MD: Wildcard Prompt Builder

The Universal Wildcard Orchestrator is a plain wildcard expander. This is its bigger sibling: it expands wildcards and produces a structured, machine-readable prompt bundle - genre tags, vocal chains, lyrics, duration, BPM, key/scale, time signature - plus a single INTERPRETED_PROMPT formatted for whatever model family you're targeting. You give it a concept, it hands your graph a complete creative brief.

Pick your target

output_target decides the vocabulary and format:

• Audio (ACE-Step) - genre, vocals, lyrics, BPM, key outputs for ACE-Step music generation.
• Image (SD/Flux) - an image prompt on INTERPRETED_PROMPT.
• Video (Wan/Mochi) - a video prompt on INTERPRETED_PROMPT.
• All Three - every output populated for unified media workflows.

For ACE-Step audio, ace_step_mode (on by default) upgrades output to the rich, Gemini-style prose captions ACE-Step 1.5 wants, and auto-strips BPM/key metadata out of the caption so the model doesn't try to read numbers as text. descriptive_mode swaps comma-tag output for full sentences when you want it.

How generation happens

generation_mode picks wildcard (instant, from internal libraries and your .txt files), llm (Ollama writes fresh content), or hybrid (the LLM picks smart combinations from curated lists). The concept field is your creative brief - be specific about genre, mood, and sonic elements and the outputs get markedly better.

The Ollama integration is worth understanding before you touch it: ollama_enabled is the master switch, and if it's False the node falls back to wildcard generation instantly and never probes the network. If you don't run Ollama, set it False - that's the author's own advice, and it's good. When it's on, you get the full sampling-knob set (temperature, top_k, top_p, min_p, keep_alive, llm_timeout). orchestrator_mode does a single-shot LLM pass for a fully coherent brief - ACE-Architect is the recommended mode for ACE-Step 1.5, and it's what outputs that big YAML_CONFIG string.

The outputs, and which ones to grab

The full list is intimidating: GENRE_TAGS, VOCAL_TAGS, LYRICS, DURATION_STRING, DURATION_FLOAT, BPM, TIME_SIG_STR, TIME_SIG_COMBO, KEY_SCALE_STR, KEY_SCALE_COMBO, SEED, TEXT_PREVIEW (a rendered dashboard image of the brief), YAML_CONFIG, and INTERPRETED_PROMPT.

But you'll use two or three. Audio users: GENRE_TAGS, VOCAL_TAGS, LYRICS into the ACE-Step text encoding, and YAML_CONFIG into the MD sampler's yaml_settings_str override - that's the whole point of the YAML_CONFIG output, a ready-made config string. Image/video users: switch output_target, grab INTERPRETED_PROMPT, run it into a CLIPTextEncode, done.

The honest take

The pin count is the worst thing about this node. Everything else is genuinely well designed - the ollama_enabled kill-switch, the ACE-Step-aware caption stripping, the YAML config output that closes the loop with the pack's samplers. If you're building ACE-Step audio workflows, this is arguably the pack's best node. If you're image-only and have zero interest in audio, the same functionality lives in simpler, more neutral packs, and the audio-centric defaults here will just be noise. But as a one-node "concept in, structured brief out" pipeline, it's the real deal.

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. Ollama is an optional extra - the node works without it as a pure wildcard engine.