🔎ESAM Model Loader

The other half of the pair. Where 🔎Yoloworld Model Loader gives you a detector that finds where things are as boxes, this node loads EfficientSAM - the model that turns those boxes into actual pixel-accurate masks. If you only want bounding boxes, you can skip this node entirely. If you want a clean cutout mask of "the dog" rather than just a rectangle around it, this is the loader that makes that happen.

What EfficientSAM actually is

SAM - Segment Anything - was Meta's model for turning a point or a box prompt into a precise object mask, and it's genuinely excellent, but the full model is heavy. EfficientSAM is a distilled, lighter version built to do the same job at a fraction of the compute cost, which matters a lot here because this pack runs it per-detection on every object YOLO-World finds. In the downstream 🔎Yoloworld ESAM node, EfficientSAM is what turns each of YOLO-World's bounding boxes into a proper mask - that's the whole reason "EfficientSAM" is half the pack's name.

The one input that matters

Just one field: device, a dropdown of CUDA or CPU. Pick CUDA if you have an NVIDIA GPU available - it's not close, GPU inference here is dramatically faster than CPU, especially once you're segmenting multiple detected objects in a single image. CPU exists as a fallback for machines without a usable GPU, or if you're deliberately keeping VRAM free for something else in the same workflow. There's nothing else to configure - no model size choice, since EfficientSAM only ships the one variant this pack wires up.

The output

One output: esam_model (type ESAMMODEL). It plugs into 🔎Yoloworld ESAM's esam_model input for full detect-and-segment, or optionally into 🔎Yoloworld ESAM Detector Provider's esam_model_opt input if you're feeding Impact-Pack and want mask-level detection rather than just boxes.

Installing it

Through ComfyUI Manager: search ComfyUI YoloWorld-EfficientSAM and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-YoloWorld-EfficientSAM
cd ComfyUI-YoloWorld-EfficientSAM
pip install -r requirements.txt

Restart ComfyUI. The part people miss: EfficientSAM's weights don't auto-download the way YOLO-World's do. Per the README, you need to manually grab efficient_sam_s_cpu.jit and efficient_sam_s_gpu.jit from the camenduru/YoloWorld-EfficientSAM HuggingFace repo and drop both files straight into the custom_nodes/ComfyUI-YoloWorld-EfficientSAM folder - not models/, not a subfolder, the pack's own root directory. That's an unusual place to stash weights and it's the single most common reason this loader fails on a fresh install: the node loads fine, then throws a file-not-found the moment you actually run it because the .jit files were never placed.

Where people get stuck

Grab both .jit files even if you only plan to use one device - the node picks the matching one based on your device selection, and a missing file for whichever mode you chose is a hard failure, not a graceful fallback. Beyond that, this pack shares the same landmines as its sibling loader: it hasn't been meaningfully updated since early 2024, users have hit install friction on newer Python versions (3.12 is a known pain point - stick to 3.10/3.11 if pip install -r requirements.txt chokes), and by 2025 some in the community were pointing newcomers toward better-maintained alternatives instead of fighting this one's dependencies. It still works for plenty of people - just go in knowing the manual model placement step is real and easy to skip past.