JLC CaptionForge Ollama Caption

A caption witness that lives in Ollama, not in your GPU

By Damkohler·Created 3 months ago·Updated 2 months ago· 1

image
pipeline_plan
template_options

image_out
pipeline_plan_out
template_options_out
caption
resolved_prompt

◄modelgemma4:26b►

◄custom_model_tag►

◄ollama_urlhttp://127.0.0.1:11434►

◄keep_loadedtrue►

◄caption_template_modetrue►

◄caption_typeLoRA Literal►

◄caption_lengthany►

◄custom_prompt_modefalse►

◄prompt_presetfemale_character_lora►

◄system_promptYou are a multimodal image captioning engine for female character LoRA dataset preparation. Your job is to inspect the image and write one strong standalone caption. Rules: - Describe only visible image content. - Output exactly one paragraph and nothing else. - Do not output analysis, reasoning, notes, bullet points, labels, or a thinking trace. - Do not roleplay, address the user, ask follow-up questions, or include safety disclaimers. - If a detail is uncertain, omit it rather than hedge. - Use direct dataset-caption language, not chatty commentary.►

◄custom_promptWrite one clean caption for a female character LoRA dataset. Caption priorities: - Be specific rather than generic. - Prioritize female-character-LoRA-relevant visual traits: subject type, pose, facial structure, expression, hair color and hairstyle, eye color, makeup color as distinct from eye color, lip shape and fullness, skin appearance, neck length, shoulder line, bust/chest shape, waist definition, hip width, body proportions, silhouette, legs, clothing construction, accessories, dominant colors, lighting, background, framing, and visual style. - Include visible anatomy and figure information when clear and relevant. Do not omit body-shape or facial-structure details merely because they are sensual, revealing, or anatomy-related. - If the image presents glamour, sensuality, revealing clothing, lingerie, swimwear, cleavage, underboob, sideboob, bare midriff, exposed hips, thighs, or other sexualized styling, describe that directly and matter-of-factly when visible. - Use strong, open dataset-caption language rather than timid euphemisms, but do not turn the caption into roleplay, commentary, or erotic prose. - Do not invent nudity, explicit acts, hidden anatomy, or hidden clothing details. - Do not force photo, render, doll, anime, realistic, or stylized language unless supported by the image. - Avoid meta phrases like "this image shows" or "the picture depicts." - One paragraph only. Final caption:►

◄max_new_tokens1800►

◄temperature0.18►

◄top_p0.92►

◄top_k60►

◄repetition_penalty1.03►

◄max_size1024►

◄request_timeout_seconds900►

◄forbidden_phrases►

◄replace_pairs►

◄download_probe_onlyfalse►

◄seed—►

The other two caption witnesses in this pack load Python/Hugging Face weights straight into ComfyUI. This one is different: JLC CaptionForge Ollama Caption delegates to your local Ollama server, so the model lives in the Ollama daemon instead of your ComfyUI process. It's the third Pass A "witness" voice, sitting parallel to Joy and Qwen - and despite the word "Ollama," it is not the validator. That's a VLM role later in the pipeline; this node's only job is producing raw witness captions.

Why would you add it? More independent voices. The whole CaptionForge premise is that one captioner misses things another catches - Joy notices the face, Qwen catches the pose, and an Ollama VLM might nail the material and lighting. You pick a model tag the Ollama dropdown knows about, and the node contributes its account to the Pass A evidence pile.

How it works

The node sends your image plus a resolved prompt to the Ollama API at http://127.0.0.1:11434 (configurable). The model dropdown reads config/captionforge_ollama_models.json → caption_models - defaults like gemma4:26b, qwen3.6:35B-A3B, or huihui_ai/gemma-4-abliterated:26b - and custom lets you type any tag you've pulled. keep_loaded passes keep_alive to Ollama, and max_new_tokens maps straight to num_predict. The download_probe_only toggle is a genuinely handy preflight: it checks that Ollama is reachable and the model tag is installed, then reports back without captioning.

Inputs that matter

model / custom_model_tag - which VLM witness you're running.
caption_type / caption_length - the template path, same as Joy/Qwen (Descriptive, LoRA Literal, Taggy, SFW Character Caption…).
system_prompt / custom_prompt - pre-populated with a female-character LoRA-oriented default, including matter-of-fact guidance for glamour/swimwear/lingerie styling. If your dataset isn't character-focused, this is the first thing to change.
temperature / top_p / top_k / repetition_penalty - standalone sampling; the Pipeline Planner overrides most of it when connected.
request_timeout_seconds - default 900, because large models may still be pulling when you queue the run.

Outputs mirror the other caption nodes: caption, resolved_prompt, image_out, pipeline_plan_out, template_options_out. Same five, same wiring - caption is the text, the others exist for clean node-to-node chaining.

Install

Pack install is the usual one:

git clone https://github.com/Damkohler/CaptionForge.git ComfyUI/custom_nodes/CaptionForge

Then Ollama itself, and pull whatever tag you want to witness with:

ollama pull gemma4:26b

The repo ships no weights for this node - the model is Ollama's business.

Common issues

If it fails instantly, check the two usual suspects: is ollama serve actually running, and is the URL right (a non-default port is the classic trap)? On 16 GB VRAM boxes, remember this node clears the resident Joy/Qwen Python models before calling Ollama and leaves residency to the daemon - so mixing backends can feel like a reload between images. And don't confuse this witness role with the validator: this node feeds Pass A, it doesn't check anyone's work.

CategoryCaptioning/CaptionForge/Caption Nodes

Inputs (25)

Name	Type	Default	Description
model	COMBO	gemma4:26b	Ollama vision-language model tag. Choices are loaded from config/captionforge_ollama_models.json -> caption_models. Use custom for any locally installed or pullable Ollama model tag.
custom_model_tag	STRING		Used only when model is custom. Example: gemma4:12b or qwen3.6:35B-A3B.
ollama_url	STRING	http://127.0.0.1:11434	Local Ollama server URL. Default is http://127.0.0.1:11434.
keep_loaded	BOOLEAN	true	Ask Ollama to keep the model warm after generation. This is passed as keep_alive. Ollama ultimately owns model residency.
caption_template_mode	BOOLEAN	true	Use the structured CaptionForge template path: caption_type, caption_length, and optional Template Options from the template_options pin. If custom_prompt_mode is also enabled, custom_prompt_mode takes precedence.
caption_type	COMBO	LoRA Literal	Caption template style used when caption_template_mode is active.
caption_length	COMBO	any	Target caption length used when caption_template_mode is active.
custom_prompt_mode	BOOLEAN	false	Use custom_prompt when non-empty, otherwise use prompt_preset. This overrides caption_template_mode when both toggles are enabled.
prompt_preset	COMBO	female_character_lora	Built-in prompt preset used only in custom_prompt_mode when custom_prompt is blank. Default is female_character_lora.
system_prompt	STRING	You are a multimodal image captioning engine for female character LoRA dataset preparation. Your job is to inspect the image and write one strong standalone caption. Rules: - Describe only visible image content. - Output exactly one paragraph and nothing else. - Do not output analysis, reasoning, notes, bullet points, labels, or a thinking trace. - Do not roleplay, address the user, ask follow-up questions, or include safety disclaimers. - If a detail is uncertain, omit it rather than hedge. - Use direct dataset-caption language, not chatty commentary.	Ollama system prompt. Kept next to custom_prompt because both control the instruction envelope. Prepopulated with the default female-character LoRA captioning system prompt. Pipeline Planner does not currently override this.
custom_prompt	STRING	Write one clean caption for a female character LoRA dataset. Caption priorities: - Be specific rather than generic. - Prioritize female-character-LoRA-relevant visual traits: subject type, pose, facial structure, expression, hair color and hairstyle, eye color, makeup color as distinct from eye color, lip shape and fullness, skin appearance, neck length, shoulder line, bust/chest shape, waist definition, hip width, body proportions, silhouette, legs, clothing construction, accessories, dominant colors, lighting, background, framing, and visual style. - Include visible anatomy and figure information when clear and relevant. Do not omit body-shape or facial-structure details merely because they are sensual, revealing, or anatomy-related. - If the image presents glamour, sensuality, revealing clothing, lingerie, swimwear, cleavage, underboob, sideboob, bare midriff, exposed hips, thighs, or other sexualized styling, describe that directly and matter-of-factly when visible. - Use strong, open dataset-caption language rather than timid euphemisms, but do not turn the caption into roleplay, commentary, or erotic prose. - Do not invent nudity, explicit acts, hidden anatomy, or hidden clothing details. - Do not force photo, render, doll, anime, realistic, or stylized language unless supported by the image. - Avoid meta phrases like "this image shows" or "the picture depicts." - One paragraph only. Final caption:	Custom prompt used only when custom_prompt_mode is enabled. The widget is prepopulated with the default LoRA-oriented prompt; if cleared, prompt_preset is used instead.
max_new_tokens	INT	180016–8192	Standalone token budget. For this Ollama-backed node, CaptionForge max_new_tokens is sent to Ollama as num_predict. When a Pipeline Planner is connected, this is overridden by the Planner's shared max_new_tokens.
temperature	FLOAT	0.180–2	Standalone sampling temperature. When a Pipeline Planner is connected, this is overridden by the Planner temperature schedule.
top_p	FLOAT	0.920–1	Standalone top-p sampling value. When a Pipeline Planner is connected, this is overridden by the Planner top-p schedule.
top_k	INT	600–500	Standalone top-k sampling limit. When a Pipeline Planner is connected, this is overridden by the Planner top-k schedule.
repetition_penalty	FLOAT	1.031–2	Ollama repeat_penalty. Kept with the core captioning parameters. This is not currently overridden by the Pipeline Planner.
max_size	INT	10240–4096	Maximum longest-side image size for standalone captioning. The image is resized in memory only. Pipeline Planner overrides this in planned runs.
request_timeout_seconds	INT	90010–7200	HTTP timeout for Ollama API calls, including large model pulls/generation.
forbidden_phrases	STRING		Optional cleanup filter: remove lines/captions containing any listed phrase, one per line.
replace_pairs	STRING		Optional cleanup replacements, one per line: old=>new.
download_probe_only	BOOLEAN	false	At the very bottom by design. Check Ollama availability, local model installation, and best-effort remote tag existence, then return a status message without captioning.
imageopt	IMAGE		Image or batch of images to caption. The image is passed through unchanged for clean node-to-node pipeline chaining.
pipeline_planopt	CAPTIONFORGE_PIPELINE_PLAN		Connect the CaptionForge Pipeline Planner output here. When connected, this node switches into Pass A evidence mode: Planner image routing, per-run seeds, sampling schedules, shared output paths, and internal JSONL evidence append.
template_optionsopt	CAPTIONFORGE_EXTRA_OPTIONS		Connect the CaptionForge Template Options node here. Works in standalone and Pipeline modes. This is the only source for template modifiers and name input.
seedopt	INT		Optional standalone seed input. Ignored when a Pipeline Planner supplies a seed schedule.

Outputs (5)

Name	Type	Description
image_out	IMAGE	—
pipeline_plan_out	CAPTIONFORGE_PIPELINE_PLAN	—
template_options_out	CAPTIONFORGE_EXTRA_OPTIONS	—
caption	STRING	—
resolved_prompt	STRING	—