My Ollama Special Generate Advance

MyOllamaSpecialGenerateAdvance is the most ambitious node in this pack and the easiest to be disappointed by. Instead of just asking an LLM to write a prompt, it asks the model to fill a structured template, then parses the reply with regex into separate ingredients - subject, style, lighting, camera angle, environment - and reassembles them into one comma-joined SD prompt. When it works it produces prompts that genuinely read like a pro's, with the whole (style:1.4) weighting vocabulary baked in. When it doesn't, you get empty fields and a shrug.

The template trick is why the README itself warns you to be careful with model choice. The default system prompt is a famous piece of prompt-engineering: a list of hand-crafted example prompts (the "artstation, greg rutkowski, trending on artstation" school) followed by an instruction to write a structured long prompt. If your model doesn't follow the JSON-ish template, every field comes back empty and you're left with an empty string on the wire.

How it works

The node appends Use the following template: {…} (a JSON blob with keys like description, long_prompt, camera_angle_word, style_words, subject_words, light_words, environment_words) to your prompt, sends it with the system instructions, then runs a handful of regexes over the response - code ported from the ComfyUI-Prompt-MZ project the author credits. Each matched field becomes an output. Then it joins the non-empty fields into a single prompt wrapped in parentheses, replaces stray periods and semicolons with commas, cleans up double commas, and optionally appends a preset style block.

That's where style_categories comes in - it's both an input and an output, which is a little odd but makes sense once you see it. Choose none, high quality, photography, or illustration on the input, and the node appends the matching quality-weight block (e.g. ((high quality:1.4), (best quality:1.4), (masterpiece:1.4), …) for photography) to the final prompt. The same value comes out the style_categories output so you can see what was applied.

The outputs

Nine of them. The one you usually wire into CLIPTextEncode is total_response - the fully assembled, comma-joined prompt. The other eight are the raw ingredients: description, long_prompt, camera_angle_word, style_words, subject_words, light_words, environment_words, and style_categories. If a regex finds nothing, that output is an empty string, so it's safe to chain them; they're mainly useful if you want to see what the model decided each part should be, or splice specific ingredients into a different prompt.

Inputs worth touching

• prompt - your short idea. The default Chinese example ("一个小女孩在森林里") is a good hint that this pack was built for multilingual input.
• style_categories - the four-way style switch described above.
• system - feel free to edit; if your model keeps ignoring the template, tightening the instruction here is the first thing to try.
• model / extra_model, temperature / top_p / top_k / num_predict / tfs_z, keep_alive - same shared sampling knobs as the rest of the pack.

Installing and caveats

Install once, works for all nodes in the pack: ComfyUI Manager (search "ComfyUi-Ollama-YN") or git clone https://github.com/wujm424606/ComfyUi-Ollama-YN.git into ComfyUI/custom_nodes. Dependencies are ollama, pandas, aiohttp, plus Ollama itself running locally.

Two honest warnings. First, model choice genuinely matters here - a small or instruction-weak model will not follow the template and you'll get empty outputs. If total_response comes back blank, try a stronger instruct model before you touch the node. Second, the pack has been quiet since September 2024, so the hardcoded template and regexes are frozen in time. They still work, but treat the whole thing as "works for the models that obey it," not a universal translator. And the usual dropdown rules apply: new models need a node re-add or restart, and type model:latest explicitly in extra_model.