☁️BizyAir Tile

Tile is the ControlNet condition behind most detail-preserving upscale workflows, and it works on an idea that sounds almost too simple to be useful: feed the model its own source image back, softened and at lower resolution, and let it add detail while staying faithful to what's already there. That's the whole mechanism - no edge detection, no depth estimation, just the original image itself as the condition.

BizyAir's Tile preprocessor is the softening/prep step before that condition gets used. It doesn't add detail itself; it prepares the image so a Tile-trained ControlNet model has something appropriately blurred to reconstruct from, rather than an image so sharp there's nothing left for the model to contribute.

Inputs and outputs

Only image is required. Two optional controls shape how the preprocessing behaves:

• pyrUp_iters - an integer from 1 to 10, defaulting to 3. This is the classic tile-preprocessor knob: it controls how many passes of blur/upscale get applied before feeding the image forward, which determines how much fine detail gets erased for the model to reinvent. More iterations means a softer starting point and more room for the model to add new detail (useful for a stronger upscale pass); fewer means the output stays closer to the original.
• resolution - 64 to 1024 in steps of 64, default 512, the working resolution for the preprocessing pass.

Output is a single IMAGE - the prepared tile condition, ready for a ControlNet Apply node running a Tile-trained model.

Worth knowing before you build a workflow around this: Tile coverage isn't guaranteed on every base model's union. Shakker Labs' widely-used FLUX.1 Union Pro 2.0 actually dropped Tile in its 2.0 revision, which was a real point of frustration for people who'd built upscale pipelines around the 1.0 version. If you're on Flux, check which union revision you're running before assuming Tile is available - Z-Image's union, by contrast, ships a dedicated Tile model trained up to 2048px.

Installing BizyAir

Through ComfyUI Manager: search "BizyAir," install. Or clone it directly:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

Restart ComfyUI afterward, or use comfy node install bizyair via Comfy-CLI. Set your API key before running anything - the README's "click to login" step is required on first use. This preprocessing runs on SiliconFlow's cloud rather than your local machine, consistent with the rest of the pack.

Common issues

If your tiled upscale results look identical to the input with no added detail, the ControlNet weight or the sampler's denoise is more likely the culprit than this node - Tile conditioning at very high control weight or very low denoise can lock the output too close to the source to add anything. Try loosening one of those before adjusting pyrUp_iters.

If results look over-smoothed or lose structure entirely, push pyrUp_iters down - you're erasing more of the original than your Tile ControlNet can meaningfully reconstruct.

And check that your base model actually has a Tile-capable union or model loaded before troubleshooting the preprocessor itself; a perfectly good tile condition with nothing downstream trained to use it will just produce a normal generation that ignores the conditioning.