Production Console
One shoot. One next action.
checking engine
PRODUCTION CONSOLE
Loading the current production state…
The console compresses the full Studio pipeline into one next action.
PRODUCT 001 · White Tee / Back Print
Four production references are explicit roles. In production these live in private Supabase Storage, not on the web server.
Hard fidelity rules
MODEL 001
Identity reference. Scene, pose and styling may move; the person should not.
Preserve facial identity; no glam transformation.
Natural skin texture; no synthetic beauty-filter surface.
Styling can change, identity cannot.
Create a coherent five-shot shoot
Front / Back / Hero / Detail / Object. Generation Gate can test the critical product anchors before spending on the full pack.
Clean SilenceQuiet studio clarity, soft material truth.
Atmospheric DistanceAir, separation, landscape and concrete.
Quiet Urban TensionArchitecture, dusk, restrained urban pressure.
Scene not previewed yet.
Shoot Bible
Preview a scene or create a shoot. Creative Director V2 will lock the concept before any rendering starts.
PREFLIGHT DIVERSITY DIRECTOR
Separate the five frames before spending on them.
This runs on the Shoot Bible before generation. It compares planned camera distance, viewpoint, subject mode, action and composition. If two frames are trying to do the same visual job, only the redundant shot brief is rewritten.
Preview or create a shoot first.
Immutable during preflight
Product DNA, prints, garment anatomy and source references.
Model identity and the chosen scene/world.
Narrative role: Front / Back / Hero / Detail / Object.
Paid calls: none. This layer is deterministic and local.
Automatic rewrites
Only camera/pose/composition fields may change. Product truth never does.
Planned pair comparison
High overlap is corrected before render; medium overlap stays visible as a warning.
Motion source
Animate only a frame worth keeping. Previewing the motion prompt is free.
Motion Director
The still is treated as the anchor. Prompt describes only the evolution so face, garment and print drift less.
Select a shot and preview motion.
Image Engine Benchmark
Render the same shot brief through up to three engines. Same product references, same scene, same instruction. Pick the winner only after checking both fidelity and beauty.
No benchmark started.
Semantic Vision status loading…
TECHNICAL
Pixel-level sanity
Decode, resolution, ratio, exposure, contrast, clipping and local detail.
PRODUCT DNA
Semantic contract
Identity, correct garment side, exact print placement, neck-label anatomy and invented-graphic detection, compared against the stored reference set.
GATE
Hard failures matter
A hard QA failure on Front/Back blocks Generation Gate even if manual scores are high.
BEAUTY CRITIC V2
A demanding advisor, not the final judge.
Critic separates originality, composition, realism, emotional force, brand fit and product readability. It also flags generic commercial-fashion grammar. Its score is advisory only: it cannot approve a frame, change Product DNA or replace your Beauty score.
Beauty Critic status loading…
Critic policy
8.0 = strong enough to publish, not merely competent.
9.0+ = hero candidate; this should be rare.
Generic premium-fashion polish is penalized.
Product QA remains completely separate from aesthetics.
Controller cannot auto-accept from Critic score.
Semantic critique is never run automatically if it may cost money.
PAIRWISE VISUAL DIRECTOR
Does every frame earn its place?
This layer compares images against each other, not against a generic quality score. It looks for repeated silhouette, camera distance, composition, light and narrative function, then nominates the frame that should change.
Pairwise status loading…
Sequence contract
FrontRecognition
BackOpposite view
HeroExpansion
DetailMacro evidence
ObjectTrace / interruption
Local comparison is free and acts as a screening signal. Optional semantic pairwise vision is manual and never runs automatically if it may cost money.
CURATOR / FINAL SELECTOR
Judge the series, not five isolated images.
Curator checks coverage, series cohesion, repetition, narrative function and hero strength across the whole shoot. Product QA is an exclusion rule: a beautiful but physically wrong frame cannot enter the final set.
Open a shoot to curate it.
Final-set policy
One Hero is recommended, never silently approved.
Front / Back / Hero / Detail / Object must all be covered.
Hard QA failures are excluded before aesthetic ranking.
Duplicate detection is conservative until pairwise vision is connected.
Locking the set requires your manual scores + Approve on all five frames.
Curator never publishes anything.
Final selection lock
This only freezes the curated five-frame set and Hero choice inside Studio. It does not publish or schedule anything.
DIAGNOSE
QA → repair scope
Typography, garment anatomy, identity, material, continuity and technical failures use different instructions.
PRESERVE
Edit, don't redesign
Face, pose, scene, camera and already-correct product areas stay locked unless they are the diagnosed problem.
SPEND GUARD
Maximum two attempts
No automatic second paid repair until real semantic vision QA can prove the first repair failed.
AUTONOMOUS SHOOT CONTROLLER
One system decides what happens next.
Controller combines Product DNA, Visual QA, manual Fidelity/Beauty, Repair Director, Generation Gate and Calibration. It can accept, hold, repair, regenerate, vary or stop a frame. Product truth always outranks aesthetics.
Open a shoot to analyze it.
Controller policy
Observe = recommendation only.
Guarded = recommendation + stored plan; no paid generation.
Autonomous = may accept clean manually-scored frames.
Paid Repair / Regenerate requires its own explicit switch.
Controller-generated decisions never train Calibration as your taste.
Run controls
Paid generation is OFF by default. Turning it on only gives the controller permission to use the selected per-wave budget; hard attempt limits still apply.
CALIBRATION / LEARNING LAYER
Learn the threshold, not the taste away.
Studio learns from your scores and Approve / Reject decisions. It may become stricter when the evidence is consistent, but it can never lower Product Fidelity below the hard IWANTTOEATAIR floor or override a QA hard failure.
Loading calibration profile…
Learning guardrails
Minimum 6 decisions globally before adaptation.
Minimum 3 decisions for a shot type before that type contributes.
One click never changes a threshold.
Fidelity can only stay at 8.0 or move upward.
Visual QA hard failures always block.
Thresholds by shot type
Observed accepted floor is based on the lower edge of frames you actually approved, not on an average aesthetic model preference.
What you reject
Optional rejection notes are converted into a small reason taxonomy. Raw notes remain available in learning history.
Recent learning evidence
Only manual scores and explicit decisions count as taste evidence. QA contributes facts, not preference.
Engine status
One backend owns UI, uploads, jobs and outputs. Provider credentials remain server-side.
Not tested.
ComfyUI / CatVTON workflow contract
Install an API-format workflow later. Studio patches named nodes, so numeric node IDs may change.
IWEA_MODEL_IMAGE
IWEA_GARMENT_IMAGE
IWEA_POSITIVE_PROMPT
IWEA_NEGATIVE_PROMPT
No workflow checked.