60+ Halloween AI Photo Prompts (2026): Costumes, Couples, Kids, Pets and Spooky Portraits

A prompt should survive a change of tool, but the controls around it do not. OpenAI's current Help Center describes ChatGPT Images 2.5 as available on web, iOS and Android, with image creation and editing, aspect-ratio choices, templates and conversational refinement. Google lists Nano Banana 2 as the product name for Gemini 3.1 Flash Image, with the API model ID `gemini-3.1-flash-image`; it is positioned as a general-purpose image generation and editing model. Midjourney's documentation says V8.2 is the current default and supports version parameters, image prompts and its newer Edit Model.

Use ChatGPT when you want to iterate in ordinary language or edit a supplied image. Use Nano Banana 2 when you want a Gemini image workflow or API route with a current model ID. Use Midjourney when its aesthetic controls, personalization and editor fit the look you want. Do not copy a parameter from one tool into another and assume it means the same thing. Keep the creative brief separate from tool syntax: write the scene first, then append the tool's current aspect-ratio, version or reference controls.

Write the prompt in the order the image needs to be understood. Start with the subject and action: 'a seated adult woman looking at the camera' is more useful than 'Halloween vibes'. Add the costume with materials and construction, then the environment, composition, camera distance, lighting, colour palette and mood. Finish with exclusions and any exact text. This gives the model a hierarchy instead of a pile of adjectives.

For a reference edit, say what must remain stable: face shape, hair length, skin tone, body position, clothing fit or the pet's markings. Avoid asking for ten changes in one pass. First establish the costume, then adjust the background, then request a crop or colour correction. If the tool changes identity or invents a hand, restore the last good version and make a smaller edit. Consistency comes from controlled iteration, not from adding more adjectives.

Replace the bracketed fields with real details. These are starting briefs, not claims about guaranteed output. Add a reference image only when you have the right to use it, and keep the identity-preservation instruction close to the subject description. For commercial work, add the brand's approved colours, logo file and usage destination yourself rather than asking the model to guess a trademark.

Costumes and portraits: (1) 'Portrait of [person] as a moonlit Victorian botanist, black velvet coat, silver fern brooch, old greenhouse, soft blue rim light, calm expression, preserve identity and natural skin texture.' (2) 'Full-body editorial portrait of [person] as a friendly retro astronaut at a pumpkin farm, orange suit details, warm sunset, playful but believable.' (3) 'A realistic 1980s arcade vampire costume for [person], red windbreaker over a dark shirt, neon reflections, waist-up composition, no fangs unless requested.' (4) 'A folklore-inspired forest ranger costume, moss-green cloak, lantern, autumn fog, grounded material detail, no weapons.' (5) 'Elegant masquerade portrait, black and copper mask held beside the face rather than covering it, candlelit ballroom, preserve facial features.' (6) 'Friendly mad-scientist costume in a tidy home lab, goggles resting on the head, purple and amber practical lights, no dangerous chemicals.' (7) 'Classic ghost costume made from translucent fabric with visible shoes and a playful pose, family-friendly studio background.' (8) 'Cyberpunk witch portrait, practical layered clothing, small glowing charms, rainy city alley, readable face, no brand logos.' (9) 'Cozy scarecrow portrait, stitched denim, straw hat, cornfield at golden hour, warm and humorous mood.' (10) 'Gothic librarian portrait, deep green dress, brass key ring, candlelit archive, rich shadows but clear eyes.' Couples and friends: (11) 'Two adults as rival potion makers, distinct costumes and poses, one emerald workstation and one amber workstation, editorial photograph, keep both faces recognisable.' (12) 'Couple in coordinated black-and-orange carnival costumes, standing under a string-light arch, candid laugh, no wedding implication.' (13) 'Three friends as a vampire, a werewolf and a ghost, modern clothes with restrained costume details, street portrait, equal prominence.' (14) 'Best-friend portrait in matching skeleton-inspired jackets, urban night, reflective pavement, fashion editorial, no skull makeup unless requested.' (15) 'Two adults in a playful haunted-house escape-room scene, flashlight beams, readable expressions, no gore.' (16) 'Couple as a celestial moon and sun, complementary fabrics, dark blue studio, subtle stars, tasteful and non-literal.' (17) 'Four colleagues in an office Halloween photo, each with a different classic-monster accessory, natural group spacing, professional and funny.' (18) 'Friends around a pumpkin-carving table, candid overhead composition, warm kitchen light, realistic hands and tools.' (19) 'Two dancers in coordinated phantom costumes on an empty theatre stage, dramatic spotlight, elegant movement, no horror violence.' (20) 'A family-friendly costume-party group photo with [number] adults, everyone looking in different natural directions, keep each person distinct.' Kids and families: (21) 'Age-appropriate child portrait as a storybook dragon rider, bright felt costume, autumn garden, gentle expression, supervised feeling, no fire.' (22) 'Two siblings as a witch and a friendly black cat, colourful porch decorations, soft daylight, playful rather than scary.' (23) 'Child in a homemade robot costume built from cardboard and foil, living-room setting, proud smile, realistic craft texture.' (24) 'Family of four in coordinated woodland costumes, pumpkin patch, golden-hour light, preserve each face and natural proportions.' (25) 'Toddler in a soft pumpkin onesie on a clean blanket, warm window light, simple background, safe seated pose.' (26) 'Child as a tiny astronomer exploring a paper-moon Halloween set, studio portrait, bright stars, no frightening creatures.' (27) 'Siblings holding lanterns on a safe, well-lit porch, cosy autumn costumes, visible adult supervision implied by the composition, no darkness hiding faces.' (28) 'Family recipe-card photo: adults and children decorating biscuits with orange icing, overhead table, clear food details, no text unless supplied.' (29) 'Child in a friendly dinosaur costume beside oversized paper pumpkins, bright park, natural candid pose, no realistic teeth.' (30) 'A multi-generation family portrait where each person wears one subtle Halloween accessory, warm living-room light, believable hands and eyes.' Pets and products: (31) 'Small dog in a soft pumpkin cape, seated on a plain autumn blanket, natural fur texture, relaxed posture, no tight clothing.' (32) 'Cat beside paper bats and a safe battery lantern, low-angle portrait, preserve whiskers and eye colour, no smoke.' (33) 'Golden retriever as a friendly mail carrier for a Halloween card, tiny hat placed safely, autumn path, tongue relaxed, no human hands.' (34) 'Two pets in coordinated bandanas, studio background, equal scale, no costume covering faces or restricting movement.' (35) 'Black cat silhouette on a moonlit windowsill, warm room behind it, realistic fur edge, family-friendly atmosphere.' (36) 'Product photo of a carved pumpkin lantern on a dark table, controlled candle glow, room for headline text on the left, no brand mark.' (37) 'E-commerce image of a handmade witch hat on a neutral cream background, accurate fabric texture, front and side views as separate images.' (38) 'Halloween baking flat lay with cookies, paper decorations and orange napkin, clean overhead composition, no invented product packaging.' (39) 'Close-up of a painted pumpkin with the exact supplied word “[TEXT]” in large readable letters, simple background, check spelling.' (40) 'A pet adoption campaign image with a calm dog in a pumpkin-themed bandana, clear eyes, generous negative space, no misleading shelter details.' Spooky but safe scenes: (41) 'Abandoned cinema lobby at blue hour, dust in the light, one flickering sign with no readable brand, cinematic wide shot, no people.' (42) 'Foggy forest path with a lantern and red leaves, suspenseful but non-violent, no monster, vertical composition.' (43) 'Old coastal lighthouse during a storm, warm window glow, dramatic clouds, realistic architecture, no disaster victims.' (44) 'Haunted library with books slightly out of place, floating paper, warm candlelight, whimsical mystery, no gore.' (45) 'VHS-era suburban Halloween night, porch lights, costumes in the distance, grainy colour treatment, no identifiable strangers.' (46) 'A whimsical skeleton parade made from paper puppets, bright stage lighting, graphic illustration, friendly expressions.' (47) 'Moon reflected in a quiet lake beside pumpkins and reeds, atmospheric landscape, no text, no people.' (48) 'A mysterious closed amusement park gate with autumn leaves, fog behind the fence, editorial photograph, no threatening figure.' (49) 'Candlelit attic with old trunks and a framed family portrait, warm shadows, tasteful gothic styling, no distressed people.' (50) 'A tiny village of illuminated pumpkins on a mossy table, macro photography, rich texture, no real flames near children.' Social and party graphics: (51) 'Vertical 9:16 Halloween party invitation background, orange moon, black cat silhouette, empty space at top for supplied text, clean edges.' (52) 'Square social post with three illustrated costume icons, purple and orange palette, blank centre panel for later typography, no fake logos.' (53) 'Wide banner for a neighbourhood trick-or-treat event, friendly pumpkins, readable empty area for date and address, no generated text.' (54) 'Newsletter header with a subtle haunted-house silhouette, cream paper texture, space for a black headline, restrained and editorial.' (55) 'Halloween sale background for [product category], dark plum and copper palette, product-free composition, room for real pricing added later.' (56) 'Podcast cover concept: a vintage microphone surrounded by paper bats, high contrast, exact title “[TITLE]” only, verify every letter.' (57) 'YouTube thumbnail concept with a surprised adult holding a pumpkin, clean left-side negative space for real headline text, no clickbait words.' (58) 'Recipe card background with autumn leaves and a small pumpkin border, empty centre, flat illustration, consistent corners.' (59) 'Office event poster background with playful ghosts in the margins, professional central space, no invented venue or date.' (60) 'A three-image campaign set with the same orange moon, paper texture and colour palette across portrait, square and landscape crops; keep the visual system consistent.'

Consistency is a workflow problem, not a magic keyword. Choose one approved reference image with a clear face or object, good lighting and no distracting background. Describe what is allowed to change and what is not. For a person, name face shape, hair, age presentation, clothing fit and expression without asking the model to infer sensitive traits. For a pet, preserve markings, ear shape and eye colour. For a product, specify dimensions, material, label placement and which sides must be visible.

Make a small change, save the output and compare it with the reference. If identity drifts, do not keep adding 'exactly the same' to a longer prompt. Return to the last good image and simplify the instruction. For a campaign, create a short style sheet: palette, lens feel, lighting direction, background texture, crop options and forbidden elements. Generate the hero image first, then use it as the reference for supporting images where the tool supports that workflow.

Ask permission before transforming a real person's photo, especially if the result could embarrass them, imply a belief, change their body or put them in a frightening context. Do not use a generated costume image to suggest that someone committed an act, attended an event or endorsed a product. Avoid identifying details in public posts. If a person says the image misrepresents them, remove it rather than arguing about whether the model made it.

Keep children's scenes age-appropriate, bright enough to inspect and free from frightening adult themes. For pets, prefer a prop beside the animal or a loose, safe bandana over restrictive clothing, and never depict distress as a joke. Keep gore, realistic injury and dangerous props out of party and family prompts. Review hands, teeth, eyes, labels, shadows, reflections and background strangers. A polished image can still contain a private detail or an unsafe implication that the prompt did not intend.

Image models are increasingly capable with short text, but a generated poster is not proof that every letter is correct. Give the exact wording in quotation marks, state the hierarchy and leave generous space. For a real event, add the date, address, price, phone number and legal copy in a design tool after the image is generated. Do not let an image model invent a venue, discount, sponsor or safety instruction.

If you need a product image or social asset, generate a clean background with negative space and add the final type yourself. This is usually faster than trying to regenerate a perfect paragraph of text. Export at the intended crop, inspect at mobile size and ask another person to read it. A spelling error that is invisible in a thumbnail can still make a campaign look careless when opened full size.

Pass one is the brief: decide the audience, destination, aspect ratio, subject, mood and one success criterion. Pass two is the base image: generate one clear composition without stuffing the prompt with alternatives. Pass three is the correction: fix the most important defect only, such as face drift, costume colour or empty space. Pass four is the set: create matching crops or supporting scenes using the approved image and style notes. Pass five is the human review: compare the final file with the brief and remove anything misleading, unsafe, private or unreadable.

Keep filenames useful: `halloween-hero-v01`, `halloween-hero-v02-background`, `halloween-social-square-approved`. Record the tool, model label, prompt, reference image and date. That small log makes a refresh possible when a model changes. It also prevents you from claiming that a prompt is universally reliable when it only worked in one interface with one reference. The goal is a repeatable creative process, not a lucky screenshot.