🎮 Smart USDU DiffDiff + ControlNet

The one USDU variant that finally looks like a proper upscaler pipeline. It takes the differential-diffusion tile upscaler and bolts on a ControlNet whose control image gets cropped per tile, so each tile's sampler sees the correct region of your canny/depth map instead of a wrong, full-canvas condition. If you've ever tried "Tiled Diffusion + ControlNet Tile" - the community's long-standing favorite recipe for photorealistic upscaling, per the upscaling knowledge base - this is that, packaged as one node: tiling, per-pixel denoise, and a spatial condition that keeps each tile faithful to the source.

That faithfulness is the whole reason to add ControlNet to an upscale. USDU alone can drift - each tile is an independent img2img pass, and without a condition, tiles start inventing their own textures. A canny edge map or depth pass says "the structure goes here," so the detail-filling happens along the source rather than replacing it. You can push denoise higher with a condition holding the structure.

The inputs that matter

• control_image (optional IMAGE) - your condition (canny, depth, tile preprocessor output). It's upscaled and cropped per tile to match the main image's tiling. Disconnect it and ControlNet is off.
• model_patch (optional MODEL_PATCH) - the ControlNet patch itself, from a loader like ModelPatchLoader. Both this and control_image are required for ControlNet to engage.
• control_strength (default 1.0) - standard ControlNet weight. Start at 1.0, drop to 0.6–0.8 if the condition is fighting the detail pass.
• control_mask (optional MASK) - lets the condition apply only where you want it, separate from the diffdiff denoise mask.
• denoise_mask + multiplier - the differential-diffusion pair from the base node (white = more denoise, black = less; multiplier <1 = stronger).

Everything else is the USDU core: upscale_by, required upscale_model, conditionings list, mode_type, tile size/padding/blur, seam fix, tiled_decode. Output is a single IMAGE.

Installing it

One of 100+ nodes in ComfyUI-ArchAi3d-Qwen by Amir Ferdos (ArchAi3d). ComfyUI Manager → search "ArchAi3d Qwen" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git
cd ComfyUI-ArchAi3d-Qwen
pip install -r requirements.txt

Then restart. The node itself needs nothing beyond ComfyUI's torch/numpy/Pillow; the heavyweight deps in requirements.txt belong to other pack nodes. License: free personal/non-commercial, paid license for commercial work.

The traps

The per-tile cropping only helps if the control image and the main image are the same size and framing - a control image from a different crop silently misaligns and you get wrong-conditioned tiles. ControlNet here is per-tile, so it isn't a cheap add: every tile runs a controlnet pass, and a 4×4 grid means 16 of them, plus seam-fix passes on top. If you're on limited VRAM, lower tile_width/tile_height rather than removing the condition. And remember the two-part enable: model_patch without control_image (or vice versa) quietly disables ControlNet and you're back to plain DiffDiff.