Swin2SR Upscaler (x2 and x4) — Custom Trained
Sharper skin. Better textures. Trained on AI-generated photorealistic images.
Swin2SR Upscaler is a series of custom-trained super-resolution models based on the Swin2SR architecture. Unlike generic upscalers that smooth out fine detail, these were trained specifically on high-resolution photorealistic AI-generated images to preserve and enhance skin textures, hair detail, and fabric weave.
Three variants are provided — all in safetensors format for easy integration with ComfyUI and other pipelines.
Gallery
 |
Available Versions
| File | Scale | Training Data | Size |
|---|---|---|---|
swin2sr-custom-x2.safetensors |
2x | Custom photorealistic dataset | 58 MB |
swin2sr-custom-x4.safetensors |
4x | Custom photorealistic dataset | 58 MB |
swin2sr-div2k-x2.safetensors |
2x | DIV2K + Custom blend | 58 MB |
Usage
These are standalone Swin2SR models, not LoRAs. Use them in any pipeline that supports Swin2SR upscale models.
In ComfyUI
Load via the UpscaleModelLoader node and connect to ImageUpscaleWithModel.
Tips
- x2 custom is the best general-purpose choice — sharp detail without artifacts
- x4 custom for maximum upscale in one pass, though quality is slightly softer
- div2k + custom x2 blends academic training data with custom images — may produce different texture characteristics
- Best for photorealistic images — these models were optimized for skin, hair, and fabric
Links
About
Created by Thalis AI.
License
Released under the Apache 2.0 license.