☁️BizyAir Load SAM Model

This loads Meta's original Segment Anything Model - the ViT-H checkpoint specifically, at 2.56GB, which the dropdown label tells you upfront so you know what you're pulling in before you commit to it. If that GB figure made you wince, that's the whole point of running it through BizyAir: the weights live on SiliconFlow's servers, not your disk or VRAM.

Worth flagging early: this pack also ships BizyAir LayerMask: Load SAM2 Model, a different node loading a different, newer model (SAM2, with video tracking support) that outputs an incompatible type. This node and that one aren't interchangeable - more on that below.

What SAM actually does here

SAM's whole trick is point-and-click (or box, or eventually text-prompted) segmentation: give it a coordinate or a bounding box on an image, and it returns a precise mask around whatever object is there. No training, no per-object model - one general-purpose segmentation model that works on basically anything you point it at. In ComfyUI workflows it typically shows up paired with GroundingDINO for text-prompted masking ("find and mask the red dress") or on its own for manual point-based masking, most commonly feeding an inpainting pass on a specific region.

The original SAM is the slow-but-thorough end of the segmentation-model family. It's since been joined by faster and newer options - FastSAM traded accuracy for speed, and SAM2 added video-frame tracking - but the original ViT-H checkpoint is still solid for single-image, high-quality masking work where you're not in a hurry.

Inputs and outputs

Just one input:

• model_name - currently a single choice, sam_vit_h (2.56GB). Nothing to decide here yet; BizyAir has one SAM variant deployed.

One output: SAM_PREDICTOR, typed specifically for downstream nodes expecting a SAM predictor object - the node you'd wire this into next is whatever BizyAir (or a compatible local pack) uses to actually run point/box-prompted segmentation.

The type mismatch to know about: this outputs SAM_PREDICTOR. The pack's other SAM loader, BizyAir_LayerMaskLoadSAM2Model, outputs LS_SAM2_MODEL - a completely different type built for the dzNodes/LayerMask node family. Don't expect to swap one loader for the other in an existing graph; they feed different downstream nodes entirely.

How to install it

Ships with the full BizyAir pack.

ComfyUI Manager: search "BizyAir," Install, restart.

git clone:

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

Set your BizyAir API key before running anything - this node included, since the SAM inference itself happens server-side.

Common issues

Wiring this into the wrong downstream node. Given the type-mismatch situation above, the most likely mistake is grabbing a masking node that actually wants LS_SAM2_MODEL and being confused when this loader's output won't connect. Check which SAM type your masking node expects before picking which loader to use.

No text-prompted masking on its own. SAM by itself doesn't understand language - "mask the dog" isn't something you can ask it directly. That needs GroundingDINO (or an equivalent text-to-box detector) upstream to find the region first, with SAM refining that box into a precise mask. If your workflow needs prompt-driven masking rather than manual point/box input, you need both pieces, not just this loader.

One fixed model, no swap. Because BizyAir only exposes what SiliconFlow has actually deployed, you can't currently point this node at FastSAM, SAM2, or SAM3 for a speed or quality trade-off - that flexibility exists in local SAM node packs, not here. If you need those newer variants specifically, look at what else this pack offers (or a local pack) rather than expecting this loader's single option to change.