Extra Parameters

This is the pack's smallest node and probably its most overhyped-adjacent one: a two-dropdown box that produces a little extra_parameters object you feed into the optional input on API Chat or the Captioner. What it controls is the thinking/reasoning behavior of models that have it - the "think before you answer" mode on GPT-5-class and Seed-1.6-class models. If your model doesn't support thinking, the node does nothing, which is worth knowing before you build a workflow around it.

How it works

Two combo inputs:

• thinking - disabled (default) or enabled.
• reasoning_effort - minimal, low, medium, or high (default medium).

The node packages those into an extra_parameters output. The real work happens downstream: whichever chat node you plug it into translates the pair into the provider's own fields. The README is specific about the reach - it only works on the OpenAI-compatible clients (chat and responses). So the honest summary is:

• On OpenAI-compatible APIs, reasoning_effort gets sent to models that accept it (GPT-5, GPT-5.1, Seed-1.6 and friends), and thinking only does anything on some Volcengine/Seed-1.6-series models, where it maps to the provider's thinking payload field.
• On Ollama, thinking is translated into the think option (and reasoning_effort into an effort hint) for the models that support it - but this isn't the OpenAI path, so don't expect identical behavior.
• On Anthropic, the effort levels are translated into thinking budget tokens (low ≈ 1024, medium ≈ 2048, high ≈ 4096).
• On Mistral, it's effectively ignored.

That's a lot of "it depends," which is the honest state of thinking-mode plumbing in 2026: there's no standard knob, so every backend names it differently.

What it's for

Thinking mode buys better answers on hard reasoning at the cost of latency and tokens. For the pack's actual jobs - captioning and grounding - the Captioner and Generate BBOXes already pass their own conservative settings, and the source comments note thinking + medium effort is even a good combo for captions. Where you'd manually wire this node in is the chat node, when a prompt needs the model to actually reason before replying rather than pattern-match. If you're just doing "rewrite this prompt," leave thinking off and save the tokens.

Install and gotchas

It ships with the pack: install ComfyUI-PDiD-LLM-Nodes via ComfyUI Manager (search the pack title) or:

cd ComfyUI/custom_nodes
git clone https://github.com/power88/ComfyUI-PDiD-LLM-Nodes

Restart, and you need ComfyUI 0.3.51+ since this pack is built on the V3 node schema. Nothing to download - it's all API-side behavior.

The traps are the two things the README warns about directly: thinking only works on some Volcengine Seed-1.6-series models, and reasoning_effort only on the newest thinking models. If your model is neither, the node is a pass-through and your workflow will run fine while silently ignoring it. Check that your model actually supports the knob before you assume the node is doing something - the biggest surprise here is when it isn't.