Generate Text Advanced BETA (Venice)

Local text generation in ComfyUI usually means finding VRAM for a quantized LLM on top of everything else your workflow already has loaded. Generate Text Advanced BETA (Venice) sidesteps all of that: it hands your prompt to venice.ai's servers and returns the model's reply as a plain string. No weights, no VRAM, no GPU at all - just an API key and a working internet connection. If you already pay for Venice, this is the text node the pack author wants you using.

It's the "Advanced" version for a reason. Where the basic Generate Text (Venice) node gives you five sampling dials, this one exposes the full chat-completions control panel, plus a couple of Venice-specific extras. The README is unambiguous: use this node for "updated text generation experience," and bolt on the Textgen Parameters (Venice) node when you want the Venice-only stuff.

How it works

Mechanically it's simple. The node builds a chat request (a system message plus your prompt), POSTs it to https://api.venice.ai/api/v1/chat/completions, and pulls the text out of choices[0].message.content. Nothing runs locally. That also means the one thing to keep in mind: this node has no chat history or memory. Every run is a stateless single exchange. If you want a multi-turn conversation, you're scripting it yourself and feeding the transcript back into system_prompt.

The inputs that actually matter

• model - a dropdown auto-filled from your Venice account (things like llama-3.1-405b). If it's stuck on the default, the model list fetch failed; that's a key problem, not a node problem.
• prompt and system_prompt - your ask, and the optional persona/instructions. enable_system_prompt is a handy kill-switch when you're testing whether the system prompt is actually changing behavior.
• temperature / top_p / top_k / min_p - the usual sampling knobs. The tooltip's advice is worth taking: tune temperature or top_p, not both, or you're fighting yourself.
• repetition_penalty - 1.0 means no penalty, and 1.2 (the default) is a gentle "please don't loop."
• max_completion_tokens - 420 by default, up to 131072. If replies are getting cut off, this is the first thing to raise.
• min_temp / max_temp - dynamic temperature scaling bounds, if you want the model to get more adventurous as it writes.
• enable_vision + image_for_vision - only works on vision-capable models, and only when both the flag and the image are present. The pack resizes your image down and base64-encodes it before sending.
• venice_parameters - the optional input you feed from Textgen Parameters (Venice) for characters, web search, and thinking control.

Output

One output: response (STRING). It lands as text on the node and wires into anything that takes a string - a text display node, or straight into an image prompt field if you're chaining generation. If nothing seems to happen, remember there's no visual preview built in; the output shows in ComfyUI's outputs panel.

Installing it

This is one node in the seven-node ComfyUI-Venice-API pack, so you install them all at once - via ComfyUI Manager (search "ComfyUI-Venice-API" or "Venice"), or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/DraconicDragon/ComfyUI-Venice-API.git
pip install -r ComfyUI-Venice-API/requirements.txt

Then restart ComfyUI, grab a key from venice.ai, and drop it in Settings > VeniceAI > API Key. The dependencies are light (requests, Pillow, numpy, configparser), and there are no model files to download - everything runs on Venice's side. It's labeled BETA, and it shows: the author openly calls the project a weekend experiment, so treat API changes as par for the course.