Together Vision Batch Node

If you've ever needed captions for a folder of images - a dataset, a Lora training batch, an archive of renders you can't remember - you know how miserable the manual route is. Preview, copy, paste, repeat. This node is the blunt instrument for the whole job: point it at a directory, run it once, and every image gets a .txt file sitting right next to it, named after the image. No dragging dozens of files through your graph.

How it works

It's the same engine as the pack's single-image TogetherVisionNode - same retries, same 1-second client-side pacing, same model list, same output-cleaning logic. The difference is the input: instead of an IMAGE tensor it takes an image_directory string, globs every *.png, *.jpg, *.jpeg, *.bmp, and *.webp in it, and processes them one at a time, waiting for each description before moving on. Each result is written to <imagename>.txt in the same folder.

The inputs that matter

• image_directory - the one you'll actually touch. It's a plain string field, no folder picker: paste an absolute path to a real, readable directory.
• model_name - five options here (no EXAONE on this one). The default is the free DeepSeek text model; pick meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo if you actually want the images seen - the free defaults are text-only.
• api_key - paste it, or set TOGETHER_API_KEY in the pack's .env.
• system_prompt / user_prompt - same art-critic defaults as the single node; rewrite them to taste.
• temperature / top_p / top_k / repetition_penalty / max_tokens - same sampling knobs, slightly different defaults (0.9 top_p, 512 tokens).
• use_cache / clean_output_text / max_retries / retry_delay - leave them alone until you have a reason.

Output

The schema says STRING, and here's the honest quirk: the node returns an empty string. It's marked as an output node because the real deliverable is the .txt files it writes to disk. If you're expecting a visible result in the UI, you'll get nothing - check the folder.

Installing it

ComfyUI Manager (search "Together Vision"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/theshubzworld/ComfyUI-TogetherVision

Restart, let Manager install together and python-dotenv, add your API key. No model downloads - everything runs on Together's servers.

Gotchas

• It's sequential and slow by design. N images means N API calls with a 1-second floor between them. A 100-image folder on the free tier is minutes of churn, and the free rate limits will push back. Run small batches.
• No skip-if-exists. Re-running re-processes every image and overwrites the .txt files. The in-memory cache only helps within a session, so don't expect reruns to be free.
• Bad path = silence. If the directory doesn't exist or has no matching images, the node logs a warning to the console and returns - no error bubble in the UI. When it "does nothing," check the terminal.
• Same deprecation caveat as the rest of the pack: the README now just points at the author's ComfyUI-NvidiaVision successor, and the code hasn't been touched since December 2025. It works; it's just not going to get fixes.