Tara Advanced LLM Node

If you're going to use this pack at all in 2026, this is the node. TaraPrompterAdvanced is the non-deprecated successor to the original TaraPrompter, and it fixes the three things that made the old one a pain: it takes a config object instead of ten raw inputs, it works with any OpenAI-compatible endpoint rather than two hardcoded providers, and it respects a seed. You hand it rough positive and negative prompts plus guidance, and it returns a refined pair you wire straight into CLIPTextEncode. The hook is simple: an LLM that rewrites your prompt with actual constraints, instead of you fighting the tag list by hand.

How it works

Mechanically it's the same trick as the legacy node - the model is instructed to reply in JSON ({"positive": "...", "negative": "..."}), the request uses response_format: json_object, the response is parsed and scrubbed of braces, brackets and quotes - but every sampling parameter now comes from the TARA_LLM_CONFIG object you feed in. That means one config can drive multiple prompt nodes, and swapping models is changing one node instead of rewiring a graph.

The inputs:

• llm_config (TARA_LLM_CONFIG) - created by TaraLLMConfig or TaraPresetLLMConfig. This carries base_url, api_key, model, temperature, seed, max_tokens, top_p, and penalties.
• guidance - your instructions to the LLM about how to refine.
• prompt_positive - the raw positive material (required).
• prompt_negative - optional; leave it empty and the system prompt just stops asking for negative guidance.

Outputs are positive and negative strings. Positive → CLIPTextEncode (positive side), negative → the negative side. The config's response_format override keeps JSON mode forced on, which matters for the gotcha below.

Pairing it

The intended graph is tiny and reads well:

TaraLLMConfig (or TaraPresetLLMConfig) → llm_config → TaraPrompterAdvanced
TaraPrompterAdvanced.positive → CLIPTextEncode (positive)
TaraPrompterAdvanced.negative → CLIPTextEncode (negative)

Point TaraLLMConfig at Groq, OpenAI, a local Ollama (http://localhost:11434/v1), LM Studio, or any OpenAI-compatible server. The config's seed is passed through to the API, which on OpenAI/Groq gives you reproducible-enough rewrites - genuinely handy when you're A/B-testing guidance text.

Gotchas

• JSON mode is mandatory. If your endpoint doesn't support response_format: json_object, the call errors. Groq and OpenAI both support it; small self-hosted servers sometimes don't. That's the main reason to pick the server before the node.
• The key prints to console. The code inherited a debug print of the api_key from the legacy node. On a shared box that's worth knowing - the author's own recommendation is to prefer the key loader over pasting keys into the graph.
• The model dropdowns in the preset node are dated (gpt-3.5-turbo, llama3-8b, gemma-7b era). Use TaraLLMConfig with a freeform llm_model field if you want something current.

Install

Standard pack install, two tiny deps, no models:

cd ComfyUI/custom_nodes
git clone https://github.com/ronniebasak/ComfyUI-Tara-LLM-Integration

Restart, or use ComfyUI Manager → Install via Git URL. It's a small alpha pack from ronniebasak and the community footprint is close to zero, so don't expect docs beyond the README - but for an LLM-in-the-graph prompt rewriter that can hit local models for free, this is the one you'd actually reach for.