Stop Overhyping AI Image Generation: AI That Can't Edit Layers Is Just a 'Painted Cake'
Toozon (Tuzhan Intelligent)
RabbitVis, a new AI design production tool from Toozon, is launched, aiming to move beyond simple AI image generation by supporting layer-based editing. Based on their UniWorld-Design model, it enables structured generation and layer-level editing, positioning AI as a tool that completes the entire design workflow.
Toozon (Tuzhan Intelligent) released RabbitVis, an AI design production tool built on their UniWorld-Design model, which integrates visual generation and editing capabilities. Unlike traditional AI image generators that produce a single final image, RabbitVis supports layer decomposition, allowing text, assets, backgrounds, and decorative elements to be edited as independent layers. This enables users to make adjustments such as repositioning logos, resizing text, or creating e-commerce versions without regenerating from scratch. The tool is based on UniWorld-Design, which was developed by Toozon with Peking University and Pengcheng Laboratory, and it features transparent material generation and image layering capabilities. In evaluations, UniWorld-Design achieved a CLIP Score of 33.03 in T2RGBA transparent material generation, and in I2L image layering tests, it scored 0.1264 in RGB L1, 0.7325 in Alpha Soft IoU, and 1.32 in editability, with a VLM subjective score of 20.43 out of 25. RabbitVis aims to transform AI design from one-time generation to continuous production, allowing design outputs to become reusable design assets, thereby shifting the competitive focus from raw model capability to practical workflow outcomes.
- Abbreviations
- I2L = Image-to-Layer — изображение-в-слои
- VLM = Vision-Language Model — мультимодальная модель
- T2RGBA = Text-to-RGBA (transparent image generation) — генерация прозрачных изображений по тексту
- CLIP = Contrastive Language-Image Pre-training — контрастивное предобучение языка и изображений
Source: QbitAI 量子位 —
original
