KoboldCpp Prompt Amplify

This is the flagship node of the pack, and it earns it. KoboldCpp Prompt Amplify takes a short, rough prompt - "a girl in the rain with a red umbrella" - and rewrites it into something a diffusion model can actually paint, using a local LLM running on your own machine. No API key, no cloud, no censorship layer, no per-call cost. The KB calls this the most common LLM-in-the-graph job, and this node is one of the better implementations of it: it's the same rewrite engine Prompt Studio's chat interface uses, exposed as a plain node you can drop into any workflow.

The typical graph is about as simple as it gets:

primitive text → KCPP_PromptAmplify → positive text encoder → sampler

Type your idea in a text primitive, let the node expand it, feed the result to the encoder. If you've ever tried to hand-format a prompt for a tag-based anime model and a natural-language photoreal model in the same week, you'll appreciate what this does.

How it works

The node talks to KoboldCpp's OpenAI-compatible Chat Completions endpoint (the raw KCPP_Apply node uses the older native endpoint instead). It builds a rewrite request from the inputs, sends it, and returns the model's answer as amplified_text. Native thinking (the thinking_mode input) is kept in Chat Completions' separate reasoning_content field, so only the clean final answer reaches your image workflow - no leaked scratch-work, which is the classic reasoner failure the KB warns about.

The inputs that matter:

• text - your rough idea. This is the input to amplify.
• model_profile - the grammar the rewrite follows: General Natural Language, Tag-Based Anime Model, or Tag-Based Photorealistic Model. This is the single most important setting; pick the one that matches your checkpoint's prompting style.
• style_preset (111 options) and framing_preset (32 options) - reusable aesthetic and composition guidance. There's real overlap with things like SDXL style/quality tags, but they're tuned as LLM instructions.
• embellishment_level - how much the rewrite expands your text, from None (no added detail) to Ultra Maximum (~120–160 words). Start at Clean and climb only if you want more prose.
• thinking_mode - Disabled through High. Reasoning modes cost context and tokens; leave Disabled unless you're chasing quality.
• secondary_instructions - optional, and important: phrases like LoRA trigger words that pass through to the second output unchanged and never enter the LLM request.
• aspect_ratio, megapixels, multiple - compute the optional width/height outputs with the same presets as ComfyUI's Resolution Selector.

The rest are KoboldCpp request knobs (URL, token allowance, temperature, penalties). max_response_tokens of 0 uses the selected profile's default.

Outputs: amplified_text (→ text encoder), secondary_instructions, and width/height (INT) for latent sizing.

Installing it

One clone for the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/tiko13/ComfyUI_PromptStudio

Restart ComfyUI, or search "ComfyUI_PromptStudio" in ComfyUI Manager. No pip dependencies - the node just needs a local KoboldCpp server at http://localhost:5001 (default), Ollama at 11434, or llama.cpp at 8080. For KoboldCpp specifically: use 1.117.1+, enable Use Jinja in the server, and restart it - the node refuses to run with generic chat formatting rather than silently degrading.

Where people get burned

• "Use Jinja" not enabled. The node errors out on purpose. Enable it in KoboldCpp and restart the server.
• VRAM contention on one GPU. This is the classic KoboldCpp+ComfyUI problem: the LLM and the diffusion model fight over your card. Prompt Studio's shared-GPU handoff (which unloads the LLM before ComfyUI queues) requires KoboldCpp Admin Mode with an Admin Directory. Without it, expect OOMs or a red status dot. Don't enable "Keep models loaded" unless the two run on separate GPUs.
• Double amplification. If this node sits in a [PS] workflow, Prompt Studio converts it to a Prompt Slot in the temporary snapshot it queues - so Studio never amplifies twice. On a normal ComfyUI queue, the amplification happens exactly once, as you'd expect.
• Wrong profile → wrong vibe. A tag-based anime model doesn't want your natural-language photoreal phrasing. If output looks off, switch model_profile before touching samplers.