English quick take
Use the image model for the photograph and the local editor for the typography. That split gives you a cleaner poster and avoids the garbled text that image models often produce.
Girigo AI Studio 10 min read
Build a vertical key-art frame, keep the cast original, and place the title with a local editor instead of asking the image model to spell it.
English quick take
Use the image model for the photograph and the local editor for the typography. That split gives you a cleaner poster and avoids the garbled text that image models often produce.
Local look lab
Upload a photo you own. The canvas, grade, grain, and caption stay in this browser.
Choose a photo to start. Nothing leaves this browser.
Prompt studio
Pick a look, set the direction, and copy a prompt that names what must stay fixed.
Direction: cinematic portrait; preserve identity; one person.
Create an original vertical K-drama key-art photograph. A fictional young adult couple stands apart in a rain-soaked Seoul alley at night, lit by a single convenience-store sign and a cold blue street lamp. One holds a phone with a blank glowing screen; the other looks back over a shoulder. Use cinematic anamorphic photography, realistic rain, shallow depth of field, and a restrained teal-and-warm-red palette. Leave clean negative space at the top and bottom for a title. The people must be original fictional characters, not real actors or celebrities. No text, no logos, no watermarks, no extra people.
Three useful outcomes
The prompt reserves negative space instead of filling every corner.
Open the studioThe local editor adds the caption after the image is generated.
Open the studioThe prompts ask for fictional characters, not celebrities.
Open the studioPrompt lab
A strong photographic base makes the text layout easier.
Create an original vertical K-drama key-art photograph. A fictional young adult couple stands apart in a rain-soaked Seoul alley at night, lit by a single convenience-store sign and a cold blue street lamp. One holds a phone with a blank glowing screen; the other looks back over a shoulder. Use cinematic anamorphic photography, realistic rain, shallow depth of field, and a restrained teal-and-warm-red palette. Leave clean negative space at the top and bottom for a title. The people must be original fictional characters, not real actors or celebrities. No text, no logos, no watermarks, no extra people.
Avoid: real actors, celebrities, text, logos, watermark, extra people
A budget edit route is enough when the composition is simple.
Create an original vertical horror-mystery poster photograph. A fictional young woman stands in a dim apartment hallway, facing a phone that casts a cold light on her face. A wall clock behind her is out of focus; the hands are not readable. Use deep shadows, a narrow pool of light, realistic skin texture, and a muted green-and-red palette. Leave dark negative space at the top for a typographic title and at the bottom for credit blocks. The subject is an original fictional character, not a real actor. No text, no logos, no watermark, no extra people.
Avoid: real actor, text, logo, watermark, extra people
Character routes help keep two original faces distinct.
Create an original vertical romance-fantasy key art. A fictional couple stands on a rooftop above a sleeping city, with a large moon and thin clouds behind them. One wears a dark coat; the other wears a pale dress with a subtle star pattern. Use soft cinematic light, realistic fabric, and a deep blue-and-silver palette. Keep the faces small enough to feel like a poster but clear enough to read as two distinct people. Leave clean negative space at the top and bottom for typography. No text, no logos, no watermark, no third person.
Avoid: text, logo, watermark, third person, blurry faces
WaveSpeed, SeeAPI, fal, Replicate, Kie
| Provider | Best for | Typical route | Docs |
|---|---|---|---|
| falSpeed-first | Fast hosted queues and strong model coverage | FLUX, Kling, and other hosted media models | Docs |
| WaveSpeedBudget-first | Low-cost image edits and image-to-video drafts | Seedream edit, FLUX Kontext, Wan image-to-video | Docs |
| KieCharacter-first | Nano Banana image edits and Kling video routes | Nano Banana, Kling video | Docs |
| SeeAPIRouting layer | One API surface for several model families and fallback routes | Image and video inference endpoints | Docs |
| ReplicateExperiment-first | Broad catalogue for experiments and batch comparisons | Community and first-party image/video models | Docs |
Step by step
Failure modes
Editorial brief
A K-drama poster is two jobs: a photograph with a clear emotional beat, and a typographic layout that survives a thumbnail. Image models are good at the first job and unreliable at the second. The practical workflow is to generate a clean vertical plate with deliberate negative space, then set the title in the browser editor where every letter is under your control.
Decide the title length before you generate. A two-word title needs a different composition than a long subtitle. Keep the faces in the upper third and the lower third dark enough for credits. If the image model adds fake text, remove it in the local editor or regenerate with a stronger "no text" constraint. A poster with no text is a better starting point than a poster with broken text.
Deeper guide
The prompt is only half the job. The field note walks through the failure modes, the identity check, and the model comparison behind this page.
Open the field noteFAQ
Use it for commentary or a clearly labeled fan concept only. Do not imply official affiliation, and do not use real cast photos without permission.
Image models often misspell words or distort letterforms. A local editor gives you control over the font, spacing, and legibility.
A vertical 2:3 ratio is a familiar poster shape. A 4:5 ratio works better for social feeds. Pick one before you generate and keep the same crop through the workflow.
Only with an original cast, a model license that permits commercial use, and a clear disclosure that the image is AI-assisted. Do not use a real person's likeness without consent.
Checked sources