Black Cat Prompts: Stop Drawing White Lines

A plump black cat in hand-carved linocut style, oversized white eyes with tiny pinpoint pupils gazing upward, one front paw stretched high in playful reach, delicate white whiskers and small nose carved out of the solid black silhouette, fine irregular scratch marks suggesting fur texture, bold matte black form against pure white void, subject anchored bottom-left with roughly 70% negative space above, childlike innocence meets vintage relief printmaking, flat two-color graphic illustration, no gradients, no shading, no gray tones --ar 3:4 --style raw --v 6

Why Black Cats Are the Hardest Simple Subject

I once spent an afternoon trying to generate a black cat illustration and got twenty versions of gray mush. The brief looked easy: one shape, two colors, no background. The lesson was structural. A solid black silhouette has no internal information, so the model invents it. Eyes, whiskers, fur edges—everything that makes the cat readable has to come from negative space or a specific mark type. If you only ask for a black cat, you are asking the model to solve a two-color print with additive habits it does not have.

I learned this after a prompt for white whiskers produced clean, even strokes painted over the black form. They looked like correction fluid, not whiskers. Diffusion models default to additive thinking: white marks on top of black. For a print-style black cat, you need subtractive language. Ask for fine scratch marks, tiny gaps, or lines carved out of the silhouette. The model then treats the whiskers as missing ink rather than added paint. That single wording change made my black cat illustration read as a cat instead of a stencil with accessories.

My first fur test failed the same way. I wrote detailed fur texture and got gray hatching that broke the two-color brief. The fix was to name a physical mark: fine irregular scratch marks suggesting fur. Short, tapered, unevenly spaced gouge cuts. The model can reproduce that mark type because it has a shape and a cause. It cannot reproduce detailed fur texture because that phrase is an outcome, not a tool. Once I switched to scratch marks, the black cat illustration kept its silhouette clean and still looked furry.

Anchoring the Style to a Physical Process

Style words like "minimalist," "cute," or "graphic" are taste descriptors. They tell the model what you want to feel, not what to make. The model has no stable visual cluster for "cute," so it averages across thousands of unrelated cute images and returns something generic. Style anchoring works when the word points to a process with consistent physical evidence: linocut, woodblock, screenprint, risograph.

"Hand-carved linocut" is especially strong because linocut has unmistakable signatures: edges that wobble slightly where the gouge changed direction, texture marks inside large inked areas where the tool skipped, and bold flat blacks with no halftone. The training data contains enormous quantities of real linocut prints, all sharing these traits, so the phrase activates a tight, coherent visual cluster. Compare that with "woodcut style," which pulls toward older, denser, cross-hatched imagery — a related but visually different tradition. Choosing between them is not a matter of preference; it is a choice between two distinct output distributions.

This principle generalizes. If you enjoy graphic, print-derived aesthetics, the same process-anchoring technique drives results like the grunge street art portrait, where "wheatpaste poster" and "spray stencil" do the heavy lifting, or the neo-expressionist portrait, which anchors to Basquiat-era mark-making rather than an abstract "edgy" mood.

Negative Space Is a Number, Not a Mood

The single biggest compositional weakness in AI illustration prompts is unquantified space. "Lots of empty background" produces random results because the model has no threshold for "lots." One generation gives you a centered cat with modest margins; the next crops the ears. The fix is to treat composition like a layout spec: subject anchored bottom-left, roughly 70% negative space above.

Two things happen when you do this. First, placement language ("bottom-left," "lower third") reliably shifts the subject off-center because the model has strong positional priors from captioned layout data. Second, the percentage gives the model a scale relationship between figure and field. The result is the tall, breathing composition you see in the reference image — the cat occupies the bottom portion, the raised paw reaches into the void, and the emptiness becomes an active design element rather than leftover canvas.

That upward reach matters more than it appears. A raised paw pointing into the negative space creates a directional vector that connects subject and void; without it, a bottom-anchored subject can feel like it is sliding off the page. When composing any high-negative-space piece, give the subject a gesture that points into the emptiness. For more on structured composition, the cubist fashion figure breakdown covers how geometric placement controls visual tension.

Controlling the Two-Color Constraint

Flat two-color output fails through a specific mechanism: anti-aliasing and soft-edge priors. Diffusion models are trained on photographs, where edges transition through intermediate values. Left alone, the model introduces gray at every boundary — around the silhouette, inside the eyes, along the whiskers. The image still reads as black and white, but zoom in and it is full of tonal mush that prints badly and scales poorly.

The countermeasure is explicit negative constraint stacking: "no gradients, no shading, no gray tones." One negative is rarely enough because the model treats a single exclusion as a soft preference. Three overlapping negatives covering the same failure mode from different angles (smooth transitions, light modeling, intermediate values) collectively close the loophole. Adding "flat" and "matte" as positive descriptors reinforces the same target from the other direction.

The --style raw parameter in Midjourney supports this by reducing the model's default beautification layer, which tends to add soft lighting and depth even to graphic prompts. For print-style work, raw mode is close to mandatory — the default styling actively works against flatness. If you are adapting this prompt for other tools, DALL-E 3 handles the negative constraints conversationally but needs the "two-color screenprint" framing restated, since it lacks a raw-style toggle.

The Eyes Carry the Entire Image

In a silhouette illustration, the eyes are the only facial feature with real estate, so their construction determines the entire emotional register. Three parameters matter: eye size relative to head, pupil size relative to eye, and gaze direction. Oversized white eyes with pinpoint pupils produce the startled-innocent expression; the same eyes with large pupils produce generic cartoon cuteness. Gazing upward adds vulnerability and justifies the raised paw narratively — the cat is reaching for something above, which is also why the negative space above it feels intentional rather than empty.

Change any one of these and the image tells a different story. That is the real lesson of this prompt: in constrained graphic work, tiny parameter shifts carry outsized narrative weight.

From here, try pushing the same system in new directions: swap the cat for another silhouette animal and keep every structural phrase identical, or move the anchor to top-right with the subject peering downward. Each variation teaches you which phrases are load-bearing and which were decoration — and that knowledge transfers directly to every flat illustration prompt you write afterward.