Hasty Briefsbeta

Qwen-Image-Layered: transparency and layer aware open diffusion model

2 days ago
  • #image editing
  • #diffusion models
  • #image decomposition
  • Qwen-Image-Layered decomposes images into semantically disentangled RGBA layers using a diffusion model.
  • Enables independent editing of each layer, improving decomposition quality and consistency.
  • Addresses the challenge of consistency in image editing faced by recent visual generative models.
  • Introduces three key components: RGBA-VAE, VLD-MMDiT architecture, and Multi-stage Training strategy.
  • Builds a pipeline to extract and annotate multilayer images from Photoshop documents (PSD) due to scarcity of training data.
  • Demonstrates superior performance in decomposition quality and establishes a new paradigm for consistent image editing.
  • Code and models are released on GitHub.