Unified asset studio decision · 16 July 2026

Build the Piexels Studio.

One owned place for client assets, generation, per-post feedback, approvals, publishing, and compounding brand memory. Keep Luma and Kimi. Put Metricool behind the surface as the publishing adapter.

Decision

Model A wins.

Extend the review hub into the operating surface. The differentiator is not another generator. It is the client-isolated learning system around the generator.

A Best Model B: Simplified One Agency. Useful benchmark, wrong system of record.

Two operating models

Model B wins setup speed. Model A wins control, memory, quality, and portability.

Recommended

Model A
Owned studio

Bella orchestrates one Piexels surface. Luma generates. Kimi reasons. Metricool publishes.

Versioned brand packs per client
Assets and prompts remain portable
Feedback becomes reviewed future rules
Best-in-class generation stays swappable

Fastest start

Model B
All-in-one product

Import, generate, approve, and schedule inside one vendor.

Little build work
Native publishing on day one
Generation stack is usually closed
Brand memory and workflow stay with the vendor

Criterion	Model A: owned studio	Model B: existing product
Brand control	Versioned rules, asset locks, prompt lineage, and explicit approval gates.	Fast brand kits and templates, constrained by the vendor's system.
Learning	Every Go, Edit, and Reject can support a reviewed rule change for the next run.	Public materials emphasize analytics, approvals, or saved brand kits. A durable per-post learning memory was not documented for the shortlist.
Multi-client	Client ID and row-level security on every record and file.	Workspaces, brands, or brand books with plan limits.
Generation	Luma UNI-1 and Ray3.2 remain canonical.	Usually locked to in-house models or a vendor-selected mix.
Publishing	Metricool Advanced provides approvals, scheduling, API, and MCP.	Native. This is Model B's clearest advantage.
Lock-in	Low. Assets, rules, feedback, and receipts remain owned.	High. Brand memory and workflow live inside the vendor.

The real Model B field

Simplified is the only shortlisted product that closely matches the complete brief. The others specialize in generation, scheduling, or ads.

Product	Strongest fit	Multi-brand and publishing	Learning signal	Public price	Verdict
Simplified One Agency Best B	Broad import, create, approve, schedule, and report workflow.	10 AI Brandbooks, 70 accounts, approval workflows, and bulk scheduling.	Brandbooks and social insights. No documented per-post learning memory.	$119/month Annual billing	Closest complete product, but it replaces the chosen generation stack.
Predis.ai	Fast AI social batches and autoposting.	Rise: 4 brands and 20 accounts. Enterprise: unlimited brands and 60 accounts.	Analytics and competitor runs, not a documented creative memory.	$79 Rise $249 Enterprise	Strong autopilot, weaker control.
Ocoya	AI agents and social automation.	Gold: 20 workspaces and 50 profiles.	Automation and analytics, no documented closed feedback memory.	$99/month Gold	Good scheduler-first suite.
Metricool	Publishing, approvals, analytics, and integrations.	Advanced: 15 brands, client management, approvals, API, and MCP.	Performance data is valuable input, but not generation memory.	$53/month	Use behind Model A.
Creatify	Ad production, testing, and media buying.	Ad-centered, not a complete organic multi-client studio.	Creative testing data.	$299/month AdMax	Paid-ad specialist.
Icon	Meta ad workflow, assets, creative analytics, and multi-model AssetGPT.	Launches and analyzes ads, not a general organic publisher.	Strong ad performance loop.	$999/month managed offer	Too narrow and expensive.
Captions	Short-form video generation and editing.	No multi-client publishing workflow documented on pricing.	No documented cross-client learning loop.	$24.99 Max $69.99 Scale	Video tool, not a studio OS.
Runway	Model breadth, generation, and asset storage.	No social scheduling or multi-client publishing layer.	No post-performance learning loop.	$28/month Pro Annual billing	Generation platform, not Model B.
Superside	Human creative team, Brand Brain, custom models, and performance iteration.	Managed service, not a solo-founder control surface.	Strong human-assisted learning.	$15,000 monthly minimum plus $1,000 software	Economically out of scope.

Unavailable sources: AdCreative.ai's official pricing page and Higgsfield's pricing content were not accessible in this research runtime. Their prices are not guessed.

One surface, replaceable engines

Wil and Bella stay in the studio. Vendor boundaries remain behind it.

ClientsAssetsBriefGenerateReviewPublishLearn

Surface: Authenticated studio routes extend the existing review hub: Clients, Asset Library, Campaigns, Review Queue, Calendar, Published, and Learnings.
Data: Supabase Postgres stores clients, versioned brand packs, campaigns, posts, feedback events, jobs, approvals, publish receipts, and performance snapshots.
Isolation: Every record and storage object is keyed by client_id. Row-level security prevents one client's assets or memory from crossing into another.
Reasoning: Kimi turns the active brand pack and campaign brief into structured post specs and Luma prompt packs.
Generation: Luma Agents API runs UNI-1 images and Ray3.2 videos or edits. The studio stores outputs with model, prompt, inputs, and source hashes.
Quality: Existing visual QA remains a release gate. A post cannot enter the approved queue without an explicit review decision.
Publishing: Metricool owns network tokens, schedules, delivery, and performance retrieval. The studio owns approval, assets, and learning.

The feedback loop is the moat.

Review each post: Go, Edit, or Reject.
Add reason tags and one short note.
Group repeated feedback by client and asset type.
Kimi proposes a rule change with evidence links.
Approve a new brand-pack version.
Use that version in the next Luma run.

Guardrail: performance can suggest the next hypothesis, but it never silently rewrites brand rules. One outlier post or ambiguous comment cannot poison the client memory.

Build effort

Useful quickly, hardened deliberately.

Pilot8 to 12focused build days

Schema, isolation, brand packs, and asset import: 2 days.
Campaign briefs, Kimi orchestration, and Luma async jobs: 2 to 3 days.
Per-post review, approval state, and feedback events: 2 days.
Metricool adapter, receipts, and status sync: 1 to 2 days.
QA, recovery, audit log, and responsive polish: 1 to 3 days.

Hardened production: 3 to 4 weeks total. Add permissions, retries, idempotency, budgets, alerting, performance ingestion, rule promotion review, export, and a client-safe approval view.

Monthly operating budget

Same planning workload as the endpoint decision: 120 five-second 720p clips and 500 2K images.

Incremental planning budget$193.81

About $19.38 per client at 10 clients, or $12.92 at 15.

Luma media workload: $81.45
25% media reroll reserve: $20.36
Kimi Allegretto: $39.00
Metricool Advanced, 15 brands: $53.00
Optional Supabase Pro: +$25.00
Optional Vercel Pro: +$20.00

Current infrastructure billing was not inspected in this lane. If both dedicated paid infrastructure lines are new, the planning ceiling is about $239/month before usage overages.

The honest tradeoff

Model B can post sooner. Model A creates maintenance responsibility.

Accept two weeks of build time. Moving into Model B now would undo the endpoint decision, trap the most valuable data inside another vendor, and still leave the central problem unsolved: learning from per-post feedback without blending client brands.

Sources

Current official product, pricing, and documentation pages fetched 16 July 2026.

Design evidence: this report preserves the approved platform-endpoints hierarchy, quiet hub shell, locked cream, ink, and lime palette, and existing Fraunces plus Inter typography. Refero MCP was not configured in this headless runtime, so no Refero pattern is claimed or substituted.