PromptDeckComposer

Type "forest ritual" into PromptDeckComposer and it hands you back a full scene description - a ritual priestess made of herbs surrounded by floating fragments in an abandoned temple, lit with soft volumetric light, in the voice of an artist you didn't pick. That's the whole pitch: the node treats prompting as a slot machine you pull instead of a paragraph you write, and it's aimed at the "I have no idea what I want, show me options" mood. For that it's genuinely handy - especially if you're feeding a modern model like Flux that wants dense natural-language descriptions rather than a comma bag of tags.

What it actually does

Under the hood it's a pool-picker with a random number generator and a pen. On startup it reads composer_deck.json, which ships with the pools: 25 primary subjects, 54 subject modifiers, 24 secondary subjects, 25 environments, 21 each of lighting/color/texture/mood/render layers, 42 artists (Caravaggio to Beeple), and 21+ full-sentence "style blocks" describing atmosphere, camera, and concept.

When you queue it, compose() builds a random.Random from your seed and walks the pools, picking each field with a probability that scales with CHAOS_LEVEL (0–1, default 0.5). Low chaos keeps the output tight and conventional; high chaos starts throwing in secondary subjects and style blocks you'd never think to combine. The picked values then get assembled by a prompt writer - there's a neutral, concise, cinematic, poetic, and technical TONE_PRESET that each phrase the same values completely differently, which is a nice touch. What comes out the prompt output is a single STRING you wire straight into your sampler's positive prompt.

The inputs that matter

The node looks intimidating - there are about seventy widgets. You really set a handful:

• randomize - the master switch. On (default) and it re-rolls on every queue, using time.time_ns() as the seed when seed is 0, so you get a different prompt each run. Off, and it uses exactly the values you've selected in the widgets.
• USER_PROMPT - your fixed anchor. It gets appended to whatever randomness produces, and unlike everything else it's never re-rolled. This is how you say "okay, but make it about this."
• CHAOS_LEVEL - how far the roll can wander. The one dial most people end up touching.
• The LOCK_ toggles* - one per field (LOCK_PRIMARY_SUBJECT, LOCK_ENVIRONMENT, LOCK_ARTIST_1, and friends). Flip the lock on the bits you like and let the rest keep rolling. This is the "evolve, don't rebuild" workflow the README talks about, and it works better than it has any right to.
• OLLAMA_ENABLE - the optional LLM expansion, off by default.

The Ollama bit

The built-in LLM hook is the most overhyped part, and also the safest one. With OLLAMA_ENABLE on, the node POSTs the composed prompt to http://127.0.0.1:11434/api/generate (the OLLAMA_URL, and OLLAMA_MODEL must name a model you've already pulled) and replaces the output with the model's expanded prose. The kicker: if Ollama isn't installed, isn't running, times out, or the model doesn't exist, it silently falls back to the structured prompt and reports the reason in OLLAMA_STATUS. The node never hard-fails on it. Where people get burned is expecting it to just work - Ollama is a separate program you have to install and run, and the default system prompt is written for Flux-style sentence prompting, so an SDXL checkpoint isn't going to love the output the same way.

Installing

No pip dependencies, no model downloads - the whole thing is pure Python stdlib plus its bundled JSON. Clone it in:

cd ComfyUI/custom_nodes
git clone https://github.com/chkeeho80/PromptDeckComposer

Then restart ComfyUI. You'll find it under prompt/PromptDeckComposer → PromptDeckComposer. It's new and obscure enough that ComfyUI Manager may or may not have it listed yet; the clone is the reliable path. The only optional extra is Ollama itself, which you install separately.

Gotchas

Two things trip people up. First, with randomize on and seed left at 0, your queue is different every single run - that's the feature, but if you land on a prompt you love and want it back, set a fixed seed or switch randomize off before you queue again. Second, the node's web UI does a lot of front-end work - reroll buttons, a saved-results list, refine/restyle/reimagine menus - and it stashes recent results and thumbnails inside its own folder (custom_nodes/PromptDeckComposer/). That's not a bug; it's just where your "kept" prompts live if you ever want to back them up.

If you want one perfect result, this isn't the node for you. If you want fifty interesting starting points in an afternoon, it's quietly excellent.