Generate Scenery Prompt

The scenery generator that never repeats

Scenery is the part of a prompt most people write last and worst. You've nailed the character, the pose, the outfit - and then it's "park, sunny" and hope. Generate Scenery Prompt (class SceneryPrompt) is the fix for exactly that. One node rolls a random setting from a bundled library of six-thousand-plus hand-written scenery prompts and hands you the finished text. One seed, done.

It comes from clouddreamfly's ComfyUI-PromptWrapper pack, and here's the thing to know up front: that pack's README is all about downloading translation models, but this node wants none of them. Everything it reads is already inside the repo.

How it works

Run it and the node opens scenery_datas.jsonl in the language you picked - Chinese or English, 6,123 entries versus 6,249 - and grabs line seed % total. Deterministic, like most of this pack: lock the seed and you get the same backdrop forever; nudge it and you roll a new one. The library is genuinely varied, too. Lawns and lunar surfaces, fireworks over a night city, Yellowstone, a cyberpunk skyline, a café, a dungeon - the kind of range that makes a batch of images feel like a coherent set instead of a rerun.

The inputs that matter

• language - Chinese or English. Flips the whole dictionary.
• seed - the pick. Your entire variation story lives here; wire it to a randomize node and you'll never copy-paste a background again.
• preset_prefix and preset_suffix - free-text blocks that get stuck around the generated scenery. They're your way to force a mood ("cinematic lighting", "dramatic low angle") without touching the random part; the rolled line lands between them.
• input_prompt (optional) - a string you can feed in from elsewhere in the graph; it's prepended ahead of everything.

The output is a single text string, ready for a CLIP Text Encode (positive) node.

Wiring it in

String goes into CLIP Text Encode. Because the output is plain natural language - no (tag:1.3) weight syntax, no booru garble - it plays nice with every encoder family, including the LLM-encoded checkpoints that silently throw away attention weights. If you want the setting to vary per image but keep the character fixed, keep the character tags in your main prompt and let this node own the backdrop. With a fixed seed it's fully reproducible, which makes it great for batches: every image in the run shares one coherent scene.

Install

ComfyUI Manager, search "ComfyUI-PromptWrapper", install, restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/clouddreamfly/ComfyUI-PromptWrapper

Then restart ComfyUI. No pip requirements and no model files for this node. Ignore the pack's translation-model download instructions unless you're actually using its Prompt Translation nodes - a 1.8B HY-MT model is a big download you don't need for scenery. If you like this one, the same pack has the same pattern for portraits, lighting, artists and styles; they all work the same way.