Visual PromptCompiler / image editionLocal-only V1
Visual systems, not prompt inflation

Visual Prompt Compiler
One brief. Two model-ready instructions.

Build once, adapt to every image model through an explicit adapter contract. Define what must stay, what may change, and how each reference should be used; V1 compiles that visual spec for Seedream and Nano Banana.

Start with a preset ↓See the signal path
LIVE SPECIMENCHAR / MARA-07
Scar placementSilver hair streakHarbor archiveNo beauty retouch
ADAPTER2 model contracts ready
BriefintentSpeclocksAdaptermodelOutputready
Working surface / no account

Compile a visual brief

Lock what matters, name what can change, then adapt the same visual spec to a real model contract.

Stays hereSaved in this browser only
Task preset
SPEC / 01
01 Outcome
02 Lock the subject
03 Direct the scene
04 Set change boundaries
05 Assign reference roles

Describe references only. Images are never uploaded or stored here.

06 Choose the target
Workflow
Ready to compile locally

The compiler contract

Separate the creative truth from the model syntax.

Your visual identity should not be rewritten every time a model changes. Visual Prompt Compiler keeps a stable source specification, then delegates model details to small, versioned adapters.

Input

Write the visual truth

Name immutable subject details, scene direction, permitted changes, prohibited drift, and the job of every reference.

Compile

Apply a model contract

The adapter changes section order, reference language, settings, edit method, and warnings without weakening your Preserve list.

Reuse

Keep the spec portable

Save up to three profiles in this browser or export JSON. No account, server, or image upload is involved.

TASK / 01

Character consistency

Lock identity, outfit anchors, and reference roles while changing action, scene, or camera.

Open character builder →TASK / 02

Product photography

Preserve geometry, material, label text, and brand codes across hero, lifestyle, and campaign shots.

Open product builder →
Adapter difference, made visible

The same intent does not mean the same prompt.

S5

Seedream 5.0 Pro

Adapter v1.0.0

  • Ordered Image 1…N reference map
  • Point, box, and annotation edit language
  • Provider: 1K / 2K; this V1 plans 2K
  • 10-reference ceiling warning
  • No seed or sequential batch on Pro
Inspect the adapter →
N2

Nano Banana 2

Adapter v1.0.0

  • Conversational reference intent
  • Multi-turn editing language
  • Provider: 0.5K–4K; this V1 plans 2K
  • Character and object fidelity warnings
  • SynthID disclosure in every result
Inspect the adapter →

Model pages remain out of the sitemap until their live six-task QA gate is completed. The local adapters are available now for transparent inspection and testing.

Privacy is an architectural choice

Your brief does not need to leave your browser.

Practical questions

Before you compile

Does this tool generate images?

No. It compiles your visual requirements into instructions for image models. It does not call a model API or create an image.

Are my prompts or references uploaded?

No. V1 runs in your browser. Reference fields describe the role of images, but the tool does not accept or upload image files.

Why not paste the same prompt into every model?

Models expose different reference, editing, resolution, and failure contracts. The adapters keep your creative intent stable while changing the delivery format.

Can a prompt guarantee character consistency?

No. A precise brief and reference roles can reduce drift, but model behavior, source images, editing workflow, and repeated testing still matter.