MD: Scene Genius Autocreator

"Scene Genius" sounds like an image-storyboarding tool. It isn't - this is a creative brief generator for ACE-Step audio workflows, and knowing that up front saves you five minutes of confusion. You feed it a concept ("Cyberpunk neon lights rain" is the default), it hands you a complete, coherent package: genre, vocals, lyrics, duration, step count, and a handful of technical ADG parameters the ACE-Step pipeline wants. One node in, everything the rest of the audio graph needs, out.

The three modes

execution_mode is the whole design:

• Fast (No LLM) - instant, picks from the node's internal dictionary libraries. No network, no latency. Use it for drafts.
• Hybrid (Smart LLM) - the LLM selects and combines curated options from the libraries. Best balance; the author recommends it for production runs.
• Full AI (Creative LLM) - the LLM writes completely new content. 5–30 seconds of API latency, and it needs a local LLM server running.

That LLM is either Ollama (http://localhost:11434, default model llama3:8b-instruct-q8_0) or LM Studio (http://localhost:1234). No cloud keys - it's all local, which is the right call for this crowd.

How the outputs get made

Behind the scenes it blends three sources: internal genre/vocal/lyrics libraries, your own .txt files dropped into the node's wildcards/genre, wildcards/vocal, and wildcards/lyrics folders (selected via load_genre_file, load_vocal_file, load_lyrics_file), and wildcard {A|B} expansion. quality_preset sets steps from Draft (60) all the way to Monster (2000) and syncs the technical ADG parameters to match. force_instrumental zeroes out the vocals and writes [Instrumental] into lyrics.

The override hierarchy is clean: genre_input, vocal_input, and lyrics_input beat everything - files, LLM, Fast mode - so you can hard-force a value while still letting the node generate the rest.

The outputs

You get GENRE, LYRICS, VOCALS (strings), DURATION (float), STEPS (int), GLIDE_POWER, MASTER_STR, ADG_ANGLE (floats), LEGACY (boolean), SEED (int), and STATUS - an IMAGE dashboard that summarizes the run. The natural wiring: feed these into MD_WorkflowContextBus to bundle them, or run GENRE/LYRICS/VOCALS straight into your ACE-Step text encoding and STEPS into the sampler.

The honest take

The default concept is very audio-culture, the library files are genre-flavored (dnb.txt, lofi_chill.txt, industrial_hard.txt), and the whole node is clearly built by someone deep in the ACE-Step music scene - which is exactly what it's good at. It is not a general-purpose scene/storyboard generator for images. If you're an audio person on ACE-Step, this is one of the best "type a vibe, get a full production brief" nodes out there. If you're not, the same pack's UniversalWildcardOrchestrator is the more neutral tool. Keep it in Fast mode until you've confirmed Ollama actually responds - nothing wastes time like a 30-second timeout on a run that used to be instant.

Installing it

cd path/to/ComfyUI/custom_nodes
git clone https://github.com/MDMAchine/ComfyUI_MD_Nodes.git
cd ComfyUI_MD_Nodes
pip install -r requirements.txt

Or search MD_Nodes in ComfyUI Manager and restart.