Feel free to post your images!!!
Maintains Dynamic Range: It keeps the intense, glowing highlights vivid and high-contrast instead of compressing them into a muddy middle-gray.
Preserves Fine Geometry: The intricate details and the sharp edges of the architectural trim stay crisp without introducing jagged staircase artifacts.
Reflections: The surface reflections retain deep, rich darks and sharp light accents that give the water a convincing wet, reflective sheen.
The model is finding a brilliant balance: it's keeping the high-frequency contrast and vibrant lighting intact where standard models either blow them out or flatten them out entirely.
How It Was Made Architecture:
Built as a 4x DAT (Dense Arbitrary Transformer) model.
Dense Arbitrary Transformer (DAT) Backbone: Built on the DAT network architecture, which uses spatial and channel self-attention mechanisms to aggregate local and global features. This allows it to understand complex image context better than traditional convolution-only models (like standard ESRGAN).
Framework: Trained using the NeoSR training framework.
NeoSR Framework: Built and configured within the NeoSR training framework, giving you access to modern loss functions, optimized PyTorch backends, and precise control over execution pipelines.
Degradation Pipeline: Configured with a custom On-The-Fly (OTF) degradation pipeline using tailored kernel settings to simulate realistic downscaling, noise, and compression artifacts during the training process.
On-The-Fly (OTF) Degradation: Utilized an OTF pipeline during the training phase. Instead of static degraded image pairs, the data was dynamically blurred, downscaled, and injected with subtle noise on the fly, forcing the transformer weights to learn robust restoration patterns rather than memorizing fixed compression artifacts.
Custom Dataset: Trained on a curated personal dataset specifically assembled to prioritize smooth gradient handling, natural shadow transitions, and clean character/background separation rather than relying on standard public upscaler training sets.
Zero Fine-Tuning Needed: Because the base architectural training run hit the targeted structural stability out-of-the-box, no secondary fine-tuning or model refining passes were required to lock in the weights.
230 epochs
813 images - Mixed dataset - Hybrid Training
Hybrid Domain Coverage: Combining clean high-resolution targets with varied real-world/synthetic image types gave the 4x DAT transformer a broad understanding of both sharp character geometry and complex, soft background lighting.
Natural Dynamic Range: The mix ensured that smooth gradients and deep shadows were preserved instead of being crushed by overly aggressive artificial sharpening (like what often happens with datasets made entirely of ultra-crisp, high-contrast images).