CTGPhrasesSimple

Same job as its sibling GR85_CTGPhrases, and the "Simple" in the name tells you the difference is in how much logic sits behind it - this is presumably the lighter, more literal splitting pass, where the other one may do more processing on the input. Both feed the same downstream consumer: Florence-2's "Caption to Phrase Grounding" task, which wants a list of individual phrases to locate in an image, not one string.

What the schema tells you. One required input, input_string - your phrase text, however you've built it (typed manually, assembled from a prompt template, or pulled from a caption). One output, STRING with is_list: true - the phrase text turned into a list, ready for GR85_Florence2RunCTPG's text_input to consume. Beyond that, the interface is identical to GR85_CTGPhrases: same input shape, same output shape, no extra toggles on either side.

Given that overlap, it's fair to treat these two nodes less as "pick the right one" and more as "the pack gives you two splitting strategies, try them both once and keep whichever matches how you actually write phrase lists" - some people write phrase one, phrase two, phrase three, others prefer one phrase per line, and a splitter tuned for one convention can mangle the other.

Installing it. Search "comfyui_gr85" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/veighnsche/comfyui_gr85

Restart ComfyUI. No models or heavy dependencies attach to this node directly - it's a string-to-list converter. The Florence-2 model it ultimately feeds is loaded by a separate node elsewhere in your graph.

Which one to actually use. With two nodes doing the same conceptual job and no description distinguishing them beyond "Simple" in the name, your best bet is to try both on the same input_string and compare with a GR85_ShowText node on each output - if they produce identical lists for your input, use whichever; if they diverge, that difference is your actual documentation for how each one splits text, and you can pick based on which result matches what you meant. This is a case where the honest answer is: the pack doesn't spell out the distinction anywhere public, so verify against your own input rather than guessing.

Where it fits in the bigger picture. Florence-2 is a real, widely-used vision-language model in the ComfyUI world - for auto-captioning LoRA training sets and for grounding/segmentation tasks alike - and phrase-list prep like this is a small but necessary step before you can run a grounding pass. If your grounding results come back missing an object you expected it to find, check the split list before you blame the model; a phrase that got mangled during the string-to-list conversion never had a chance to match anything in the image.