Current price
$0.0018 / image
Product use cases · Studio pipelines · Images and video
Keep the same character recognizable across every scene.
Use an approved character pack and the same test scenes to compare reference-capable models. Record the settings that preserve identity best, then review the estimate before producing a larger set.
- Best for
- Teams creating recurring fictional characters across image and video content.
- What you get
- A reusable character pack and tested settings that make identity more consistent across supported models.
Concept illustration
A creative brief to try
Change one character detail at a time
Start with one main portrait of a fictional 30-year-old adult with short auburn hair, green eyes, and freckles. Make clothing, location, and profile variants from that original.
- 01
From the main portrait, change only the outfit to a cream sweater. Keep the pose and background for an easier comparison.
- 02
Start again from the main portrait and change only the setting to a café. Save this separately from the sweater variant.
- 03
Use the main portrait again for a profile view. Avoid feeding each edited result into the next edit.
Before you keep the result
Compare face shape, eyes, hair, and freckles with the original. Keep variants that remain recognizable; a reference does not promise identical pixels.
Start with your own reference images, follow the scene plan and review each result.
Open this scene in the Playground →Recommended models
73 matching models
Matches use the tasks and required capabilities published in the catalogue. Availability and current prices come from each model page.
Current price
$0.0018 / image
Current price
$0.0018 / image
FLUX.1 Schnell LoRA · Image Editing
Image editing
Current price
$0.006 / image
Current price
$0.011 / output second
Anima Turbo · Image Editing
Image editing
Current price
$0.012 / image
P-Image LoRA · Image Editing
Image editing
Current price
$0.012 / image
Z-Image Turbo LoRA · Image Editing
Image editing
Current price
$0.012 / image
What one run costs
Worked out from the live catalogue, at the settings named under each figure.
Catalog estimates refresh hourly. Check the estimate for your settings before running; final charges follow the model’s usage and billing policy.
How to build it
Keep the versioned character originals in your own storage and reuse accessible HTTPS references. Temporary uploads last 1 day, so refresh them when needed. Record reference order and settings for comparison, then check the face, hair and identifying details in each result; repeated inputs can still produce variation.
- 01
Ask for an upload ticket
POST
/api/v1/common/upload-urlThe ticket is valid only for the byte count you declared: JPEG, PNG, WebP and GIF up to 10 MiB each. Media that already has a public HTTPS URL skips this step entirely.
- 02
Commit the uploaded file
POST
/api/v1/files/{fileId}/commitThe commit checks the bytes and hands back a spicy:// reference, which is what a model's media fields accept. Committing the same file again changes nothing.
- 03
Create the task
POST
/api/v1/jobs/createTask202 means accepted, not finished. Send an Idempotency-Key, and if the price must not move, send the jobs/quote result together with the amount you agreed to.
- 04
Read the task back
GET
/api/v1/jobs/recordInfoOne task ID, one record. Queued, running and terminal states share the same shape, and this response is never cached.
- 05
Move the file into your own storage
POST
/api/v1/common/download-urlMint a short-lived download URL for the output and copy it into your own bucket within the same job. Results stay on our CDN for 14 days.
Copy a request
This example starts with the scene described on this page. Fields come from the selected endpoint’s schema. Replace any example.com reference URL with your own image or committed file URI before sending.
GPT Image 2.5 Sunburst · Image Editing →
Where this usually snags
Platform behaviour that is cheaper to design around than to discover in production.
Reference fields need a model that takes references
Input is validated against that endpoint's own schema and extra properties are refused, so a reference sent to a model without the capability is a 400, not a field quietly ignored.
Inline images are the narrow path
A data URI is capped at 1 MiB decoded per image, 16 images per request and 2 MiB for the whole JSON body. Anything larger goes through an upload ticket or a public HTTPS URL.
202 is a commitment
An accepted task runs to the end and the hold stands; there is no cancel. While the parameters are still moving, quote first — the same checks, nothing created.
Launch checklist
- 01Version the approved character pack and keep the original source assets unchanged.
- 02Run the same representative scenes through candidate models before choosing one for production.
- 03Record reference order, crop, resolution, seed and model version for approved results.
- 04Review the estimate before a larger batch; once a task is accepted, it cannot be canceled.
Models that match this workflow
Explore model families for this workflow. Open a model page to review its inputs, pricing and current availability.
Frequently asked questions
73 published models match the listed tasks and required capabilities. Compare their inputs, examples, availability and current prices before choosing one.
A first frame is useful when the composition is already established. Reference images are useful when the same character needs to appear in different compositions. Choose for the shot and the inputs your model supports.
Check the reference-count limit in the selected endpoint’s input schema. The Playground uses that schema to display its controls. Do not assume another model or version accepts the same number of images.
A clear first frame and reference images help reduce character drift. You can also reuse the last frame of a clip as the next clip’s first frame, or use a model that supports video extension. Check continuity across each generated segment.
Changing poses, settings and reference material at once can increase variation. Adjust one aspect at a time, version your references and record the seed when supported. The same seed does not guarantee an identical result.
About this workflow
Character consistency depends on both the model and the workflow. Establish reference assets, reuse them across image editing and video generation, and record asset versions and settings. Check the current input schema for reference counts, duration and other limits, and review each result for continuity.
