Files

T

MNeMoNiCuZ b269aab0ed Update ideogram4_random_prompter.md

2026-06-14 12:25:53 +02:00

19 KiB

Raw Permalink Blame History

🎲 Ideogram 4 Random Prompter

An experimental companion to the Ideogram 4 Prompt Builder. Instead of drawing regions by hand, this node automatically generates a complete composition from scratch by scattering weighted-random regions across the canvas and filling each with descriptions built from real dictionary words.

Requires the wonderwords package:

pip install wonderwords

What makes this node unique

Every descriptive word in the caption is pulled live from the wonderwords dictionary — there are no hardcoded or curated word lists. This means every run produces genuinely novel, varied compositions:

Random region count — minimum to maximum regions, weighted across size tiers.
Random placement — centers scattered, clustered, arranged in spirals/bursts/grids, or pushed to compass edges.
Random sizes — background (70–100% canvas area), large (34–62%), medium (17–40%), small (5–17%).
Random descriptions — each region gets a randomized content-word budget filled from the dictionary.
Random styling — aesthetics, lighting, medium, color palette, and harmony are all randomized per run.
Scene framing — descriptions can be bare words or woven together with articles and spatial connectors into one coherent sentence.
Freeform regions — some regions can drop their hard bounding boxes to blend softly into the scene.
Text regions — some elements can render as actual in-image text (drawn words) instead of objects.

The output matches the Ideogram 4 Prompt Builder exactly — same JSON format, same preview, same bounding-box style — so the two nodes can be swapped freely or used side by side.

Canvas settings

Setting	Type	Description
`width`	INT	Canvas width in pixels. Ideogram 4 prefers multiples of 16. Affects aspect ratio and the pixel grid for bboxes.
`height`	INT	Canvas height in pixels. Ideogram 4 prefers multiples of 16.

Core randomization: regions and size tiers

The node randomly chooses how many regions to generate, then weighs each one across four size tiers.

Setting	Type	Default	Description
`seed`	INT	0	Random seed. The same seed + same settings always produce identical output. Change the seed to roll a completely new composition.
`region_count_min`	INT	10	Minimum number of element regions (inclusive). Set equal to max for a fixed count.
`region_count_max`	INT	20	Maximum number of element regions (inclusive).
`background_weight`	FLOAT	0.4	Likelihood that any region is a background tier (covers ≥70% of canvas area, acts as a base layer).
`large_weight`	FLOAT	0.6	Likelihood that a region is large (34–62% of canvas per axis).
`medium_weight`	FLOAT	0.4	Likelihood that a region is medium (17–40% of canvas per axis).
`small_weight`	FLOAT	0.2	Likelihood that a region is small (5–17% of canvas per axis).

All four weights are summed and each region picks a tier proportionally. Set all to 0 to get equal weighting (25% each).

Region descriptions: words and content budgets

Each region gets a content-word budget — a random integer between min and max, filled with real dictionary words pulled live.

Setting	Type	Default	Description
`region_word_min`	INT	5	Minimum content words per region. Connector/article words (in scene mode) are NOT counted.
`region_word_max`	INT	15	Maximum content words per region.
`word_length_bias`	INT	0	Preferred dictionary word length (characters). 0 = no preference; >0 = bias toward that length. Example: 4 favours short punchy words, 11 favours long ornate words.
`word_length_randomness`	INT	2	Spread around `word_length_bias`. Words are drawn from `[bias - this, bias + this]`. Ignored when bias is 0. Example: bias 8, randomness 2 → words 6–10 chars.

Scene framing: pure vs. scene mode

OFF (pure mode)

Each region's description is a bare comma-separated list of dictionary words:

vivid tower, hollow stone.

Maximum randomness, but Ideogram tends to render this as a collage or asset sheet of separate items.

ON (scene mode)

The same dictionary words are woven together with articles (a / an) and spatial connectors (beside, near, behind, above, below, etc.) into one continuous sentence:

a vivid tower beside a hollow stone against an amber cloud.

This tells Ideogram it is one coherent scene, so a photograph actually looks like a photograph instead of a grid. The connector and article words are structural and do NOT count toward region_word_min/max.

Setting	Type	Default	Description
`scene_framing`	BOOLEAN	True	False = pure (bare words); True = scene (woven sentence).

Placement and arrangement

Regions are positioned either by a positioning bias (scattered, center-weighted, edge-weighted, grid-aligned, or a compass direction) or by a structured arrangement (spiral, burst, grid).

Setting	Type	Default	Description
`positioning_bias`	ENUM	`scattered`	How regions tend to cluster: `scattered` (anywhere), `center_weighted`, `edge_weighted`, `grid_aligned`, `random_weighted`, or compass directions (`north`, `south`, `east`, `west`, and diagonals). Ignored when `arrangement` is not `none`.
`arrangement`	ENUM	`none`	Structured placement pattern for region centres: `none` (use positioning_bias), `spiral` (wind outward), `burst` (explode from center), `grid` (orderly tile). Overrides positioning_bias when not `none`.

Special region modes

Freeform regions

Some regions can drop their hard bounding box and blend softly into the scene instead of rendering as a pinned rectangle. This removes the "cut-out collage" look.

Setting	Type	Default	Description
`freeform_chance`	FLOAT	0.0	Per-region probability that a region becomes freeform (no hard bbox). Freeform regions are drawn dashed in the preview but excluded from the bounding-box output (they have no fixed location). 0.0 = every element keeps a hard box; 1.0 = nothing is boxed (fully painterly).

Text regions

Some regions can render as actual in-image text (a dictionary word drawn into the picture) instead of an object.

Setting	Type	Default	Description
`text_region_min`	INT	0	Minimum number of regions rendered as in-image TEXT. The text count is a random integer between min and max (clamped to the total region count). This is an EXACT count, not a per-region probability.
`text_region_max`	INT	2	Maximum number of regions rendered as in-image TEXT. Example: min 1, max 2 → always 1 or 2 text words in the image.

Styling: medium, palette, and harmony

Every run randomizes the image medium, color palette family, and color-harmony rule.

Setting	Type	Default	Description
`medium`	ENUM	`photograph`	Image medium (an Ideogram 4 schema value). `photograph` emits a `photo` style key (focal length / aperture); every other medium emits an `art_style` key. `random` picks one per run. Options: `photograph`, `illustration`, `3d_render`, `painting`, `graphic_design`.
`color_palette`	ENUM	`none`	Colour palette family for the image-level palette. `none` = emit no palette; `random_color` = any RGB; themed options: `muted`, `grayscale`, `binary`, `neon`, `pastel`, `colorized` (grayscale tinted with one hue); `random` = pick one family per run.
`color_harmony`	ENUM	`none`	Colour-harmony rule applied to generated colours: `none` (unrelated), `complementary` (two opposite hues), `analogous` (neighbouring), `triadic` (three evenly spaced), `tetradic` (four evenly spaced), `random` (pick one per run). Ignored when `color_palette` is `none`.
`element_palette_chance`	FLOAT	0.0	Per-region probability that a region carries its OWN small colour palette (a subset of the image-level palette) instead of inheriting the global one. 0.0 = all elements share the image palette; 1.0 = every element gets its own colour sub-set. Ignored when `color_palette` is `none`.

High-level description (image overview)

The caption includes a one-line high_level_description (an overview of the whole image). This can be auto-generated or fully overridden.

Setting	Type	Default	Description
`description_override`	STRING	``	Full replacement for the high_level_description. When non-empty, this exact string is used verbatim and generation is skipped. Leave blank to auto-generate.
`description_prefix`	STRING	`A close-up photography of`	PREFIX prepended to the auto-generated description. Ignored when `description_override` is set. Final value: `<prefix> <generated words>`.
`description_length`	INT	35	Target length (words) for the auto-generated description. The generator keeps adding dictionary word-groups until reaching ~this many words. Ignored when `description_override` is set.

Background description

The compositional_deconstruction block includes a background field that sets a base scene layer.

Setting	Type	Default	Description
`description_background_prefix`	STRING	`an environment photography background of`	PREFIX prepended to the auto-generated background description. Leave blank for a fully random background. Final value: `<prefix> <generated words>`.

Inputs summary

Required

seed, width, height
region_count_min, region_count_max
background_weight, large_weight, medium_weight, small_weight
word_length_bias, word_length_randomness
scene_framing, region_word_min, region_word_max
freeform_chance, text_region_min, text_region_max, element_palette_chance
medium, color_palette, color_harmony, positioning_bias, arrangement

Optional

description_length, description_override, description_prefix, description_background_prefix

Outputs

Output	Type	Description
`prompt`	STRING	The assembled Ideogram 4 caption JSON, ready to pass to an API node.
`preview`	IMAGE	Rendered preview of all regions (solid rects for boxed regions, dashed rects for freeform).
`bboxes`	BOUNDING_BOX	Pixel-space bounding boxes `{x, y, width, height}` for each boxed region (freeform regions are excluded). In the format expected by SAM3 and crop nodes.
`width`	INT	Canvas width (pass-through).
`height`	INT	Canvas height (pass-through).

The bboxes output uses per-frame nesting (list[list[dict]]) — the standard shape that SAM3 and other bounding-box consumers expect.

JSON caption structure

The node outputs the same Ideogram 4 caption JSON as the Prompt Builder:

{
    "high_level_description": "A close-up photography of ...",
    "style_description": {
        "aesthetics": "vivid mysterious, ...",
        "lighting": "soft luminous, ...",
        "photo": "85mm, f/5.6",
        "medium": "photograph",
        "color_palette": ["#RRGGBB", ...]
    },
    "compositional_deconstruction": {
        "background": "an environment photography background of ...",
        "elements": [
            {
                "type": "obj",
                "bbox": [ymin, xmin, ymax, xmax],
                "desc": "a vivid tower beside ...",
                "color_palette": ["#RRGGBB", ...]
            },
            {
                "type": "text",
                "bbox": [ymin, xmin, ymax, xmax],
                "text": "TOWER",
                "desc": "..."
            }
        ]
    }
}

bbox coordinates are on a 0–1000 grid as [ymin, xmin, ymax, xmax].
style_description is omitted when color_palette is none.
color_palette is omitted from an element when no colours are set.
Freeform elements (with nobbox=True in generation) are omitted from the elements array.

Spatial connectors (scene framing mode)

When scene_framing is ON, the node weaves descriptions together with these structural words (which do NOT count toward content-word budgets):

beside, near, behind, before, above, below, amid, atop, against, beyond, framing, facing

These are randomly selected to join phrases into one coherent sentence.

Example workflows

High-quality realistic photograph

scene_framing = ON
medium = photograph
region_count_min = 3, region_count_max = 6 (sparse)
background_weight = 0.8, large_weight = 0.2, medium_weight = 0, small_weight = 0 (mostly large shapes)
freeform_chance = 0.3 (some soft blending)
text_region_min = 0, text_region_max = 0 (no text overlays)

Busy abstract collage

scene_framing = OFF (pure mode for maximum randomness)
region_count_min = 25, region_count_max = 40 (dense)
medium = graphic_design
arrangement = grid (orderly)
color_palette = neon, color_harmony = tetradic (vivid 4-colour scheme)
text_region_min = 2, text_region_max = 4 (lots of text elements)

Surreal painted scene

scene_framing = ON
medium = painting
arrangement = spiral (energetic)
positioning_bias = center_weighted (elements cluster toward the middle)
freeform_chance = 0.6 (lots of soft blending)
color_palette = pastel, color_harmony = analogous (soft, harmonious)

19 KiB Raw Permalink Blame History Unescape Escape