nightknocker/u-slider-text-encoder-docling-258m

UNet

A lightweight UNet with single-block levels and sliding window attention.

• Pixel-space model in CIELAB color space
• LAB input, RGB output
• Decompose the input images into their frequency-domain components
• Docling as text encoder
• Token efficient visual text inputs
• Variable head in the attention modules across the layers

Retrospection

Reconstruction quality, from good to worst:

• U-Docling (this repo)
• U-DAE
• U-DAE-NLL
• EQ-SAE-CIELAB
• EQ-SAE-CIELAB-c8
• VAE-f16-c4-kv
• VAE-f16-c4
• VAE-f16-c8

References

• 2411.17459
• 2503.11576
• 2510.17800
• 2510.18279