Cubist AI Prompt: Bas-Relief for Faceted Depth

A faceless businessman figure in a pinstripe suit, constructed entirely from overlapping faceted angular geometric planes, analytical cubism style, bronze and copper metallic tones with green-blue oxidized patina accumulating in recessed facet angles, thick oil paint texture with visible palette knife strokes and raised impasto edges following each facet boundary, fragmented background of rectangular color blocks in muted ochre, burnt sienna, teal and burgundy, figure standing with hands in pockets, strong vertical composition, dramatic chiaroscuro lighting from upper left at 45 degrees, hard-edged geometric cast shadows between facet layers, bas-relief sculpture effect with measurable physical depth between overlapping planes, each facet edge raised 2-3mm from the surface, gallery fine art quality --ar 9:16 --style raw

Bas-relief is a sculptural constraint, not a stylistic adjective. In a cubist prompt, it tells the model that the surface has measurable physical depth: the facets are carved from a single block rather than painted onto a flat plane. This distinction matters because cubist prompts default to flat geometry unless a physical property is explicitly named. The word shifts the render target from a decorated plane to a carved object, and that shift is what produces the layered shadow structure the rest of the prompt depends on.

Analytical cubism already specifies "overlapping faceted angular geometric planes" in most working prompts. Without a depth anchor, the model renders those planes as adjacent shapes sharing one surface. The eye reads pattern, not structure. The cubist fashion figure prompt demonstrates the same principle in a different subject, and the failure mode is identical: facets collapse into a mosaic unless something forces physical separation between them.

What Bas-Relief Actually Tells the Model

Midjourney's training data associates "bas-relief" with carved stone, bronze plaques, and architectural friezes. When the token appears, the model allocates shading gradients to facet edges — not just color changes. This is why the prompt's existing request for "hard-edged geometric cast shadows between facet layers" becomes achievable at all. Without the token, the model has no reason to generate side walls on each facet, so it renders color transitions instead of shadow transitions.

The mechanism is token proximity. "Bas-relief" sits near "3D" and "physical depth" in the prompt, reinforcing the same latent concept across multiple tokens. When these cluster, the model treats each facet as a raised element with a visible edge, a side wall, and a cast shadow. When they drift apart, the model averages them into a generic "3D-ish" look that reads as decorative texture rather than carved structure.

Before moving forward, generate two versions of the same prompt: one with "bas-relief" and one without. Compare the shadow falloff at facet boundaries. If the shadows are uniform in width and direction, the depth anchor failed and the model treated the word as a style label rather than a physical property.

How the Model Interprets Depth Between Facets

The model does not simulate geometry. It predicts pixel patterns that historically co-occur with the tokens in the prompt. "Bas-relief" co-occurs with hard-edged shadow lines, visible edge walls, and directional light falloff. That is why the prompt's "dramatic chiaroscuro lighting from upper left at 45 degrees" becomes more legible when bas-relief is present: the model has a reason to assign directional shadows to each facet edge instead of flattening them into ambient occlusion.

The effect is not binary. Bas-relief strength scales with how many depth-related tokens surround it. A prompt with "bas-relief," "physical depth," "3D," and "cast shadows" produces deeper apparent relief than a prompt with only "bas-relief." The tokens do not duplicate meaning; they triangulate a single concept from multiple angles, and the model responds to the density of that triangulation.

Check this by removing "physical depth" from the prompt while keeping "bas-relief." The facets will still have edges, but the shadows will lose their directional logic. This tells you which token was doing the structural work versus which one was reinforcing it.

The Parameter Interaction That Makes or Breaks the Effect

--style raw removes Midjourney's default aesthetic polish. This matters for bas-relief because stylized rendering flattens surface depth into decorative texture. Raw mode preserves the harsh edge transitions that make carved facets legible, and it prevents the model from smoothing the shadow gradients into a painterly blur. Without raw mode, bas-relief still functions, but the edges soften and the figure reads as painted rather than carved.

The 9:16 aspect ratio is not neutral either. Vertical composition stacks facets along the Y-axis, which gives the model more vertical pixels to render shadow gradients per facet. A square crop compresses the same number of facets into less vertical space, reducing the visible depth cue. The effect is subtle but measurable: the same prompt at 1:1 produces shallower apparent relief than at 9:16.

The interaction is non-obvious. --style raw plus 9:16 plus bas-relief produces deeper apparent relief than any one setting alone. Removing raw mode while keeping the word still works, but the facet edges lose roughly half their visible side-wall thickness. This is the parameter trade-off most tutorials skip: the word and the flags are not independent.

Why the Word Alone Fails Without Texture Anchors

"Bas-relief" tells the model about geometry, not material. Without a texture anchor, the model renders smooth, plastic-looking facets that read as folded paper. The original prompt includes "thick oil paint texture with visible palette knife strokes and raised impasto edges" — this is the material anchor that makes the depth believable. Impasto and bas-relief are complementary: impasto describes paint pushed up from a canvas, bas-relief describes stone carved down from a block.

Together they tell the model the surface has relief in both directions. This is why the prompt specifies "raised impasto edges following each facet boundary." The impasto handles the micro-texture within each facet; bas-relief handles the macro-structure between facets. Remove either one and the other loses its reference point. A deconstructed portrait sketch uses the same two-layer logic with line work instead of impasto, and the principle transfers directly.

Check: if the facets look like folded paper rather than carved metal, the texture anchor is too weak. Add "weathered bronze patina" or "oxidized copper surface" to reinforce material weight. The green-blue oxidized patina already in the prompt does this work, but it needs the bas-relief token to have something to accumulate inside.

Testing the Boundary — When Bas-Relief Collapses

Bas-relief fails when the model reads the word as a style label rather than a physical property. This happens when the prompt contains too many competing style tokens. If "bas-relief" sits next to "flat design," "minimalist," or "graphic poster," the model averages the concepts and produces shallow geometry. The token loses its physical reference and becomes decorative.

The failure mode is visible in three stages. First, facet edges lose their side walls. Second, shadows become gradients instead of hard-edged shapes. Third, the figure reads as a mosaic rather than a sculpture. This is the boundary: the point where the word stops describing a physical object and starts describing a look. Once crossed, no amount of additional depth tokens will restore the carved quality.

Leonardo.ai's prompt weighting system makes this boundary easier to test because you can isolate the token and measure its effect independently. In Midjourney, the equivalent test is removing competing style tokens one at a time and observing which removal restores the facet side walls. The token that restores them is the one that was suppressing the depth anchor.

The Signal That Tells You the Boundary Was Crossed

The primary signal is shadow behavior. In a successful bas-relief render, every facet boundary has a distinct cast shadow with a visible direction. When the boundary is crossed, shadows become ambient occlusion — soft darkening at edges with no directional logic. The shift from directional to ambient shadow is the clearest indicator that the model stopped treating the facets as physical objects.

The secondary signal is background separation. The prompt specifies "fragmented background of rectangular color blocks." In a working bas-relief render, those blocks sit behind the figure with clear depth separation. When the effect collapses, the background merges with the figure's facets and the entire image reads as one flat plane. The background is not just decoration; it is a depth reference that the model uses to calibrate the figure's relief.

The practical limit is token count and token order. Beyond roughly 60 tokens, style descriptors start competing for attention. Bas-relief needs to sit near the depth-related tokens to survive. If it drifts toward the end of a long prompt, the model deprioritizes it and the facets flatten. The signal is immediate: the first render after the word drifts will show ambient shadows instead of directional ones. That is where this approach stops working, and the fix is repositioning, not adding more words.