Adaptive Beta Sampling for Diffusion Models
This repository contains the trained components for Method 3, an adaptive timestep selection approach for diffusion models based on:
- Semantic embeddings (SentenceTransformer)
- Frequency-domain features (FFT)
- MLP-based region classification
Components
model.pt: trained MLP classifierscaler.joblib: feature normalizationlabel_map.json: class mappingconfig.json: inference configuration
Method Overview
- Run short DDIM probe
- Extract FFT-based features
- Combine with text embedding
- Predict region (structure / balanced / detail)
- Select Beta schedule
- Generate final image
Usage
See the accompanying notebook for full inference pipeline.
