VNCCS Prompt Designer

Random prompts are the cheapest source of image variety in ComfyUI, and {brown|blonde|black} is old A1111 wildcard syntax that most people who arrived after 2024 never learned. The catch was never the syntax - it's that a wildcard node hands the resolved text straight to the encoder, so you can't see what it picked until the image comes out, and afterwards you can rarely tell which prompt made which file.

VNCCS Prompt Designer fixes that and then keeps going: a prompt workbench inside one node, with tabs of drafts, reusable fragment cards you drag in, {a|b|c} variants, weighted and nested choices, variables, If/Else conditions, and a preview of the exact text the next queue will send. The author is AHEKOT, the VNCCS project behind Pose Studio and UniCanvas - well known in the character/pose corner of ComfyUI, though this node is new enough that there's no chatter about it yet.

How it works

The Python side is almost insultingly simple: one input, one job - parse a JSON blob and render text.

That blob is node_state. It's a STRING, hidden behind the custom DOM widget you actually interact with - the widget serializes your tabs, block cards, conditions, seed and sampling mode into that single string so the whole composition travels inside the workflow. No CLIP, no model, no image.

Rendering is done with dynamicprompts in process, using a deliberately bounded sampling context: cached, budgeted, capped at 256 KB of state and 64 KB of resolved prompt. Each block insertion samples once and conditions reuse that resolved value instead of re-rolling it, so an If/Else on a block can't randomly disagree with the block visible in the output. Blocks sample in random or cycle mode, and the cycle position persists across runs.

The seed lives in the widget, not your KSampler. Leave it fixed and every queue sends the same prompt; set After generate to Randomize and each queued run draws fresh. Batch of 20 suspiciously identical? That's the switch you forgot, and re-rolling the sampler seed won't change it.

The input and the outputs that matter

You'll set exactly one thing: the composition itself, through the widget. Seed, After generate, per-block Variants and the random/cycle sampling mode are widget controls, not sockets.

Outputs are all STRING, named prompt through prompt16. In practice you wire prompt into a CLIP Text Encode, a prompt-combining node, or anything else that eats a string - the same as any other prompt node. The extra sockets only matter if you build multi-prompt fragments: one composition can emit up to 16 separate strings, which is how you drive a positive/negative pair or parallel branches from one authored prompt. The widget exposes sockets as fragments need them and preserves existing connections when you edit.

Practical tip: prompt resolves at execution time, so wire it into a Show Text or Save Text node to record what each run actually sent. That's the standing workaround for wildcard workflows everywhere.

Installing it

ComfyUI Manager → Custom Nodes Manager → search VNCCS Utils → Install → restart. Or manually:

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

Now the honest warning: requirements.txt is the pack's requirements, not this node's. Installing it pulls opencv-python, av, kornia, pytorch-lightning, roma, timm, color-matcher and friends, because the same repo ships 3D Factory, UniCanvas and Pose Studio. Prompt Designer itself needs dynamicprompts>=0.31,<1 and little else. The TripoSplat weights only download if you open 3D Factory's Model setup, so you're not forced into gigabytes - but you do share a Python environment with a much bigger project. On a machine already full of node packs, that's the real cost.

Where people get burned

• Missing dynamicprompts. Requirements not installed throws a clear RuntimeError: Prompt Designer requires dynamicprompts the moment the workflow runs. Fix: pip install "dynamicprompts>=0.31,<1" into ComfyUI's environment and restart.
• A dead widget. The editor talks to backend routes registered at startup. If the panel is blank, check the console for [VNCCS] Failed to register Prompt Designer API - usually a route clash - then hard-reload the browser. Custom DOM widgets cache aggressively.
• Accidental re-rolls. Setting After generate to randomize without realising it makes every queued run a new composition, so a 30-image batch is 30 different prompts, not 30 takes on one.
• Backups, and huge libraries. Your block library lives in PromptLibrary/ inside the pack folder, with revision history and backups - updating in place is fine, deleting the pack directory takes it with you. State is also capped at 256 KB and the resolved prompt at 64 KB, so a composition with thousands of fragments errors out instead of quietly truncating.