image-refinement

TBG Enhanced Tiled Upscaler and Refiner upscales up to 200MP with precise control. It features dual-model processing (structure + detail), per-pixel denoising to control where creativity is applied, and per-tile and per-object prompts, denoise, and seeds. Advanced fast previews. New generative tile fusion seamless joining, even on high-denoise.

Seamless tiled refinement and upscaling for ComfyUI. The VL nodes step all tiles on one shared latent, conditioned by vision tokens and per-tile captions instead of a prompt, so objects never repeat across tiles. Refine whole images or just a masked region. Any GUIDER model on the base node; Krea 2 on the VL nodes. Tested at 8K+.

GPU-accelerated iterative bilateral solver for edge-aware refinement.

Image Captioning Refinement Using Deep Learning and computer-vision — Flickr8k dataset with BLIP, Jina, Qwen and Majority Voting