TTPlanetPig/Comfyui_TTP_Toolset

What it solves

ComfyUI TTP Toolset addresses the limitations of traditional fixed-grid tiled upscaling in image generation. Standard grids often split critical visual elements (like faces, hands, or text) in half and apply uniform detail levels across the entire image, regardless of whether a region is a complex focal point or a simple background.

How it works

The core of the project is Smart Tile 2.0, an object-aware tiled img2img workflow. Instead of equal squares, it creates variable-size tiles based on visual structure.

1. Tile Creation: Users can manually define tiles, paint masks, or use automatic detection via SAM3.1 or QwenVL3 to identify specific objects (faces, eyes, hands, etc.).
2. Loop Processing: A specialized loop system (Loop Source and Loop Collect) sends one tile at a time through the img2img sampler. This allows for per-tile prompts (generated automatically via QwenVL) and individual upscaling of each tile before sampling to manage VRAM via megapixel caps.
3. Intelligent Assembly: Processed tiles are pasted back together using a semantic ranking system. This ensures that high-detail focus tiles (like eyes) are layered on top of larger context tiles (like a face or background) with feathering, color correction, and mask-aware blending to avoid seams.

Who it’s for

Digital artists and AI image creators using ComfyUI who need high-resolution upscaling with precise control over detail density and object integrity, particularly for complex portraits or detailed scenes.

Highlights

• Object-Aware Tiling: Variable-size tiles that wrap around visual structures rather than cutting through them.
• Automated Prompting: Integration with QwenVL to generate specific prompts for each individual tile.
• Semantic Layering: A priority system that automatically places detail-heavy tiles (faces, text) above background tiles during assembly.
• Efficient Loop Workflow: Processes tiles sequentially to avoid duplicating sampler chains in the ComfyUI graph.
• Flexible Assembly: Supports GPU-accelerated pasting, histogram color correction, and multiple composite modes (e.g., safe_auto or replace_tile).