DTPSR: Disentangled Textual Priors for Diffusion-based Image Super-Resolution

March 22, 2026 · View on GitHub

Official implementation of DTPSR (Disentangled Textual Priors for Diffusion-based Image Super-Resolution), a diffusion-based SR framework that injects disentangled textual priors along:

  • Spatial hierarchy: global vs. local
  • Frequency semantics: low-frequency (structure/layout/color) vs. high-frequency (texture/edges/details)

DTPSR also introduces multi-branch classifier-free guidance (CFG) with frequency-aware negative prompts, and releases DisText-SR, a structured dataset with ~95k image–text groups.
(See paper for full details.) :

🔥 Highlights

  • Disentangled textual priors: global caption + per-segment LF/HF descriptions, enabling controllable, interpretable guidance.
  • Progressive semantic injection via specialized cross-attention modules:
    • GTCA (Global Text Cross-Attention)
    • LFCA (Local Low-Frequency Cross-Attention)
    • HFCA (Local High-Frequency Cross-Attention)
    • LRCA (LR Feature Cross-Attention for identity anchoring) :
  • DisText-SR dataset (~95k): global + region-wise LF/HF descriptions generated using panoptic segmentation + MLLM prompting.

🧠 Method Overview

Given an LR image, DTPSR:

  1. Encodes LR to latent space (VAE encoder)
  2. Runs diffusion denoising with sequential semantic guidance:
    • Global prior guides scene-level layout (GTCA)
    • Local LF priors refine object-level structure (LFCA)
    • Local HF priors enhance textures and edges (HFCA)
  3. Fuses LR visual features through LRCA to maintain image identity and consistency

📦 DisText-SR Dataset

DisText-SR contains ~95,000 groups of structured annotations:

  • Global description (c_g)
  • For each segment S_i (from panoptic segmentation):
    • Low-frequency text (c_lf^(i)): shape/size/colors/orientation (no fine details)
    • High-frequency text (c_hf^(i)): texture/material/edges/subtle details

Texts are encoded by a frozen CLIP text encoder to form embeddings for guidance.

TODO: Provide dataset download link / license / checksum.