Gemma4 Prompt tool.

Gemma4 Prompt Engineer — ComfyUI Node

A local AI prompt generation node for ComfyUI powered by Gemma 4 running entirely on your own hardware via llama-server. No API keys, no cloud, nothing leaves your machine.

You type a rough idea. The node writes a fully structured LTX 2.3 prompt — camera moves, lighting, layered audio, physical detail, dialogue — without you needing to know any of that vocabulary. Put in a little, get a full scene. Put in a lot, it cleans it up and makes it coherent.

---

What it generates prompts for: LTX 2.3 • Wan 2.2 • Flux.1 • SDXL • Pony XL • SD 1.5

---

Key features:

80 environment presets — from a rain-soaked Tokyo street to an active volcano observatory to a fake taxi parked in a quiet layby. Each one pre-loaded with location detail, lighting, and sound design. Pick one and build your scene on top.

14 animation presets — SpongeBob, Rick and Morty, BoJack Horseman, Avatar, Scooby-Doo, Looney Tunes, and more. The model already knows the characters, locations, and tone. Just use the names.

POV mode — Off / POV Female / POV Male. Camera becomes the viewer's eyes. Hands visible in frame, physical sensations described, no third-person cutaways. Works with any scene including explicit content.

Dialogue system — auto-detects singing, ASMR, and talking from your instruction. Generates actual spoken words in quotes, contextually relevant to your scene. Back and forth exchanges, moaning, dirty talk, whispered ASMR — whatever fits.

Full NSFW support — anatomically direct, no euphemisms, no sanitising. Explicit scenes, undressing sequences, sex positions, physical response described correctly. Pairs well with LoRAs like the Rogue AI Missionary/Cowgirl — describe the action, wire in your start frame, let I2V do the rest.

I2V grounding — connect a start frame image and toggle use_image. The node describes what's in the image coming to life. Vision enabled automatically when mmproj file is present.

Auto-setup — includes a bat file that detects or installs llama-server, downloads the recommended GGUF, and gets you running in one click.

PREVIEW / SEND workflow — generate and read the prompt before committing. PREVIEW halts the pipeline, SEND outputs and kills llama-server to free VRAM for generation.

---

Recommended model: Gemma 4 26B heretic Q4_K_M (~15.7GB) — uncensored, vision capable with mmproj

Requires: Windows, CUDA GPU, llama-server (auto-installed), 16GB+ VRAM recommended

Quick Install

1. Download and drop the node folder into:

ComfyUI/custom_nodes/

1. Run setup_gemma4_promptld.bat — it will:

   • Detect or auto-install llama-server to C:\llama\

   • Prompt you to download the recommended GGUF

   • Install Python dependencies

2. Restart ComfyUI — the node appears under LoRa-Daddy/Gemma4

3. Hit PREVIEW — the node handles everything from there

---

Manual downloads if you prefer:

llama.cpp CUDA build: https://github.com/ggml-org/llama.cpp/releases/tag/b8664

Recommended model (text + vision): https://huggingface.co/nohurry/gemma-4-26B-A4B-it-heretic-GUFF → grab gemma-4-26b-a4b-it-heretic.q4_k_m.gguf + gemma-4-26B-A4B-it-heretic-mmproj-bf16.gguf → drop both in C:\models\

Alternative (text only, slightly larger): https://huggingface.co/amarck/gemma-4-31b-it-abliterated-GGUF

---

GGUFs always go in C:\models\ — the node scans that folder on startup and populates the dropdown automatically.

Github link
workflow
Reddit post