🦙 Ollama Captioner Extra Options 🦙

This node doesn't do anything on its own - it's a checklist of toggles that plugs into OllamaImageCaptioner's optional extra_options input, and each toggle you flip gets appended to the captioning instructions as its own sentence. If you've used JoyCaption's ComfyUI node before, this will look familiar: it's the same pattern, a menu of "include this / exclude that" rules instead of you hand-writing a paragraph of instructions every time.

The reason it exists as a separate node rather than a wall of text on the captioner itself is reuse - build one set of toggles that matches how you caption for a given LoRA (say, always flag lighting and composition, never mention resolution or watermarks), and wire the same Extra Options node into a batch job without retyping anything.

The options that matter

There are 17 boolean toggles total, every one off by default, and the literal field label is the instruction - there's no separate description to write. A few worth knowing about specifically:

• "If there is a person/character in the image you must refer to them as {name}." - pairs with the name field on the captioner node; useful for LoRA datasets where you want a consistent trigger token instead of "a woman" or "a man" in every caption.
• "Do NOT include information about people/characters that cannot be changed (like ethnicity, gender, etc), but do still include changeable attributes (like hair style)." - this is the identity-lock toggle you want on for most character LoRA work: it's the same principle the training community has settled on - describe what should stay adjustable, leave what should be permanent undescribed.
• "Include information about whether there is a watermark or not." and "Include information about whether there are JPEG artifacts or not." - genuinely useful for filtering a scraped dataset, less useful for clean renders.
• "You MUST include information about the subjective aesthetic quality of the image from low to very high." and "Include information on the image's composition style, such as leading lines, rule of thirds, or symmetry." - for building richer captions when you want the model to reason about photographic quality, not just contents.
• "Do NOT include anything sexual; keep it PG." and "Include whether the image is sfw, suggestive, or nsfw." - content-classification toggles for filtering or labeling a mixed dataset.

The rest cover lighting, camera angle, likely camera/aperture/shutter/ISO details for photos, depth of field, ambiguous-language avoidance, in-image text, and a "only describe the most important elements" catch-all for terser output. Flip on whichever match what you actually want out of your captions - most people use two or three, not all seventeen, since stacking too many instructions tends to make the model's captions read like a checklist rather than a description.

Inputs and outputs

There's no image, prompt, or model selection here - every input is one of the 17 booleans, and the single output is an Extra_Options value that only means anything wired into the captioner's extra_options slot.

How to install it

Ships as part of the same pack as OllamaImageCaptioner - install once, get both. Via ComfyUI Manager, search "ComfyUI-Ollama-Describer" and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/alisson-anjos/ComfyUI-Ollama-Describer.git
pip install -r requirements.txt   # or install.bat on Windows

Restart ComfyUI. No separate model download is needed for this node specifically - it's pure logic that shapes the prompt sent by whichever captioner it's attached to; the actual vision model requirement lives on OllamaImageCaptioner.

Common issues & troubleshooting

This node has no failure mode of its own worth naming - it can't error out, since it's just booleans feeding a string. If your captions aren't reflecting a toggle you flipped, check the connection actually reaches extra_options on the captioner node (it's an easy wire to forget), and remember these are instructions to the LLM, not hard guarantees - a smaller or weaker vision model will sometimes ignore an instruction it doesn't fully understand, especially the more technical ones like camera settings. If captions come back generic despite several toggles being on, that's usually the model, not this node - try a stronger vision model on the captioner side.