Prompt-free Tiled Upscaling with Krea 2: Slicing Vision Tokens for 4K Upscales

blakeem · reddit · 2026-08-06

The author shared a new image upscaling workflow using the Krea 2 model and open-sourced the ComfyUI nodes.

Pain Points of Traditional Tiled Upscaling

During standard tiled upscaling, since the prompt describes the whole image, a strongly prompt-adherent model tries to recreate the entire prompt inside each tile, causing artifacts or duplicated objects near edges.

Innovative Solution

Krea 2 uses Qwen3-VL, a vision-language model, as its text encoder. The author runs a single vision encode of the full image to produce a grid of vision tokens with spatial position information.
For each tile, only the tokens covering that tile's area are sliced out to replace the positive prompt. No text prompts are used at all.
This method tells each tile exactly what it contains. Since all tiles share the same whole-image encode, they agree on tone, palette, and structures crossing seams, eliminating duplicated objects and drift between tiles.

Results & Open Source

At a denoise strength of 0.42-0.5, this method can upscale images past 4K in a single pass with highly detailed and coherent results.
The relevant nodes are free and open-source on GitHub. Theoretically, this method can be adapted to any model with a VLM encoder.

More from coding & agent