Replicate Prompting

Prompting image and video models on Replicate

Distilled from Replicate's blog posts (2024-2026). Techniques are model-agnostic. For model selection and execution, see replicate-models.

§ Image

Writing prompts

• Natural language, not keywords: "A woman standing in a Tokyo alleyway at dusk, neon signs reflecting off wet pavement" beats "woman, Tokyo, alleyway, dusk, neon, wet pavement."
• Be specific: name exact colors, materials, lighting, camera equipment, spatial relationships.
• Name subjects directly: "the woman with short black hair" — avoid pronouns.
• Long prompts win: most models accept thousands of tokens. 12+ specific requirements work if each is stated clearly.
• Start simple, iterate: test small edits first, then build.

Photographic language

• Film stocks: Kodak Portra 800, Fuji Velvia 50, Ilford HP5
• Lens: 50mm Summilux wide open, 85mm f/1.4, 24mm wide-angle
• Depth of field: shallow vs deep
• Techniques: golden hour, blue hour, long exposure, double exposure
• Lighting: Rembrandt (triangle on cheek), soft diffused studio, rim/backlight, flat diffused, volumetric
• Composition: rule of thirds, symmetry, wide/medium/close-up/macro, high/low angle, tilt-shift

Text rendering

• Wrap text in double quotes: "Design a poster with the title \"BLUE NOTE SESSIONS\""
• Stick to readable fonts
• Editing: "Change 'old text' to 'new text'"
• Match text length to avoid layout shifts
• Some models inpaint text: mask region, prompt with new text

Style transfer

• Name exact style: "impressionist painting," "1960s pop art," "Sumi-e ink wash"
• Describe traits if label fails: "visible brushstrokes, thick paint texture, rich color depth"
• State what stays: "keep the original composition"
• Example-based editing: provide before/after pair, model infers transformation

Character consistency

• Clear reference description: "the woman with short black hair and green eyes wearing a navy blazer"
• Say what changes (setting, activity) and what stays (face, clothing)
• Use reference images when supported
• Break complex changes into steps

Image editing

• General: specify what to keep. "keeping the pose and expression unchanged"
• Verbs: "change the clothes to a blue jacket" not "transform"
• Object removal: describe what fills the space, not just what to remove
• Background: describe new background in detail; specify subject stays
• Inpainting: mask region, prompt fill. Describe whole scene if magic prompt is off; describe only masked region if on
• Outpainting: ControlNet-style conditioning (edge detection, depth maps) preserves structure

Multi-image and storyboard

• Ask for "a series" or specify grid (e.g., "2x2 storyboard grid")
• Describe each panel with consistent character descriptions
• Repeat exact descriptions for style/character continuity

Product photography

• Materials: "brushed steel," "matte aluminum," "kraft paper," "frosted glass"
• Lighting: "soft diffused studio lighting, crisp highlights and gentle shadows"
• For brand assets: look for models with native SVG output
• For layouts: look for models with strong typography and composition

Fine-tuning and LoRAs

• Use trigger words from trained model in every prompt
• Balance multiple LoRAs with scale parameters (0.9-1.1)
• Generate synthetic training data: generate many images, pick best, retrain

Common image pitfalls

1. Keyword-stuffed prompts — write sentences, not tags
2. "Transform" for small edits — use targeted language
3. Not specifying what to keep
4. Negative prompts on models not trained for them
5. Too-high guidance scale → burnt contrast
6. Short prompts for complex scenes
7. Ignoring aspect ratio — most models work best at ~1 megapixel
8. Wrong model for the task — if it struggles, try another
9. Not iterating — best results come from iterative refinement

§ Video

Scene description

A good video prompt is a scene description, not a caption. Layer these elements:

1. Subject: who or what
2. Context: where
3. Action: what happens
4. Style: cinematic, animated, stop-motion, documentary
5. Camera: dolly, tracking, static, handheld
6. Composition: wide shot, close-up, over-the-shoulder
7. Ambiance: mood and lighting

Be specific: "A high-speed car chase on a rain-drenched highway at night. Two muscle cars weave through heavy traffic at 140mph..." beats "A car chase."

Overdescribe: "a desperate man in a weathered green trench coat picks up a rotary phone... bathed in the eerie glow of a green neon sign" beats "a man on the phone."

Camera and cinematography

• Shot types: wide/establishing, medium, close-up, extreme close-up
• Motion: static, pan, tilt, dolly, tracking, crane, handheld, drone, dolly zoom
• Position: eye level, low angle (powerful), high angle (vulnerable), over-the-shoulder, POV
• Lens: shallow depth of field, deep focus, macro, wide-angle, tilt-shift
• Escalation pattern: wide → medium → close-up → extreme close-up (maps well to 8-15s clips)

Audio and dialogue

Prompt all four layers:

1. Dialogue: exact words or described intent
2. Ambient sound: rain on metal, city traffic, forest birds
3. Sound effects: door slamming, glass breaking, sword drawn
4. Music: genre, mood, instrumentation

Dialogue tips:

• Explicit: "The man says: My name is Ben."
• Implicit: "The man introduces himself."
• Match dialogue length to clip duration — 8s fits ~2-3 short sentences
• Use colons (not quotes) to avoid subtitle contamination: "She says: Hello there"
• Add "(no subtitles)" and repeat if needed
• Phonetic spelling for mispronounced names: "foh-fur" not "fofr"
• Tie dialogue to visual descriptions in multi-character scenes

Multi-shot and time-coded prompting

Write timestamps directly:

[0-4s]: Wide establishing shot, static camera, misty bamboo forest at dawn
[4-9s]: Medium shot, slow push-in, the fighter steps forward
[9-15s]: Close-up, orbit shot, the fighter strikes, slow motion

Transitions: "Hard cut to...", "Seamless morph into...", "Whip pan to...", "Snap cut to..."

Reference inputs

• Image-to-video: feed starting image; describe motion and changes
• First/last frame interpolation: provide both images; model generates transition
• Subject references: bracket syntax like [Image1] or [Image2]
• Audio-driven: sync video to audio — transcribe audio in prompt, match duration
• Multi-reference: image for appearance + video for motion + audio for rhythm + text for direction

Style control

• Name style explicitly: "claymation," "Pixar animation," "anime," "stop-motion," "8-bit retro," "LEGO"
• Quality anchors: "hyper-realistic, 8k," "cinematic"
• Genre language: "Michael Mann cinematography" (neon, night), "Wes Anderson" (symmetrical, pastel), "Blade Runner 2049" (atmospheric, orange/teal)
• Use input images for pixel-level style control
• Grain/texture: "slightly grainy, film-like" or "VHS aesthetic" avoids too-clean AI look

Character consistency

• Repeat descriptions verbatim across prompts
• Character sheet: "John, a man in his 40s with short brown hair, wearing a blue jacket and glasses, looking thoughtful"
• Vary scene, not character
• Reference images more reliable than text alone for facial features

Common video pitfalls

1. Not describing audio → hallucinates inappropriate sounds (e.g., studio audience laughter)
2. Too much dialogue for clip length → unnaturally fast speech
3. Too little dialogue → gibberish or awkward pauses
4. Not specifying what to keep unchanged
5. Expecting variation from identical prompts — change the prompt, don't just rerun
6. Not prompting camera motion → static or unpredictable movement
7. Subtitle contamination — use colons, add "(no subtitles)"
8. Vague prompts for complex scenes
9. Ignoring aspect ratio and resolution (480p, 720p, 1080p)
10. Expecting real-time knowledge — models work from training data

Sources

Image:

• How to prompt Seedream 5.0 (Feb 2026)
• Recraft V4 (Feb 2026)
• Run FLUX.2 on Replicate (Nov 2025)
• FLUX.1 Tools (Nov 2024)
• Ideogram 3.0 on Replicate (May 2025)

Video:

• How to make remarkable videos with Seedance 2.0 (Apr 2026)
• How to prompt Veo 3.1 (Oct 2025)
• Wan 2.2 (Jul 2025)
• Compare AI video models (Jul 2025)