Semantic Prism: A New Framework for Generative Semantic Segmentation
Researchers have introduced a novel framework named Semantic Prism for generative semantic segmentation, aiming to resolve issues like color drift and boundary mixing in direct color decoding. This framework is detailed in a paper on arXiv (2608.11537) and employs a conditional method for generating and refining semantic images with deterministic inference. It includes a one-step generator that creates a semantic RGB image via diffusion distillation. The approach also establishes a clear probabilistic interface using per-pixel distances to a fixed class-color codebook. Moreover, it organizes multi-level generator features spatially with Hierarchical Generator Evidence Alignment and uses a zero-initialized output projection for predicting an additive residual in the interface logit space, relying on the image-defined interface for stability.
Key facts
- Semantic Prism is a conditional semantic-image generation-and-refinement framework with deterministic inference.
- It uses a diffusion-distilled one-step generator to render a semantic RGB image.
- Per-pixel distances from rendered colors to a fixed class-color codebook define an explicit probabilistic interface.
- Hierarchical Generator Evidence Alignment spatially aligns multi-level generator features.
- A zero-initialized output projection predicts an additive residual in the interface logit space.
- The framework retains the image-defined interface as the reference for the final distribution.
- The paper is available on arXiv under identifier 2608.11537.
- The announcement type is cross.
Entities
Institutions
- arXiv