers (0.1414).
- GPT-6 Astra Ultra recovers slightly more valid vector layers (2,345 vs 2,233).
### What each system returns
| | Image layers | Editable text | Native shapes and lines |
| --- | :---: | :---: | :---: |
| **Aixio Layer v2** | **Yes** | **Yes** | **Yes** |
| GPT-6 Astra Ultra | Yes | Yes | Yes |
| Canva Magic Layers | Yes | Yes | No |
| Qwen-Image-Layer | Yes (16 full-canvas layers) | No | No |
### How we measured
All four systems received the same 150 designs: posters, slides, infographics,
social graphics, product pages and interfaces, both human-designed and
AI-generated. For each design we counted the editable elements that came back
usable:
- an **image layer** counts when it holds real content, isn't duplicated, and has no obvious artifacts;
- a **text layer** counts when it matches real text and isn't left baked into an image;
- a **vector layer** counts when it matches a real shape or line in appearance, colour and size.
Validity was judged by an evaluation skill built on GPT-6 Astra; GPT-6 Astra
Ultra is also one of the systems compared. Seconds per image is the reported
mean; Canva Magic Layers runs as a hosted product and was not timed. This is an
Aixio-reported benchmark.
## Samples
Each panel shows every image layer a system returned for the same design, laid
out at