Muse Advanced (GGUF)

"Muse Advanced (GGUF)" is the plain Muse brainstorming node with three more constraint dials and the full llama.cpp sampler panel. Same concept-inventing engine, same prompt + json outputs - just more control over what comes out and how it's sampled. You move to it when the five-dropdown version keeps landing in the same visual neighborhood and you need to force a look.

The three new creative dials are the reason to upgrade:

• lighting (25 options) - Golden Hour, Neon Glow, Bioluminescent, Fluorescent / Clinical…
• color_palette (22 options) - Teal & Orange Cinematic, Pastel Dreamlike, Monochrome Green (Night Vision)…
• era (19 options) - Contemporary, 1980s, 1950s Atomic Age, Victorian / 1880s, Far Future…

With those you can go from "moody photo" to "moody teal-and-orange bioluminescent 1980s photo" in two clicks, which is exactly the specificity a diffusion prompt wants.

The rest of the panel

It inherits genre, mood, shot_type, artist, subject, free_hint, aesthetic_bias, output_format, reference_mode, and the optional image/video style references from Muse. On top of that:

• device - auto or cpu.
• temperature / top_p / repetition_penalty - sampler defaults here are a bit looser than the prompt nodes (0.85 / 0.92 / 1.1), which fits the creative use: more temperature means more varied concepts. Drop temperature when you want consistency.
• gpu_layers - -1 puts everything on the GPU; lower it to trade speed for VRAM.
• ctx / n_batch / top_k / pool_size - the performance and search knobs. Leave them alone unless you're chasing a specific problem.
• frame_count (1–32) and image_max_tokens - both for the video reference path: how many frames get analyzed and how big a token budget the vision side gets.
• max_tokens (default 1024), keep_model_loaded, seed.

Outputs stay the same: prompt (STRING) with the extracted generation prompt, and json (STRING) with the full concept.

Install

Pack-level install, same as everything in ComfyUI-VLMPrompt:

cd ComfyUI/custom_nodes
git clone https://github.com/mexxmillion/ComfyUI-VLMPrompt.git
pip install -r ComfyUI-VLMPrompt/requirements.txt

Restart ComfyUI, and remember the pack-wide GGUF caveat: llama-cpp-python needs a multimodal build with the VLM chat handlers or the GGUF nodes - Muse included - won't load. Models auto-download to ComfyUI/models/llm/GGUF, and the dropdown pools are data-driven from muse_data.json if you want to curate them.

Common issues

• Concepts still feel random - that's the design; the era and color_palette dials are what narrow it. Set them.
• Repetitive concepts - raise temperature a notch or two (it's a creative node, 0.9+ is fine) instead of fighting the seed.
• VRAM tight - lower gpu_layers before you reduce quality; a 4B GGUF is small, but it adds up in a big workflow.