Yogurt Sa2VA Image Segmentation

This is the node that makes prompt-based segmentation feel like cheating: "the woman in the red dress on the right" and Sa2VA just... cuts her out. No Grounding DINO phrases, no SAM boxes, no second pass. The old Grounded SAM 2 / Grounding DINO pipeline fell apart on anything more specific than "dog on right" - it's a detection model, not a text understander. Sa2VA is a full vision-language model, so it actually parses a whole sentence and segments what you meant. That's the gap this node fills, and it's the reason Sa2VA got the community's attention in the first place.

Feed it a model object from Yogurt Sa2VA Model Loader and a single image, describe the target, and it returns two things: the model's text response and the masks.

How it works

The node takes your image, wraps the prompt as <image>..., and calls Sa2VA's predict_forward in a single pass. The model returns probability masks per detected object (it was fine-tuned on SAM2 outputs, so it natively emits masks - no separate detector). Then a few steps happen in node-land:

• threshold converts the probability mask to a binary mask - 0.5 default, raise it for stricter (smaller, harder) masks, lower it to catch soft edges.
• morph cleans up the binary result: opening/closing to smooth noise, erode/dilate to shrink or grow, with erode_kernel/dilate_kernel/iterations controlling how aggressive. none is fine for clean subjects.
• Each object comes back as its own MASK, stacked into a batch - so "all objects" gives you one mask per object, and you can process them independently downstream.

There's one honest caveat, straight from the README: the configurable threshold only truly works on the Qwen-series Sa2VA models (like the Qwen3-VL-4B default), which get a patched predict_forward that returns raw sigmoid probabilities. On InternVL or kumuji models the node falls back to the model's built-in binarization and the threshold slider goes quiet. If your slider seems to do nothing, that's why - it's not you.

Inputs and outputs in one breath

• sa2va_model - from the Model Loader.
• image - RGB tensor, the thing to cut up.
• segmentation_prompt - describe it like you'd ask a person. Specific beats generic; "the wooden chair in the foreground" beats "chairs".
• threshold / morph / kernels / iterations - mask binarization and cleanup, defaults are sensible.
• Out: text_output (STRING - the model's caption/answer) and masks (MASK batch).

Wire masks into InvertMask → Join Image with Alpha for an instant transparent cutout, or straight into an inpainting node where you only want the masked region touched. The text_output is handy for agentic workflows where you want the model's reasoning recorded.

Install & the fine print

Same pack install as everything here: clone the repo into custom_nodes, pip install -r requirements.txt (transformers, qwen_vl_utils, pillow, numpy), restart. Models in ComfyUI/models/sa2va with a config.json, else transformers downloads them. Gotchas from the README that'll actually hit you:

• Dropdown empty after copying models - wrong structure (no config.json) or ComfyUI not restarted.
• 8-bit load fails - bitsandbytes not installed; install it or turn use_8bit off in the loader.

This node gives you masks, not mattes. Edges are object-level: good enough for inpainting and compositing a person, ragged on hair. If hair/fur/glass is your use case, skip to Yogurt Sa2VA Image Segmentation V2, which adds VITMatte refinement and gets you alpha-level edges.