llama.cpp ADV++ Prompt

ADV Prompt does text-plus-images. ADV++ takes that and stacks on the three things that turn a chatty LLM into a reliable pipeline component: bundled prompt templates, token bans, and structured-output constraints. If you've ever gotten a caption back that starts with "Here is a detailed description of your image," this node is the cure.

It's the node you reach for when the LLM isn't just generating text - it's doing a job with strict output requirements. Caption → image prompt. Rough idea → structured prompt. Raw text → valid JSON. Those jobs are exactly where a plain chat model fails, because it hands you its conversational habits as literal output.

How it works

ADV++ is ADV Prompt's generation engine with three extra plug-in points, each coming from a sibling node or a bundled file:

• Templates - the template dropdown applies a bundled prompt template before generation. The pack ships Image2Prompt (rewrite an image into a single final image-generation prompt) and Prompt Enhancer (turn rough ideas into detailed prompts), defined in web/templates.json. Each template injects a system_prompt and an optional default user prompt. Templates are applied in Python, so API-format and headless workflows behave the same as the UI.
• Token bans - a token_ban input takes a list from a llama.cpp Token Ban node. The bans are sent as llama.cpp text-form logit-bias entries, so generation is steered away from specific tokens the same way the LTX-style hard-stop nodes block role delimiters.
• Structured output - a structured_output input accepts a constraint built by llama.cpp Structured Output (JSON Schema, JSON object, or GBNF grammar), which forces decoding into the shape you asked for.

The inputs that matter

Most of the sampling controls match ADV Prompt, so I'll skip those and hit the distinctive ones:

• template - Empty, Image2Prompt, or Prompt Enhancer. Restart ComfyUI after editing web/templates.json to add your own.
• prompt - the user prompt. With Image2Prompt this is where an image description or output: marker goes.
• image_amount / image_1…image_10 / include_image_batch - the full vision stack, same as ADV Prompt.
• token_ban + enable_token_ban - plug in a Token Ban node and toggle it.
• structured_output - plug in a Structured Output node. enable on that node is what bypasses the constraint when you're experimenting.
• stop_sequences - still useful even with structured output; belt and braces.

Outputs: response (the final generated or structured text), thinking, and success.

Why this matters

The KB's LLM-in-ComfyUI essay is blunt about the failure mode: a chat LLM doesn't emit a clean prompt by default - it emits role delimiters, preamble, markdown scaffolding. ADV++ gives you the three-layer fix in one node: a template that tells it the exact job, token bans that physically stop it from writing certain tokens, and structured output that constrains decoding itself. That combination is the difference between an enhancer that makes your prompts worse and one that earns its place in the graph.

Common issues

• Template output looks wrong - the bundled templates are opinionated (Image2Prompt is explicitly NSFW-tolerant and refuses to omit visible anatomy). Edit web/templates.json and restart if you want your own rules.
• Structured output and token bans fighting - a ban that contradicts the grammar can stall generation. Disable one while you debug.
• Model adds text outside the JSON - make sure strict is on in the Structured Output node and system_prompt is empty or minimal; the template's system prompt can override your formatting intent.