Scene Weaver - Fantasy

Drop the node on the canvas, change nothing, hit queue. You get a sentence:

A fantasy scene set in a frostbitten mountain pass. A cave troll is climbing toward a roost on the cliff face. It is hulking, covered in matted grey-green hide. It has a single shattered tusk and four clawed fingers.

That's the pitch. No wildcard files, no LLM call, no API key. Just prose your encoder can follow.

Why you'd reach for it

The usual way to get variety is a wildcard pack, and the usual result is a comma pile whose parts fight each other: a sleeping dragon that's meanwhile mid-pounce, a wizard tower nine hundred storeys tall. The model spends the image reconciling the contradiction, and you get mush.

SceneWeaver's bet is the opposite: a prompt whose every word agrees with every other word leaves the model free to draw one believable scene. It's a prompt writer that knows what it's describing - a longship is clad in its material, a dragon is covered in one, and a thing with no surface is never handed one to describe. That grammar is data, not a template.

What it is not: the pack writes subject description and nothing else. No lighting, lens, film stock or mood. You bring the style, and in exchange it composes cleanly with a style LoRA or a style prefix instead of arguing with one.

How it works

Every genre is the same engine plus a data module, which is why Fantasy, Horror and Sci-Fi are three node pairs and not three codebases. Four things matter:

• Coherence is enforced while it draws. kind scopes every other field on that slot - a longship asks for open water, a wagon asks for a road, a pixie is six inches and a hill giant twenty feet, nothing burns under the sea.
• Locking is the workflow. Every dropdown is Random / a value / None. A value pins that field, Random rolls around it, None omits it. A locked field is never cut and never overruled.
• There's a detail budget. Four fully-described entities is roughly a 400-token prompt, and encoders drop detail across one - mush exactly when you asked for the most. Slot 1 is described in full, support slots briefly, and anything cut is written back to None. Locks are never cut.
• The node face tells you what you actually rolled - two lines naming the scene and its rough token count, because every widget says "Random" and "Random" is not a description.

Everything is seeded, so a fixed seed reproduces a scene exactly. The node re-runs on every queue by design, which is what makes randomize behave the way you expect.

The controls that matter

• seed - leave the control on randomize to explore, flip to fixed once you like something.
• entity_count - 1 to 4, and the authority on how many subjects exist: slots past it are dropped, slots inside it are filled even if their widgets still say None. Two is the sweet spot; three and four read as a crowd.
• scene_filter - Any, Peaceful or Conflict. Peaceful keeps weapons and hostile actions out of the description entirely. A value you locked is always kept.
• set_all_fields - a one-shot bulk edit, not a mode: Randomize all turns every None into Random, Clear all does the reverse, and it snaps back to Off. Clear all plus one locked environment is how you get a pure landscape.
• environment, then per slot the morphology fields: entity1_kind (which scopes the rest), subkind, scale, condition, form, material, primary_color, accent_color, markings, surface_detail, appendages + count, emitters + count + emitter_color, armament + count, sensors + count, aperture, extras, and entity1_situation - an event in progress, never a pose. Slots 2–4 expose only kind, subkind, scale, condition and situation.
• relation_1_2 and friends, plus the _position variants - opt-in, default None, rendered only when both entities exist.
• context - one secondary detail for scale and story. A thing in the world, never a lighting instruction.
• Four optional entity_1_in … entity_4_in sockets, for a SCENE_ENTITY payload.

Outputs: prompt_text goes into your text encoder (and into a Show Text node, so you can read it), and prompt_json is structured scene data - don't feed that one to an encoder.

Install

ComfyUI Manager: search SceneWeaver. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/EnragedAntelope/comfyui-sceneweaver

Keep the folder named comfyui-sceneweaver and restart. No dependencies - standard library only, no requirements.txt step - but it needs ComfyUI 0.25.0+ and Python 3.10+. The nodes appear under conditioning/sceneweaver.

Your own vocabulary goes in a user_options.json beside the pack; it's gitignored, so it survives a git pull, and fantasy values live in a "fantasy" section.

python scripts/builtin_options.py subkind --kind dragon --pack fantasy

Where people get burned

You forgot the style. "The images look flat" is the pack doing its job - with no rendering frame in the prompt, models draw bare content. The maintainer's own graphs prefix every prompt with:

cinematic film-quality slightly-grainy analog film still.

Four entities, one subject's worth of attention. That's the model, not the node.

A locked slot that still reads short. A structure or vessel carries only its silhouette; a creature carries its whole component list, so a slot full of locks can come out shorter than you expected.

user_options.json additions that do nothing. Fields like form, material, markings and situation are overridden by every kind, so a _default pool for them is never read. Put those values under a kind name - the console warns you if you get it backwards.

Old ComfyUI. Below 0.25.0 you get an import failure on comfy_api.latest, not a node.