LLM Dialogue Cycle (Simple)

LLM Dialogue Cycle (Simple) is the same A→B→A→B model-to-model loop as the full LLM Dialogue Cycle node, with the parameter panel slimmed down to what you actually touch. The full node exposes a dozen sampling and caching knobs on the canvas; this one gives you the two models, the system prompts, the cycle count - and pushes everything else into a JSON config file. Same engine underneath: it literally instantiates the full node and hands off.

That split is the pack's whole design philosophy. Standard nodes show you the runtime controls; Simple nodes keep the UI small and let advanced tuning live in config/simple_defaults.json, which you can either edit in place or point at with your own file via config_path. So if the full node feels like a cockpit, this is the one you put in a workflow you share with people who just want two models to argue.

The inputs that matter

Almost everything here is required, but it's all stuff you'd set anyway:

• modelA / modelB - your two GGUF files from ComfyUI/models/LLM/.
• cycles - round trips, one cycle being A then B. The Simple variant allows up to 100, double the full node's cap.
• system, systemA, systemB - shared prompt plus per-role overrides. Leave empty to fall back to whatever the config file sets.
• history_dir - where histories, summaries, the transcript, and session caches live; empty uses ComfyUI's output folder.

The two optional inputs are worth knowing. force_text_only disables mmproj auto-detection, and for today's text-only dialogue you want it True - that's the author's own recommendation, spelled out in the tooltip and the README. reset_session overwrites the session history with a fresh one while keeping the disk cache, so you can clean the slate without throwing away the speed benefit of cached prompts.

Output is a single transcript_text string with the whole exchange. Session naming works exactly like the full node: model A's history is {id}_A, B's is {id}_B, and the transcript appends to {id}.txt.

Installing and running it

Same pack as everything else here - ComfyUI Manager (search ComfyUI-LLM-Session) or clone into custom_nodes, install pillow, numpy, and llama-cpp-python, and restart:

cd ComfyUI/custom_nodes
git clone https://github.com/kantan-kanto/ComfyUI-LLM-Session.git
pip install pillow numpy llama-cpp-python

Model files go in ComfyUI/models/LLM/. The recurring gotchas all apply: two models means twice the VRAM, so pick your quants with care, and plain PyPI llama-cpp-python is fine for text-only - you only chase the JamePeng build if you're doing vision. If the node ever seems to ignore a setting you swore you changed, you've hit the config-file fallback: an empty UI field means the JSON default wins, and that's by design.