☁️BizyAir Load VITMatte Model

Quick distinction that matters before you use this: segmentation and matting are not the same job. A segmentation model like SAM or BiRefNet's plain weights labels every pixel foreground or background - a hard yes/no. ViTMatte does matting: it predicts fractional alpha, so a strand of hair or the edge of a wine glass can come out 40% transparent instead of forced into "in" or "out." If your cutouts keep looking harsh around hair or anything semi-transparent, that's usually a segmentation model being asked to do a matting model's job - this node is the fix for that specific problem.

This is the loader half of a two-node pair. It just loads the model and hands you the pieces the actual prediction node needs; the work happens in BizyAir VITMatte Predict, downstream.

What it does

ViTMatte pairs a plain Vision Transformer backbone with a lightweight matting head, and unlike the trimap-free matting weights you'll find elsewhere in the ecosystem (BiRefNet's matting variants, for instance), it needs a trimap - a rough three-zone map marking "definitely foreground," "definitely background," and "unknown, figure it out." You typically build that trimap from a coarse segmentation mask (SAM or similar), eroded hard for the definite-foreground zone and dilated hard for the definite-background zone, leaving the boundary as the unknown region ViTMatte actually solves.

Inputs and outputs that matter

• detail_method - VITMatte or VITMatte(local). Given that this whole pack's premise is cloud-hosted execution with an option to fall back to weights you've cached yourself, the practical read is that the plain option runs on BizyAir's cloud and the (local) variant uses a copy you've downloaded locally instead. The README doesn't spell this specific choice out in detail, so treat that as the reasonable inference it is rather than a documented fact - if you're unsure which one you need, the plain VITMatte option is the one that doesn't require you to have fetched any weights yourself.

Outputs are two custom types you won't see anywhere else in ComfyUI: VitMatte_MODEL and VitMatte_predictor. Both are required inputs on the predict node - this loader exists purely to feed that node, it doesn't do anything useful on its own.

How to install it

• ComfyUI Manager - search "BizyAir", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git, restart.
• Comfy CLI - comfy node install bizyair.

Set your API key first, either through the in-app login prompt or at bizyair.siliconflow.cn. Unlike the pack's pure API-passthrough nodes (Veo, Vidu), the ViTMatte pair is one of BizyAir's cloud-hosted model executions - you're not paying a third party's API markup, but you're still running against BizyAir's account and credits rather than your local GPU unless you're specifically using the local variant.

Common issues & troubleshooting

Nothing downstream connects. VitMatte_MODEL and VitMatte_predictor are BizyAir-specific types - they won't plug into a generic matting node from a different pack, and a generic model/predictor pair from elsewhere won't plug into BizyAir VITMattePredict. Keep the loader and predict node from this pack paired.

You expected this to remove a background outright. It doesn't, by itself - it produces a matte given an image and a trimap you supply. If you don't already have a trimap-generation step in your workflow (typically a segmentation node feeding a mask-erode/dilate pair), matting alone won't get you a finished cutout.

BizyAir showed up in your install and you're not sure why. A ComfyUI Manager bug used to mis-resolve missing nodes onto this pack, which is why you'll find "is this safe" threads about it online. It's a legitimate pack from SiliconFlow, not malware - uninstall it through the Manager if you genuinely don't need it, keep it if you do want nodes like this one.