Costume sketches, fabric swatches and a tablet on a dark work table

Prompt library

Halloween AI prompts you can use and adapt

Each recipe separates the reference photo, the transformation, the scene and the failure checks. Copy one into another assistant or open the linked generator with the same starting prompt already loaded.

Open a costume editor Browse all looks

Halloween 2026 guide, checked October 20, 2026

Structure: Keep, change, avoid; Identity first, transformation second, failure clauses last.
Reuse: Any capable model; The wording is plain enough to adapt outside SeedVR2.
Testing: One variable at a time; Change a costume, material or background between runs.

Choose a Halloween look

How to adapt a copied prompt

Do not add more Halloween adjectives. Add concrete decisions: what the person wears, what material it is, where they stand and how the scene is lit.

1
Keep the preservation clause
The first sentence should tell an image editor that the uploaded person, face, pose and proportions are the reference.
2
Replace one visual block
Swap the costume paragraph, the location paragraph or the lighting paragraph separately.
3
Keep the failure checklist
Extra fingers, duplicated people, changed identity, text and watermarks are practical review points.

Copy-ready Halloween prompts

These are original directions rather than named film or game characters. Adapt materials, colours and settings freely.

Haunted unicorn selfie

Use the uploaded photo as the primary reference. Preserve the person's recognizable face, skin tone, age and pose. Create an original haunted-unicorn mascot portrait with a sculpted ivory mask, one silver horn, a plum velvet cape and tiny moon charms. Keep the eyes visible. Place the subject in a foggy moonlit garden. Photorealistic editorial photography, realistic fabric and cinematic light. Avoid an existing character design, extra limbs, changed identity, text and watermarks.

Open the generator

Midnight braids portrait

Edit the uploaded portrait into a gothic Halloween fashion image. Preserve the subject's identity, facial proportions, skin texture, pose and body. Change the hair into long violet braids with delicate silver ornaments. Replace the clothing with an elegant black-and-purple velvet outfit. Place the subject in a moonlit Victorian garden with soft fog and lanterns. Use photorealistic editorial photography. Keep one person. Avoid altered identity, extra fingers, distorted anatomy, text and logos.

Open the generator

Pumpkin pet studio

Use the uploaded pet photo as the primary reference. Preserve the animal's breed, fur colour, markings, eye colour, face and natural anatomy. Dress the pet in a soft pumpkin costume with a small green leaf accessory that does not cover the eyes. Use a cozy autumn studio with warm light and miniature pumpkins. Create a photorealistic professional pet portrait with realistic fur. Avoid extra animals, duplicated paws, changed species, text and watermarks.

Open the generator

Cowboy and angel couple

Use both uploaded photos as identity references. Preserve both people's recognizable faces, skin tones, body proportions and relative identities. Dress one person in a refined western costume with a dark hat and embroidered jacket; dress the other in an ethereal ivory costume with soft feathered wings. Place them together in a warm dusk field. Keep exactly two people and do not exchange faces. Avoid a third person, duplicated hands, distorted anatomy, text and watermarks.

Open the generator

Five-second liminal hallway

Create a five-second cinematic shot inside an original empty liminal corridor with cream walls, worn carpet and one flickering ceiling light. The camera moves forward slowly and steadily. A faint fog crawls along the floor. Keep the architecture stable, the lighting consistent and the motion subtle. No people, no sudden cuts, no camera shake, no text, no logos and no watermarks.

Open the generator

Choose a Halloween look

Every card opens a page with a tested prompt, an editable photo, and a named starting model. Swap the model in the workspace without losing your upload or prompt.

Haunted unicorn

A mascot silhouette, one silver horn and a moonlit garden.

Adult with long violet braids in a dark velvet costume in a moonlit garden

Midnight braids

Violet braids, silver details and a Victorian garden at night.

Cosmic princess

Midnight fabric, silver stars and a moonlit observatory.

Robot costume

A real face inside a wearable armour costume, not a mechanical replacement.

Dog in a soft pumpkin costume sitting in a warm autumn studio

Pumpkin pet

Warm autumn light, soft costume fabric and the same pet underneath.

Cowboy and angel

Two coordinated costumes, two identities, one dusk scene.

Original empty liminal corridor with cream walls, worn carpet and a ceiling light

Backrooms video

A slow push through an empty corridor, with fog and a failing light.

A prompt is a checklist, not a mood board

Put the reference and identity requirement first.
Use one paragraph for costume and one for scene and light.
End with visible checks you can review, not vague quality claims.

The strongest recipes make decisions. They say what reference wins, which garment replaces the original and which scene supports the subject. A list of atmospheric words alone leaves the model to guess the important parts.

Keep a short version of each prompt next to its output. When a later run goes wrong, compare the wording before changing the model. That separates prompt problems from model limitations.

Halloween prompt FAQ

Can I copy these prompts into Gemini or another tool?

Yes. They are written as plain instructions. Some tools handle uploaded reference images differently, so keep the preservation and avoid clauses and adjust the wording to the tool's supported features.

Why do the prompts repeat identity instructions?

Image editors can drift when several visual changes compete. Repeating the face, pose and body requirements makes the invariant explicit and gives you a clear standard for reviewing the result.

Can I use a famous character name?

The library uses original descriptions to avoid copying protected character designs. You can describe a hero, mascot, witch or ghost archetype and specify the silhouette, materials and colours you want.

How should I change a prompt when the result is wrong?

Change one block. If the costume is right but the background is wrong, leave the costume paragraph alone. If identity drifts, remove secondary subjects and simplify the scene before increasing model quality.

Keep exploring

Costume generator

Preserve a real face while changing clothing, make-up and atmosphere.

Halloween photo editor

Relight, restyle and clean up a costume photo without replacing the person.

Halloween video generator

Short atmospheric clips built from one action, one camera move and controlled light.

Haunted unicorn

A mascot silhouette, one silver horn and a moonlit garden.

Midnight braids

Violet braids, silver details and a Victorian garden at night.

Copy-ready Halloween prompts

These are original directions rather than named film or game characters. Adapt materials, colours and settings freely.

Halloween prompt FAQ

Can I copy these prompts into Gemini or another tool?

Why do the prompts repeat identity instructions?

Image editors can drift when several visual changes compete. Repeating the face, pose and body requirements makes the invariant explicit and gives you a clear standard for reviewing the result.

Can I use a famous character name?

How should I change a prompt when the result is wrong?

Halloween AI prompts you can use and adapt

How to adapt a copied prompt

Keep the preservation clause

Replace one visual block

Keep the failure checklist