TogetherVisionNode

Here's the loop: generate an image, describe it, feed the description back into the prompt, generate again. It's how a lot of people push a render from "good" to "this is exactly what I meant." TogetherVisionNode is the describe step - a hosted vision LLM, no local model, no VRAM. Connect any IMAGE and it hands you back a STRING you can pipe into a positive prompt, a ShowText node, or a text file.

How it works

The node encodes your image tensor to a base64 PNG - auto-resizing anything larger than 1024px on the long edge, so a 2K render still goes through - then calls Together AI's chat.completions endpoint with your system and user prompts, streaming the reply back. It's not a thin wrapper, either. The source ships real handling: a 1-second client-side pacing delay, retries on timeouts and connection errors, specific error messages that distinguish a bad key from a rate limit, and an in-memory cache (last ~100 responses) keyed on model + prompts + sampling params + an image hash. Leave the image input disconnected and it degrades into a plain text LLM node - handy for rewriting prompts without building a second workflow.

The inputs that matter

• model_name - this is the one to get right. The default is deepseek-ai/DeepSeek-R1-Distill-Llama-70B-free, a free text model. The genuinely vision-capable model on the list is meta-llama/Llama-3.2-90B-Vision-Instruct-Turbo - that's the one that actually sees your image, and it's the paid tier. Want image descriptions? Pick a vision model, or use Other (Custom) and type any model Together hosts.
• system_prompt / user_prompt - the defaults are an "ekphrasis art critic" persona in British English with hashtags. Rewrite them; this is 80% of output quality.
• seed_mode + seed - fixed, random, or increment control the sampling seed sent to the API.
• max_tokens - default 1024, up to 4096. Your description budget.
• temperature / top_p / top_k / repetition_penalty - the usual sampling knobs; defaults are sane.
• use_cache - in-memory cache, leave it on.
• clean_output_text - strips <think> reasoning blocks (the default DeepSeek model loves those), "Here is the description:" preambles, and code fences.

Output

One description STRING. It wires straight into a text prompt input, or into ShowText/SaveText if you want it visible or on disk.

Installing it

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

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

Restart, then install the two deps (together, python-dotenv) when Manager asks. No model files. You need a Together API key - paste it in the node or set TOGETHER_API_KEY in the pack's .env.

Gotchas

• Only the first image of a batch is described - encode_image takes image[0] and silently ignores the rest. If you're feeding a batched output, expect one description.
• Free-tier models are rate-limited and can be slow; the 90B vision model costs per token.
• The pack is deprecated - the README now just points at the author's ComfyUI-NvidiaVision successor. It still works (last commit December 2025), it's just unmaintained.

Troubleshooting

The node writes readable errors to the UI: auth failures say so, rate limits tell you to wait and roughly how long, "Model not found" means your custom model name is wrong, and timeouts retry up to max_retries times before giving up. If you get an empty description, check whether your chosen model can actually see images - that's the most common beginner trip here.