Tenos.ai LLM Image Describe

A swiss‑army ComfyUI node that lets any vision‑capable LLM (OpenAI GPT‑4o, Gemini 1.5 Pro Vision, Groq‑mix… you name it) rip through a reference image and return a prompt you can feed straight back into your diffusion pipeline—no hand‑crafting required.

---

Why you’ll want it

⚡	Auto‑prompt in three flavors – “Style”, “Subject”, or “Describe” modes let you yank out pure style cues, pure content, or a balanced mix—all enforced by detailed system prompts baked into the node.
🔑	API‑key juggling – Reads keys from api_keys.json, but you can hot‑swap a key per‑node at runtime if you’ve got a second account (or if someone else is paying).
🖼️	Hash‑based caching – It won’t re‑hit the LLM unless the image or model settings change, saving your tokens (and sanity).
🧩	CLIP‑agnostic conditioning – Detects whatever encoding method your CLIP loader exposes (encode_from_tokens, encode_conditioning, encode, encode_text, or even a bare callable) and formats the output so the rest of your graph stays happy.
🚑	Fails graceful‑ish – No API key? Vision‑less model? Timeout? You get meaningful error text (and a blank conditioning tensor so the graph doesn’t hard‑crash).

---

File roster

custom_nodes/tenos_nodes/
├─ llm_describe_node.py          ← the node itself
├─ llm_models.json               ← list of provider → model names
└─ api_keys.json                 ← stash your secrets here (optional)

Both JSON helpers are optional but highly recommended—skip them and you’ll feed keys/models manually.

---

Quick install

# Inside your ComfyUI install:
mkdir -p custom_nodes/tenos_nodes
cp llm_describe_node.py llm_models.json api_keys.json custom_nodes/tenos_nodes/
# restart ComfyUI

Heads‑up: the file expects to sit next to its two JSON sidekicks. If you rename the folder, update the import path in the code (line right after the imports).

---

JSON cheat‑sheets

api_keys.json

{
  "openai_api_key":  "sk‑...",
  "google_api_key":  "AIza...",
  "groq_api_key":    "gsk_..."
}

Keys map 1‑to‑1 with provider names via an internal lookup table.

llm_models.json

{
  "providers": {
    "openai":  { "models": ["gpt-4o-mini", "gpt-4o-vision-preview"] },
    "gemini":  { "models": ["gemini-pro-vision", "gemini-1.5-pro"] },
    "groq":    { "models": ["llama-3-70b-vision"] }
  }
}

Anything not listed still works—type the full model string into the node’s dropdown or override box. The JSON just keeps the UI tidy.

---

Node I/O spec

Field	Type / UI	Description
clip	CLIP	Your text encoder. The node figures out how to talk to it.
image	IMAGE	4‑D tensor (B,H,W,C) from any loader.
llm_provider	dropdown	Values sourced from llm_models.json; "None" skips the call.
llm_model	dropdown	"provider: model" strings; editable.
description_mode	["Style","Subject","Describe"]	System‑prompt preset.
auto_describe_on_image_change	boolean	If false, you must nudge the node (toggle a field) to refresh.
override_api_key	string (optional)	Beats whatever’s in api_keys.json for this one call.
Returns	CONDITIONING	List [[embeddings, {"pooled_output": pooled}]] ready for SDXL / Flux.

---

Example ComfyUI graph (pseudo)

Loader → Tenos LLM Image Describe (Style, GPT‑4o‑mini) → SDXL Style LO → Merge → ...

Or go all‑in:

Loader → Split → Tenos LLM Image Describe (Subject) ─┐
         |                                          ├→ Combine Conditioning → Diffusion
         └→ Tenos LLM Image Describe (Style) ───────┘

---

Pro tips

• Batching: Node loops images internally, hitting the LLM once per frame. Keep batches small if you value tokens.
• Model sanity: Some providers don’t expose vision models yet. If you point “groq: llama‑3‑70b” at a JPEG, it’ll just ignore the picture and hallucinate—watch the console warnings.
• Clip weirdness: If your custom CLIP lacks pooled output, Flux will still run—just with less global juice.

---

Troubleshooting crib notes

Symptom	Likely Cause	Fix
Error: No API key provided… in prompt output	Empty api_keys.json or bad override field	Paste key, restart
Node prints Unknown LLM provider	Typo in dropdown or JSON	Match provider names exactly (openai, gemini, groq)
Blank conditioning but no error	auto_describe_on_image_change is off and nothing toggled	Tweak any field or feed a new image
CLIP object … does not have a recognized text encoding method	You forgot a CLIP loader	Add a CLIP‑Loader node upstream

---

Made with love and a dash of “why hasn’t ComfyUI done this yet?” ✨