Precision image
GPT 2
- Input
- Prompt, or prompt plus images
- Output
- Square, wide, or vertical
- Quality
- Medium or High detail
Kiara model catalog
Image models for cinematic creators.
A production-style view of every image model currently exposed in Kiara X: what it is for, what it accepts, what it can output, and how to get the best result from it.
Realism polish
Kiara Realism
- Input
- One source image plus a short style note
- Output
- Source shape preserved
- Quality
- 2K realism polish
Overview
Pick by job, not by name.
The catalog is split into generation, reference editing, enhancement, and roadmap lanes. The important decision is whether the user needs a clean prompt-first image, a reference transformation, or a finishing pass on an existing image.
9
image lanes
9
active or beta
7
reference-first
6
source-sized
Creator workflow
Choose by intent. Finish by quality.
Start with the job the image has to do. New concept, controlled reference, heavy edit, or final polish: each lane has a different strength.
Prompt Reference Edit Finish
- 01
Open brief
Start with text when the image does not exist yet.
Use a prompt-first lane for a new character, product shot, mood piece, or cinematic still. The prompt should define subject, setting, camera, light, wardrobe, and finish before any reference is added. - 02
Locked source
Use reference lanes when identity or pose must survive.
Reference-aware models are best when the user already has a face, outfit, product, pose, or style anchor. Every uploaded image needs a job: identity, body, product, pose, lighting, or mood. - 03
Rebuild scene
Use edit lanes when the image needs a new reality.
Editing models are for changing action, outfit, environment, pose, or combining subjects. Treat the prompt like a production note: who stays, what changes, where the scene happens, and what the final camera sees. - 04
Final pass
Use enhancement lanes after the composition is approved.
Realism, sharpening, upscale, and relight tools preserve the existing image. They are the final-mile pass for texture, skin detail, cleaner edges, stronger light, and delivery quality.
Image models
Current Kiara image stack.
Generate
Prompt-first image models for clean stills, concepts, product scenes, and cinematic looks.
Default Precision image
GPT 2
The clean instruction follower for exact details and multi-reference composition.
The most instruction-accurate general image lane. It is the best choice when the user needs exact objects, text-like layout discipline, product cleanliness, or several references blended into one controlled result.
- Prompt field
- Ratio picker
- Up to 8 refs
- 2 quality paths
- 15-120s estimate
- Input
- Prompt, or prompt plus images
- References
- Up to 8 optional references
- Output
- Square, wide, or vertical
- Quality
- Medium or High detail
How it works
Instruction-first image lane with optional reference support. It is strongest when the brief contains exact layout, object, identity, color, and composition rules.
Input contract
- Prompt is always the main instruction.
- Add up to 8 references when identity, product shape, outfit, or style must be controlled.
- Medium is the default balance; High is for the cleanest detail.
Prompt shape
- Name the final image in one sentence.
- List reference roles clearly: face, outfit, product, pose, room, or style.
- Specify what must stay unchanged and what should be changed.
Use it for
Product shots, exact scene requests, clean edits, multi-reference compositions.
Best practice
Be explicit about layout, object positions, exact colors, and what must stay unchanged.
- Name every object that must appear.
- Describe placement from left to right when layout matters.
- Use references for identity, product shape, or exact visual language.
Avoid
Do not upload references without explaining their role. Do not rely on vague words when exact placement matters.
- Best default image model
- Handles optional references
- Clean product and edit direction
Available Cinematic image
Wan2.7 Image Pro
A cinematic still model for beauty, fashion, people, and detailed scene direction.
A cinematic still lane for beauty, fashion, people, atmosphere, and scene direction. It responds well to cinematic language and can use references when the user wants a specific person, outfit, or location feeling.
- Prompt field
- Ratio picker
- Up to 8 refs
- 1 quality path
- 15-180s estimate
- Input
- Prompt, or prompt plus images
- References
- Up to 8 optional references
- Output
- Square, wide, or vertical
- Quality
- Balanced production finish
How it works
Cinematic prompt and reference lane for people, beauty, fashion, and environment-heavy stills. It benefits from director-style language and clean visual hierarchy.
Input contract
- Works from prompt alone or with up to 8 optional references.
- Uses square, wide, or vertical output.
- The current production lane exposes a single balanced quality path.
Prompt shape
- Write the shot like a film still.
- Define subject, location, action, wardrobe, camera distance, lens, light source, color palette, and texture.
- Use references for the exact face, clothing, location mood, or set design.
Use it for
Fashion editorials, beauty stills, lifestyle scenes, atmospheric portraits.
Best practice
Lead with cinematic context, then subject details, then lighting and camera language.
- Write the scene like a short director brief.
- Specify lighting source, lens feel, palette, texture, and mood.
- Use references for wardrobe, face, or set design, not for every tiny detail.
Avoid
Do not mix too many incompatible style directions. Avoid weak prompts that only say beautiful, cinematic, or realistic.
- Fashion and beauty tone
- Detailed people prompts
- Prompt or reference driven
Reference and Edit
Models that need one or more uploaded images and transform them into the final result.
Beta Reference transform
Nano Banana
One primary reference plus a scene prompt for outfit, setting, and action changes.
A reference transformation lane. Upload one primary image, then describe the new scene, outfit, environment, pose, or mood. It is built for taking a person or object from the source and placing it into a fresh composition.
- Prompt field
- Ratio picker
- 1 required ref
- 2 quality paths
- 20-90s estimate
- Input
- One image plus a scene prompt
- References
- 1 required primary reference
- Output
- Square, wide, or vertical
- Quality
- High-resolution composed result
How it works
Reference transformation lane. One primary image anchors the subject, then the prompt moves that subject into a new scene, outfit, action, or visual style.
Input contract
- One primary reference is required.
- The prompt should say what to keep from the reference and what to replace.
- Aspect ratio can be square, wide, or vertical.
Prompt shape
- Start with: keep the same person or object from the reference.
- Then describe the new setting, pose, wardrobe, lighting, and camera.
- Mention identity-sensitive details that must not drift.
Use it for
Put this person into a new scene, outfit, pose, or visual concept.
Best practice
Tell it what to keep from the reference and what to replace in the final scene.
- Say exactly what must stay from the reference.
- Describe the new setting and wardrobe separately.
- Use clean source images where the subject is not hidden or heavily cropped.
Avoid
Do not use blurry or heavily cropped references. Do not expect it to combine several unrelated people from one image with perfect separation.
- Primary reference required
- Scene transformation
- Character restyling
Available Scene edit
Kiara Edit
Reference-based scene composition for one subject, multiple subjects, or pose control.
A scene-composition editor. It can use one subject, add a second subject, and optionally follow a pose reference while rebuilding the final scene around the written action.
- Prompt field
- Source-sized
- 1 required ref
- Up to 3 refs
- 60-240s estimate
- Input
- Image references plus a scene prompt
- References
- 1 required, 2 optional
- Output
- Source shape preserved with high-end finish
- Quality
- Detailed composite output
How it works
Scene composition and edit lane. It accepts a main source, optional second image, optional pose guide, then rebuilds the final scene around the written action.
Input contract
- Reference 1 is the main source or character and is required.
- Reference 2 can add a second person, object, or visual source.
- Reference 3 can guide pose or framing when body position matters.
Prompt shape
- Write it like a production note.
- State who stays, what changes, where the scene happens, what the action is, and what the final camera should see.
- Call out any identity, outfit, or background details that must stay locked.
Use it for
Changing a subject into a new action, scene, outfit, or multi-person composition.
Best practice
Separate subject identity, new action, environment, and what should stay unchanged.
- Use reference 1 for the main subject.
- Use reference 2 only when a second subject or object matters.
- Use the pose reference to guide body angle and framing, not identity.
Avoid
Do not upload optional references without saying what each one controls. Avoid asking for too many simultaneous body, scene, and identity changes in one pass.
- Source character required
- Optional second character
- Optional pose reference
Available Face studio
FaceSwap PRO
Premium head swap with a 4K finish for higher-end identity work.
The premium face-swap lane. It follows the same body-plus-face workflow as the fast swap, then applies a stronger finishing pass for cleaner high-resolution delivery.
- Prompt field
- Source-sized
- 2 required refs
- 45-180s estimate
- Input
- Body image plus face image
- References
- 2 required references
- Output
- Matches body image shape
- Quality
- 4K premium finish
How it works
Premium face transfer lane with a stronger finishing pass. It follows the same two-image contract as the fast swap but targets cleaner high-resolution delivery.
Input contract
- Two images are required: body or scene first, face second.
- Best when the body image already has the final pose and lighting.
- Use it when the result is meant to be published or sold.
Prompt shape
- Keep the body image as the final composition.
- Pick the cleanest face reference available.
- Use prompt text only for subtle expression, realism, or finish notes.
Use it for
Final-ready face swaps, creator identity work, high-resolution outputs.
Best practice
Use clean, front-readable face refs and a body image with the final pose.
- Use this when the face swap is the final asset.
- Choose body images with clean light and visible facial structure.
- Avoid tiny, blurry, side-profile face references.
Avoid
Do not use it to invent a new body pose from text. Avoid mismatched extreme lighting between the face and body references.
- 4K finish
- Body plus face references
- Higher delivery quality
Enhance
Finishing models for realism, detail, sharpness, and lighting control.
Available Realism polish
Kiara Realism
Texture and realism enhancement for uploaded images with source dimensions preserved.
A realism and texture finisher for existing images. It is not for rebuilding the whole composition; it is for making an image feel more natural, detailed, less artificial, and more production-ready.
- No prompt field
- Source-sized
- 1 required ref
- 2 live controls
- 20-130s estimate
- Input
- One source image plus a short style note
- References
- 1 required source image
- Output
- Source shape preserved
- Quality
- 2K realism polish
How it works
Realism finisher for an existing image. The source image drives structure and size; the short text note only tells the model what kind of realism pass to apply.
Input contract
- One source image is required.
- No aspect-ratio picker: output follows the source image shape.
- Use it after the composition is already correct.
Prompt shape
- Keep the text note short.
- Ask for skin texture, natural pores, realistic fabric, better light, film texture, or clean editorial realism.
- Use it as a polish step, not a rewrite step.
Use it for
Fixing AI-looking skin, improving texture, polishing a finished reference.
Best practice
Keep the note short: film texture, skin realism, natural light, or clean editorial finish.
- Use it after the composition is already good.
- Keep the instruction short and finish-oriented.
- Ask for natural skin, fabric texture, realistic light, or artifact cleanup.
Avoid
Do not ask it to change the entire pose or scene. Do not use it to fix a fundamentally wrong composition.
- Skin texture repair
- AI artifact cleanup
- 2K realism finish
Available Enhancement
Kiara Sharp X
Upload one image and sharpen, denoise, upscale, and finish it for delivery.
A final delivery enhancer. It sharpens, denoises, and upscales an existing image while keeping the original idea intact. Use it as the last step before publishing or sending an image to a user.
- No prompt field
- Source-sized
- 1 required ref
- 1 live control
- 20-60s estimate
- Input
- One source image
- References
- 1 required source image
- Output
- Source shape preserved
- Quality
- 2K fast finish or 4K final finish
How it works
Pure enhancement lane. The source image is the brief; the model sharpens, denoises, upscales, and adds a controlled final finish.
Input contract
- One source image is required.
- No prompt field is used.
- Pick 2K for normal delivery or 4K for final high-resolution output.
Prompt shape
- No prompt needed.
- Decide whether the image deserves 2K or 4K.
- Use it only after the creative image is already approved.
Use it for
Final pass on good images, delivery polish, cleaner social or print output.
Best practice
The source image is the brief. Pick 2K for speed, 4K for final delivery.
- Use 2K for normal posts and fast review.
- Use 4K only when the source is worth preserving.
- Do not use it to fix a bad pose or wrong composition.
Avoid
Do not expect pose, wardrobe, or background changes. Avoid sending low-value drafts through 4K just because the option exists.
- No prompt needed
- 2K or 4K finish
- Subtle denoise and grain
Available Light edit
Light Editor
Preserve the image while changing direction, intensity, and color of light.
A lighting-control tool for finished images. It keeps the subject and composition intact while changing key-light direction, strength, color, and cinematic bloom.
- No prompt field
- Source-sized
- 1 required ref
- 4 live controls
- 25-90s estimate
- Input
- One finished image
- References
- 1 required source image
- Output
- Source shape preserved
- Quality
- Controlled relight pass
How it works
Lighting-control lane. It preserves the subject and composition while rewriting the direction, intensity, color, and cinematic bloom of the light.
Input contract
- One finished source image is required.
- No prompt is needed.
- Use direction, intensity, color, and cinematic mode to control the look.
Prompt shape
- Choose the light direction first.
- Then choose the strength and color temperature.
- Use cinematic mode for bloom, halation, and a softer premium light rolloff.
Use it for
Turning flat images into studio, rim light, warm gold, neon, or moody looks.
Best practice
Choose light direction first, then intensity and color.
- Use it on images that already have a good pose and face.
- Pick one clear lighting direction before changing color.
- Use cinematic mode for bloom, halation, and a softer premium finish.
Avoid
Do not use it before the image composition is approved. Do not expect identity, outfit, or background edits from the light controls.
- No prompt needed
- Lighting direction controls
- Cinematic bloom toggle
Coming next
Visible roadmap lanes that will become production tools after testing.
Video models & pricing
Powered by Seedance 2.0.
Every Kiara X video runs on Seedance 2.0 — a multimodal engine that turns text, reference images, clips and audio into footage. Pick a tier for the quality you need, and pay only by the second you generate. Here are the models, their specs, and exactly how the price is built — with a live calculator.
Three tiers, one engine
Every tier is Seedance 2.0 under the hood — the same multimodal inputs. What changes is the quality ceiling, the top resolution, and the price.
Seedance 2.0 Mini
CheapestFast, low-cost drafts and high-volume runs. Native 480p/720p, upscaled higher.
- Top resolution
- 1080p (upscaled)
- Length
- 4–15s
- From (720p)
- 56 cr/s
Seedance 2.0
Most popularThe balanced standard — sharp, fast multimodal video for everyday work.
- Top resolution
- 1080p (upscaled)
- Length
- 4–15s
- From (720p)
- 73 cr/s
Kiara Dream Pro
FlagshipHighest-fidelity renders for hero shots and finals. Adds true 2K output.
- Top resolution
- 2K (upscaled)
- Length
- 4–15s
- From (720p)
- 91 cr/s
| Capability | Mini | Seedance 2.0 | Dream Pro |
|---|---|---|---|
| Native resolution | 480p / 720p | 480p / 720p | 720p |
| Top resolution | 1080p | 1080p | 2K |
| Clip length | 4–15s | 4–15s | 4–15s |
| Text-to-video | ✓ | ✓ | ✓ |
| Reference images | up to 9 | up to 9 | up to 9 |
| Reference clips | up to 3 | up to 3 | up to 3 |
| Reference audio | ✓ | ✓ | ✓ |
| Relative speed | Fastest | Fast | Standard |
| Price tier | Lowest | Standard | Premium |
Credits by clip length
Longer clips cost proportionally more, and a higher tier costs more per second. Same length, three price points.
Dream Pro Seedance 2.0 Mini
What you're paying for
01
Per-second billing
Every second of generated video has a rate. A 10-second clip costs twice a 5-second one. Higher resolution and higher tiers raise the per-second rate.
02
Reference clips count
When you guide a render with a reference video, we process its full length too — so its seconds are added to the billed total on top of your output length.
03
A small minimum
Very short clips carry a fixed minimum billed length, because every render has a baseline cost. Past that, you simply pay for the seconds you use.
Price calculator
Set it up exactly like you would in the studio. This lands on the same number you'll be charged.
Total seconds of clips you attach to guide the render. Leave at 0 if you're not using reference video.
You'll pay
364 credits
Seedance 2.0 · 720p
Generated footage5s
Billed length5s
Per-clip examples
A standard clip with no reference video, at 720p, in credits. Mini is the cheapest way to generate.
| Model | Resolution | 5 sec | 10 sec | 15 sec |
|---|
Prices include everything — no hidden fees, and reference images are always free. You see the exact cost before you generate.
Production stacks
How the lanes combine in real jobs.
Prompt to publish
New concept, clean final.
Best when the creator starts with only an idea and needs a polished image quickly.- Draft in GPT 2 or Kiara Z MAX 2.
- Run Kiara Realism once the composition is right.
- Finish with Kiara Sharp X only for approved outputs.
Reference redesign
Same subject, new scene.
Best when a face, product, outfit, or pose already exists and the user wants a new visual world around it.- Use Nano Banana for one-reference transformations.
- Use Kiara Edit when a second subject or pose guide is needed.
- Relight or sharpen after the scene is locked.
Identity swap
Body scene plus face identity.
Best when the final pose, outfit, and background already exist but identity must be transferred cleanly.- Use Face Swap Base for fast testing.
- Use FaceSwap PRO for the final high-resolution pass.
- Keep source angles and lighting close for the best face match.
Light direction
Turn a flat image into a cinematic frame.
Best when the subject and scene are correct but the image needs mood, rim light, colored light, or studio polish.- Start from any approved image.
- Use Light Editor to choose direction, color, intensity, and bloom.
- Use Sharp X last if the image needs delivery-grade resolution.
Support overview
What each model is built to handle.
| Model | Best use | Input | References | Output | Quality path |
|---|---|---|---|---|---|
| GPT 2 | Product shots, exact scene requests, clean edits, multi-reference compositions. | Prompt, or prompt plus images | Up to 8 optional references | Square, wide, or vertical | Medium or High detail |
| Wan2.7 Image Pro | Fashion editorials, beauty stills, lifestyle scenes, atmospheric portraits. | Prompt, or prompt plus images | Up to 8 optional references | Square, wide, or vertical | Balanced production finish |
| Nano Banana | Put this person into a new scene, outfit, pose, or visual concept. | One image plus a scene prompt | 1 required primary reference | Square, wide, or vertical | High-resolution composed result |
| Kiara Realism | Fixing AI-looking skin, improving texture, polishing a finished reference. | One source image plus a short style note | 1 required source image | Source shape preserved | 2K realism polish |
| Kiara Sharp X | Final pass on good images, delivery polish, cleaner social or print output. | One source image | 1 required source image | Source shape preserved | 2K fast finish or 4K final finish |
| Kiara Edit | Changing a subject into a new action, scene, outfit, or multi-person composition. | Image references plus a scene prompt | 1 required, 2 optional | Source shape preserved with high-end finish | Detailed composite output |
| Kiara Style | Outfit experiments, wardrobe try-ons, look variations from a single character. | Two references — character first, outfit source second | 2 required references in order | Matches character image shape | Fast clean swap |
| FaceSwap PRO | Final-ready face swaps, creator identity work, high-resolution outputs. | Body image plus face image | 2 required references | Matches body image shape | 4K premium finish |
| Light Editor | Turning flat images into studio, rim light, warm gold, neon, or moody looks. | One finished image | 1 required source image | Source shape preserved | Controlled relight pass |
Prompting
Model-aware prompt rules.
Precision image
GPT 2
Be explicit about layout, object positions, exact colors, and what must stay unchanged.Cinematic image
Wan2.7 Image Pro
Lead with cinematic context, then subject details, then lighting and camera language.Reference transform
Nano Banana
Tell it what to keep from the reference and what to replace in the final scene.Realism polish
Kiara Realism
Keep the note short: film texture, skin realism, natural light, or clean editorial finish.Enhancement
Kiara Sharp X
The source image is the brief. Pick 2K for speed, 4K for final delivery.Scene edit
Kiara Edit
Separate subject identity, new action, environment, and what should stay unchanged.Outfit swap
Kiara Style
The direction anchor is locked — slot 1 always keeps face + hair, slot 2 always provides the clothing. Leave the prompt empty for a full outfit swap. Use a short instruction only to scope it ("only the top", "keep the skirt", "and the shoes too").Face studio
FaceSwap PRO
Use clean, front-readable face refs and a body image with the final pose.Light edit
Light Editor
Choose light direction first, then intensity and color.Roadmap
The next version turns this into a live model guide.
This first page is the static catalog. The next pass can wire it to the live `/api/dream/image` response, expose availability status, add example outputs per model, and let Lexa recommend the right lane from a user brief.