The influence stack

The panel on the right is the whole product in one column: which layers are contributing to this image, how much each one weighs, and exactly which words each one put in the prompt.

How the stack resolves

Given the entity you have selected, Wobu walks outward and collects sources in a fixed order — outermost first, subject last:

#	Layer	Source	Contributes
1	Style	The Art Style entity	Medium, rendering, line quality, lighting model, global negatives
2	World	The World Canon entity	Era, tone, tech and magic level, material vocabulary
3	Ancestry	the `species_of` chain	Body plan, anatomy, skin and hide, scale
4	Culture	the `member_of` chain	Costume, ornament, iconography, weapon language
5	Place	the `located_in` chain	Climate, architecture, ambient light, wear
6	Subject	the entity itself, plus anything `related_to`	Everything specific
7	Shot	the controls below	Framing, pose, output preset, aspect

Art Style and World Canon are seeded into every stack whether or not anything links to them. A entity's parent is an implicit edge of full weight, so a district inherits its city which inherits its region. Each source collected becomes one card in the inspector, and the card says why it is there.

Fragments, not paragraphs

A layer does not contribute a block of prose — it contributes fragments. Each fragment knows its source entity, which description section it came from, its weight, and where it should be routed:

prompt · negative · style_ref · structure_ref · palette · moodboard_only

A fragment's weight is the product of three things:

link weight  ×  section priority (per preset)  ×  your slider

Section priority is why presets matter. A material study boosts materials and drops silhouette; a turnaround does the reverse. The same description produces a different prompt depending on what kind of sheet you asked for.

Compiling

Resolve and collect. Walk the stack, gather every fragment.
Filter. Drop muted layers and anything marked mood-board-only. Mood-board fragments are not reported as dropped, because they were never candidates.
Score and sort by weight, stable within layer order.
Budget. Trim to the model's usable prompt length, dropping the lowest-weight fragments first — and report what was dropped rather than truncating in silence. Images are budgeted separately and more tightly, per bucket, against the provider's declared limits.
Emit. Fragments go out in arrival order — layer by layer, subject last, the shot's framing after that — rather than in the order they were dropped. The negative prompt is built from every layer's Never section; style references go to style conditioning, structure references to structure conditioning, palettes to colour conditioning.
Snapshot. The entire resolved stack — every weight, every mute, every dropped fragment — is saved with the generation.

Reading the compiled prompt

Every fragment keeps its layer identity all the way to the screen. Show sources turns the compiled prompt from prose into tinted spans, one colour per layer; hovering one names the layer and its weight, and clicking it opens the entity that produced it.

style     painterly gouache, heavy brush texture, strong single key light, deep shadow
world     late iron age, forty years post-eruption, scavenged and repaired materials
ancestry  tall narrow-shouldered digitigrade humanoid, four-jointed legs, ash-grey hide
culture   ember guild kiln-glaze plate over oiled leather, rope fastenings, collarbone signet
place     cinder bay harbour light, salt-bleached timber, ash haze
subject   Kael Vantris — armour cut for a broader frame, cinched with rope; burn scar left jaw
          to temple; signet ground flat; unlit ashglass lantern at the belt

Attribution is not a debug feature. It is the main feedback loop for learning to write good upstream notes. When an image comes out wrong, the tint tells you which entity to go and fix — and over a few weeks it teaches you what belongs at which altitude far more effectively than any documentation could.

Both channels — the prompt and the negative — show their fragment and character counts and can be copied.

What was left out, and why

Beneath the prompt are up to two lists, and they are worded differently on purpose:

Turned down — a slider or a link weight is at zero. Raise it and these come straight back; nothing has been edited.
Did not fit — these were the lightest fragments when the prompt ran over budget. Write leaner notes upstream, or weight these sources higher so something else goes first.

If everything was cut, Wobu says so plainly rather than quietly sending an empty prompt. If a single fragment is longer than the whole budget it is sent long rather than sent blank, and the panel tells you by how much.

Layer cards

Each card in the stack gives you:

Thumbnail and layer name	Which of the seven layers this is, and what the source looks like.
Source entity	With an Open source button that jumps to it so you can edit it.
Counts	`4 text · 2 images · 1 sent · 1 dropped` — what this layer actually contributed.
Weight slider	Scales every fragment from this layer, from 0 to 1.
Mute toggle	Removes the layer from this generation entirely.
Fragment list	The exact text and images it contributes, verbatim, each with its own weight.

Inspector edits are per-generation Muting and reweighting here affect this generation only — they never edit the world. That separation is deliberate and absolute. If you find yourself muting the same layer on every render, that is not a workflow, that is a signal: go and change the entity, or change the link.

Link weights, by contrast, are part of the world. Set them from the entity's Relations tab when a character should genuinely be only loosely of its culture.

Practical steering

Style is overpowering the subject	Drop the Style layer to about 0.6 for this shot. If it happens constantly, your Art Style notes are too specific — move detail down to World or Culture.
Character keeps looking generic	Check the Subject layer's fragment count. A thin subject layer usually means the notes are backstory rather than description — Enhance drops anything not visible on the body.
Two characters look identical	Their subject layers are not saying enough that differs. The ancestry and culture layers are doing all the work, which is exactly what they are supposed to do.
Something anachronistic keeps appearing	Add it to Never on the highest layer where it is always wrong — usually World Canon. It compiles into the negative prompt for every descendant.
Fragments are being dropped	You are over prompt budget. Mute a layer you do not need for this shot, or find the duplication — usually a child restating what its parent already said.
A reference is not being used	Check its role, and check the provider. `mood` is never sent anywhere, structure roles are inert on providers that do not accept them, and the shipped ComfyUI workflows take no images at all.
The negative prompt vanished	Some image models have no negative conditioning. Wobu withholds it and reports that rather than pasting your Never list into the positive prompt.

The snapshot

Every generation stores its influence snapshot — the exact resolved stack, weights and all, at the moment you pressed Generate. That is what makes a result reproducible six months later, after the world has moved on and three of those layers have been rewritten.

Open a result and Wobu shows the recorded stack next to today's, and names the difference: The prompt has changed, or The parts that can be compared are unchanged. Run these settings again sends exactly what the receipt holds, without reading today's world at all.

← References and assets · Generating →

This page is rendered from its source in the repository — the same file the application reads for its built-in guide.