Yogurt Sa2VA Model Loader

Every workflow using this pack starts here. Yogurt Sa2VA Model Loader loads ByteDance's Sa2VA - the vision-language model fine-tuned on SAM2 outputs that lets you describe what to segment in plain English - once, caches it, and hands you a reusable model object instead of reloading a multi-GB model every time a segmentation node runs.

That last part is the entire reason this pack exists. The upstream alexjx/ComfyUI-Sa2VA-XJ loaded Sa2VA inside each segmentation node, so every node in your graph paid the full load cost. This fork splits loading out: one loader, one load, N downstream nodes. If you're running a Sa2VA-XJ workflow where three segmentation nodes each hit the disk, switching to this pack is the upgrade.

How it works

The loader builds a YogurtSa2VAModel bundle - model + processor + an inference lock - and caches it in a dict keyed by (model_name, use_8bit, use_flash_attn). Same config, same model object back. The underlying torch module is wrapped and registered with ComfyUI's model_management, so when some other node needs VRAM, ComfyUI can offload Sa2VA and drag it back before your segmentation node runs. No manual unloading.

Two memory paths, worth knowing:

• 8-bit (use_8bit, default on): needs bitsandbytes. The model can't reliably do .to("cpu"), so on VRAM pressure it's deleted and reloaded on demand. Frees the most memory; costs a reload every time something else wants the GPU.
• fp16/bf16: whole-model CPU/GPU offload. Reloads are faster, but a 4B model is still 4B - it eats more VRAM.

Inputs that matter

• model_name - seven candidates in the dropdown, default ByteDance/Sa2VA-Qwen3-VL-4B. That default is the community darling (the Qwen3-VL-4B rev is what made Sa2VA mainstream in late 2025). Sa2VA-Qwen2_5-VL-7B or InternVL3-8B/14B if you want more capacity and have the VRAM; kumuji/Sa2VA-i-1B if you're on a potato.
• use_8bit - on by default, and it's a good default on 8GB cards. Turn it off if you'd rather eat VRAM than eat reload latency.
• use_flash_attn - on by default but graceful: if flash-attn isn't installed, it just logs and continues. You don't need it; it just makes generation faster when present.
• force_reload - discards the cache for this config and reloads. For after you swap model files.

Output is a single sa2va_model (YOGURT_SA2VA_MODEL), which feeds Yogurt Sa2VA Image Segmentation, ... Video Segmentation, and ... V2.

Install

Whole-pack install - one clone gets you all five nodes:

cd ComfyUI\custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtSa2VA
cd ComfyUI-YogurtSa2VA
python -m pip install -r requirements.txt

requirements.txt is minimal: transformers>=4.57.0, torch, qwen_vl_utils, pillow, numpy. Optional extras per the README:

python -m pip install bitsandbytes      # for use_8bit
python -m pip install flash-attn --no-build-isolation   # for use_flash_attn

Models go in ComfyUI/models/sa2va as plain folders containing config.json - use huggingface-cli download --local-dir rather than dumping HF cache. If nothing's there, transformers pulls it from Hugging Face on first run.

Gotchas

• 8-bit load failure: bitsandbytes missing. Install it or flip use_8bit off.
• Dropdown empty after copying models: wrong directory structure or no config.json, then not restarted. Check both.
• First run is slow, embarrassingly so: the 4B model download plus load is a multi-GB affair. That's why you cache it once and reuse - and it's exactly what this node is for.

If you only ever segment one image in a whole session, a loader feels like ceremony. The moment you run two nodes, or a video, it's the difference between a snappy workflow and watching models load in a queue.