ComfyUI_KleinTiledUpscaler Debuts Seamless Upscaling For Flux2 Klein

    
        By vramkickedin    
     | 
    
            May 31, 2026 at 4:20 pm        
    
     | 
    
        2 min read

ComfyUI_KleinTiledUpscaler is a custom node for ComfyUI that performs creative image upscaling using a tiled, inpainting-based approach designed specifically for the Flux2.Klein model. Rather than just sharpening existing pixels, the tool lets the AI generate new high-frequency details like textures and grain within each tile. It processes images tile-by-tile to keep video memory usage low while producing large, detailed outputs.

Developer Gavr728 created this node after noticing that existing tile upscaling workflows for Flux2.Klein either failed to use reference latents properly or produced visible grid seams in results. The entire project was "vibecoded," meaning Gavr728 built it by prompting large language models without personal coding knowledge. The node was released freely for hobbyist and non-commercial use.

Smart tiling that prevents seams

Key Features

Auto slices canvas into perfectly equal tiles
Built-in color matching across tile boundaries
Adaptive tiling saves steps on flat areas
Dynamic RoPE alignment for Flux2.Klein coordinates
Detail-first strategy processes textured zones first
VRAM-friendly sequential tile-by-tile processing
Full LoRA support for upscaling and style
Laplacian detail analysis avoids smooth area artifacts

This node fits well for hobbyists and small studios running local AI on consumer GPUs who need creative upscales with added detail. The tile-by-tile architecture means you can upscale massive images even on hardware with limited VRAM. Privacy-conscious professionals working offline will appreciate that everything runs locally without cloud dependency.

Prompt sensitivity and testing notes

The developer warns that complex, highly descriptive prompts can cause objects to hallucinate in wrong tiles, so simple prompts like "upscale this image" work best. Most testing used a configuration of 4 steps, Euler sampler, and 2x upscale at 1024 tile size, and users straying far from these defaults may encounter unexpected behaviors. The node also includes a tiled decode option for extreme resolutions above 8K to prevent memory crashes, though it might increase color differences between tiles.