An ad factory for DTC ecommerce is an end-to-end creative system that ships on-brand creatives at volume from a documented Brand DNA spec, generation pipeline (Midjourney, Runway, brand-fine-tuned LoRAs), and a fast-loop QA gate. Done-for-you delivery, priced after the free audit.
TL;DR
- The leak is brand drift. Off-brand assets creep in as volume grows.
- Brand DNA + LoRA fixes drift. A documented spec plus brand-fine-tuned LoRA keeps assets on-brand.
- Volume + velocity beat bespoke. Done-for-you delivery, priced after the free audit.
- Disclose the AI on Meta and TikTok. Both auto-detect; hiding it kills ad-account standing.
- FTC compliance is non-negotiable for performance claims. $43,792 per violation in 2026.
Why creative budgets leak · Brand DNA Sprint · Generation pipeline · QA gate · Monthly breakdown · Ad factory vs agency · Platform-specific creative · FTC + claim substantiation · Cost + ROI math · What goes wrong · FAQ
1. Why DTC creative budgets leak
Most DTC brands at $5-30M revenue spend 8-15% of revenue on creative production. The Lucidpress State of Brand Consistency report estimates 23-35% of those assets drift from the brand book. Multiply by an annual creative budget and you have $50-150k in pure brand-drift cost in the median mid-market DTC brand.
Drift compounds across three surfaces. First, the channel surface - what works on Meta does not work on TikTok and gets reformatted off-spec. Second, the freelancer surface - rotating freelancers each apply slightly different palette, slightly different type, slightly different photography rules. Third, the urgency surface - "we need a Black Friday creative by Wednesday" gets shipped without QA. The third is usually the largest drift driver: rush jobs skip the checks that keep assets on brand.
The leak is not just brand inconsistency. It is also test velocity loss. Agency stacks ship 5-15 creatives per month and run them as test cells; ad factory stacks ship far more and run them through the same media spend. The unlock is not "better creative" - it is more shots on goal at the same budget. Ask any creative partner how they decide which creative to scale; the brands that ran ad factories had the answer because they had the data.
2. The Brand DNA Sprint (7 business days)
The Brand DNA file is the operational artifact that prevents drift. Not a brand book PDF that lives on a Notion page nobody reads. A 12-15 page operational document that the generation pipeline and the QA gate both reference programmatically.
2.1 Day 1 - Intake call (45 minutes)
One call with the founder, the head of marketing, and ideally the original brand designer if they are available. Topics: brand origin story, current sales surfaces, top 3 brand concerns, top 3 creative tests they wish they could run, the assets they wish they could ship faster.
2.2 Day 2-4 - Documentation build
Brand DNA file structured into seven sections: typography (primary, secondary, system fallbacks; weights; tracking; pairing rules), palette (hex, RGB, CMYK, Pantone, accessibility-AA contrast pairs), photography (hero, product, lifestyle, mood; lighting rules; angle library), voice (positioning statement, tonal scales, banned phrases, signature phrases), layout (grid, spacing scale, hierarchy rules), motion (timing curves, transition library, dwell rules), composition (rule-of-thirds variants, negative-space rules, focal-point hierarchy). Lives in Notion for editorial access plus Figma for visual reference.
2.3 Day 5 - LoRA training
Fine-tune a brand-specific LoRA on the existing creative archive. Minimum training set: 80 high-quality assets across the brand surfaces (paid social, organic, email, retail, packaging, lifestyle photography). Training time: 2-4 hours on a rented GPU; iteration to acceptable fidelity typically 1-3 cycles.
2.4 Day 6 - Pipeline assembly
Wire Midjourney with the brand LoRA, Runway for motion, ComfyUI for batch generation, brand-DNA-aware prompt template library, and a quality-controlled output bucket. Test with 5 reference assets - generate 5 variants of each, score against the QA dimensions, iterate until 4-of-5 pass at 80%+ fidelity.
2.5 Day 7 - First monthly batch
Generate the first month's batch (paid social, organic, email, retail, motion). Each asset runs through the QA gate before delivery. Day 7 ends with a delivery review call where the founder approves the final 40 (or rejects 5-10 for revision before going live).
3. The generation pipeline in detail
The pipeline is three layers: input layer (prompts plus reference imagery plus brand LoRA), generation layer (Midjourney for stills, Runway for motion, occasional Flux or Stable Diffusion for surfaces where Midjourney loses brand fidelity), output layer (export plus retouching plus QA scoring).
Prompt template structure. Every prompt is built from four blocks: subject (the actual product or scene), style anchor (locked from the Brand DNA palette and photography rules), composition rule (grid plus negative-space plus focal-point spec), constraint block (aspect ratio, output dimensions, stylistic guardrails). Templates live in a versioned library; A/B testing happens at the template level, not the individual prompt level.
Brand LoRA versioning. Every brand LoRA carries a version tag. New LoRA versions train on additional approved assets (typically every 60-90 days). Old prompt templates re-test against the new LoRA before promotion to ensure fidelity has not regressed. The discipline mirrors the SSOT pattern described in why AI pilots fail to reach production - the brand DNA file plus LoRA is the SSOT for creative.
Retouching pass. Generated assets get a 30-90 second retouching pass in Photoshop or Affinity for cleanup (eye corrections, hand fixes, edge clean), text overlay (typography from the Brand DNA), and platform-specific export. Roughly 70% of stills ship without retouching; 30% need light cleanup; under 5% require regeneration.
4. The QA gate: 5-dimension scoring
The QA gate is what separates an ad factory from a generation farm. Every asset is scored on five dimensions before it ships. Anything below 80% on any dimension goes back for regeneration; anything below 60% on two or more triggers a prompt-template review.
| Dimension | What it measures | How scored |
|---|---|---|
| Brand fidelity | Match to Brand DNA spec on palette, type, photography rules | Manual + LoRA-fidelity classifier (0-100) |
| Composition | Grid alignment, focal point, negative space, rule-of-thirds variant | Manual against template (0-100) |
| Conversion potential | Hook strength, CTA legibility, scroll-stop in feed mockup | Internal scorecard (0-100) |
| Compliance | FTC claim flag, platform-policy flag, model-release flag | Boolean (pass / hold for review) |
| Technical | Resolution, aspect ratio, file size, format, colour profile | Automated check (pass/fail) |
The brand-fidelity classifier is the unique technical surface. We train a lightweight classifier on the same dataset as the LoRA, then score every output against expected fidelity. Drift below 80% triggers a regeneration. Drift below 60% triggers a prompt-template audit. The pattern catches LoRA decay early - usually before the human reviewer notices.
5. What an on-brand monthly mix looks like for a $5-15M DTC brand
The monthly breakdown is calibrated for the typical mid-market DTC brand running paid plus organic across Meta, TikTok, email, and one retail or print surface. The composition shifts for brands heavy on one surface - a TikTok-first brand might ship 60% motion versus the typical 35%.
| Asset type | Spec | Primary use |
|---|---|---|
| Paid social statics | Meta + TikTok creative testing | |
| Short-form motion | Reels, TikTok, Shorts | |
| Organic carousels + posts | Instagram, Pinterest, LinkedIn for B2B-leaning | |
| Email banners + hero images | Klaviyo drops + flows | |
| Print or OOH | Retail signage, packaging, OOH |
The count per asset type is set after the audit, against your channels and media budget. Brands in launch windows or category-expansion phases ship more in a single month. The constraint is not generation capacity but QA throughput.
6. Ad factory vs creative agency vs in-house
The mistake most $5-30M DTC brands make is choosing one model. The correct answer is usually all three, in different roles.
| Model | Best for | Cost band | Output band | Watch out for |
|---|---|---|---|---|
| Creative agency | Brand repositioning, hero photoshoots, big-bet campaigns | $8-15k/month or $25-80k/project | 5-15 assets/month | Ask for a live system, not a deck (see audit-first AI consultancy) |
| Ad factory (kratt or similar) | Weekly creative testing, paid variants, organic cadence, retail at scale | Priced after the audit | Volume set after the audit | Skip the Brand DNA, get slop |
| In-house creative team | Day-to-day brand judgment, design system maintenance, executive presence | $120-300k+ all-in | Variable | Capacity caps; bus factor |
The pattern that works for most $5-30M DTC brands: 1-2 in-house creative leads (judgment + brand custodianship), 1 creative agency on retainer for big-bet work (twice a year), 1 ad factory running the always-on production surface. The factory carries the always-on volume, so the agency retainer stops being stretched across work it was never priced for.
7. Platform-specific creative considerations
The same brand DNA generates platform-specific output. The differences matter at the prompt-template level, not the brand-DNA level.
7.1 Meta (Facebook + Instagram)
Meta's algorithm now favours short hooks (under 1.5 seconds), face-forward imagery on Reels, and authentic-feeling motion over polished. Meta business help ships an updated creative best-practices guide twice a year; the latest emphasises captions burned in (90% of users watch with sound off) and faster scene cuts (1-2 sec scenes test better than 3-5 sec).
7.2 TikTok
TikTok algorithm favours native-feeling content over polished. Aspect 9:16 only. Captions native (not burned-in) for accessibility. Sound on by default. Strong hook in first 0.8 seconds. TikTok Ads Manager publishes a weekly creative trend digest worth subscribing to. Disclosure rule: AI-generated content must be tagged at upload.
7.3 Pinterest
Pinterest favours portrait (2:3 or 1000×1500), still imagery over motion, search-intent context, and longer-form descriptive captions. The most under-used surface in DTC because it does not feel "performance" - but the conversion math is real for home, beauty, fashion, and food categories.
7.4 Email (Klaviyo, Customer.io)
Email banners need 600×400 or 1200×600 retina, sub-100kb file size, alt text written by hand (not auto-generated), and accessibility-AA contrast. Test brand-DNA email assets against stock template imagery and keep whichever earns the click.
8. FTC + claim substantiation compliance
Performance-claim ads (before-after, customer-result, weight loss, skin transformation, time-saving claims) require substantiation under the FTC Endorsement Guides. The 2024 update tightened disclosure requirements; the 2026 enforcement environment is materially stricter than 2022 was.
Three rules to bake into the QA gate. First, every claim image (before-after, result transformation, time-elapsed) must reference documented backup in the brand-asset wiki. The reference is a wiki ID number written into the asset metadata. Second, every endorsement (customer testimonial, influencer review, expert quote) needs documented consent, paid-relationship disclosure where required, and a working backup link. Third, every "results not typical" or material-disclaimer line needs to be readable - not 4pt grey text in the corner.
FTC fines for non-compliant DTC ads ran from $43,792 per violation as the standard penalty band in 2026. The bigger risk for DTC brands is platform action: Meta, TikTok, and Google now share enforcement signal across networks, so a violation on Meta can trigger TikTok review automatically.
Build the compliance dimension into the QA gate from Day 1. Every asset gets a boolean compliance check before it ships, and any flag holds the asset for human review. The pattern is the same SSOT discipline applied to creative; the operational side is covered in our automation work.
9. Cost + ROI math at three brand sizes
| Brand size (revenue) | Current creative spend | Ad factory cost | Output increase | Annual savings |
|---|---|---|---|---|
| $5M DTC | $45-65k/year creative agency | €55-72k/year ad factory | 3-5x assets, 6-8x test velocity | ~$0 cash, large velocity wedge |
| $15M DTC | $140-180k/year | €72-110k/year | 3-5x assets, 6-10x test velocity | $60-100k |
| $50M DTC | $420-650k/year | €110-180k/year | 4-6x assets, 8-12x test velocity | $300-500k |
The savings band widens with size because larger brands spend disproportionately on agency overhead per asset. The velocity wedge (more creative shots at the same media spend) is often a bigger lift than the cash savings - DTC brands that test 10x more creative variants per month find winning ads they would not have surfaced otherwise. Run the Revenue Leak Heatmap for your brand-specific number.
10. Five things that break ad factories
- Skipping the Brand DNA Sprint. Vendors that go straight to "generate assets" produce slop. The Brand DNA is not optional; it is the SSOT for creative.
- No QA gate. Without a 5-dimension scoring rubric, drift creeps in by month 3. The brand-fidelity classifier should automate the bulk of QA scoring.
- Single-platform optimisation. Optimising only for Meta means your TikTok creative dies on arrival. Build platform-specific prompt templates from Day 1.
- LoRA staleness. Brand LoRAs need retraining every 60-90 days as new approved assets accumulate. Stale LoRAs drift quietly; the classifier catches it before the human reviewer does.
- FTC + platform compliance gap. Performance claims without documented backup, AI-generated content without disclosure - both kill ad-account standing fast. Bake compliance into the QA gate.
Cross-vertical patterns: the same five failure modes show up in voice agents (voice-agent failure patterns) and automation (why AI pilots fail to reach production). The unifying theme is end-to-end discipline, not creative judgment.
11. Companion services for DTC brands
The ad factory closes the creative production surface. Three companion services close the operational and revenue surface:
- Ecommerce automation. The ecom automation pillar covers post-purchase flows, abandoned cart recovery, customer journey orchestration around the creative testing loop.
- DTC website. The ecom website-generation pillar builds landing pages and PDPs on the same brand DNA; scope and timing are set after the free audit.
- Voice agent for ecommerce support. The ecommerce-support voice agent handles inbound order, return, and product questions around the clock - critical for DTC brands running paid traffic that generates after-hours interest.
- Listing video at scale. The listing video service generates short-form product demonstration video that performs against Reels, TikTok, and Shopify product pages.
Cross-pillar pattern: each surface uses the same end-to-end discipline (SSOT, QA gate, runbook, on-call). Sibling ad-factory verticals: real-estate ad factory, hospitality ad factory, B2B SaaS ad factory, health and wellness ad factory, franchise ad factory, fashion ad factory, food and beverage ad factory.
12. What to ship this week
Document your Brand DNA. One page is fine for v1. Hand it to your next freelancer or agency and watch quality improve 30-50% the same week. Then run the Agency Audit on your current creative vendor stack to surface where the spend is leaking. Or book a 30-minute review with a kratt operator who has run an ad factory for a brand at your stage.
13. Frequently asked questions
What is an ad factory for DTC ecommerce?
An end-to-end creative system that ships on-brand creatives at volume from a documented Brand DNA spec. Three components: brand documentation, generation pipeline (Midjourney + Runway + brand-fine-tuned LoRA), and a 5-dimension QA gate.
How is this different from a creative agency?
Volume and velocity. Agencies: bespoke concepts at €8-15k/month for 5-15 assets. Ad factories: branded production at volume, priced after the free audit. Complementary, not substitutes.
How long does the Brand DNA Sprint take?
One scoped sprint, delivery date agreed in writing. 45-min intake, documentation + visual spec, LoRA training + pipeline, first monthly batch.
Will the assets look like AI?
Only if your brand looks like AI. Inputs are your existing brand. Generic output means the DNA work was skipped or the LoRA training set was too small.
What about print, OOH, and packaging?
Same brand DNA generates 1080×1080 statics, 9:16 motion, 50×70 posters, 300dpi CMYK print-ready PDFs. Lock the colour profile spec on Day 1.
How does FTC compliance affect output?
Performance claims require substantiation. QA gate includes a compliance dimension that flags any claim or testimonial needing documented backup. Fines run $43,792 per violation in 2026.
Will Meta and TikTok approve AI-generated creative?
Yes, with disclosure. Both auto-detect via watermarks plus content classifiers. Hiding the AI tag kills ad-account standing.
What is the realistic creative-testing velocity?
An agency-only stack typically caps out at 8-15 creative variants per month. An ad factory removes that cap, and the volume is set after the audit against your media budget. More test velocity is the actual unlock.
14. Five patterns to check in DTC creative ops
These are patterns to check for in your own brand, not measured benchmarks. They track the structural specifics of DTC operations - the cadence of paid creative testing, the founder's personal taste, the platform-specific creative dialects.
Note 1 - the founder's eye is the bottleneck. In many DTC brands the founder personally approves every creative asset. That works while the brand is small; it does not scale. The way through the ceiling is for the founder transferred their eye to a documented Brand DNA file plus a designated creative custodian. Until that transfer happens, the velocity ceiling is whatever the founder can review in a week.
Note 2 - paid social is the harshest QA mechanism. The fastest way to find out a creative is off-brand is to run it as a paid ad and watch the comments. Brands running ad factories with high test velocity surface drift in days; brands running agency stacks at 5-15 assets/month surface drift in months. The lesson is to test more, not test less.
Note 3 - retail and packaging anchor the DNA. Include retail and packaging in the Brand DNA file rather than treating them as separate disciplines. The reason: retail and packaging are physical and harder to drift; using them as the anchor for digital surfaces keeps the digital consistent with the brand the customer sees in their hand.
Note 4 - email creative is dramatically under-resourced. Email often gets a small slice of the creative budget even when it drives a large share of revenue. Ad factories that include email banners + hero images in the standard monthly batch surface this under-investment immediately and rebalance creative spend toward the highest-ROI surface.
Note 5 - the comparable-brand reference is non-optional. Brands that come into the Brand DNA Sprint with 5-10 reference brands they admire produce a sharper DNA file than brands that say "make it look on-brand". The reference brands need not be from the same category - a sustainable cosmetics brand might reference Patagonia, Aesop, and Allbirds despite none of them being in cosmetics. Diverse references produce a more robust DNA file.
The fix in every case is operational discipline applied to creative. Documented Brand DNA, versioned LoRA, automated QA scoring, weekly creative review cadence, designated brand custodian. The end-to-end pattern that works in voice agents (voice-agent failure patterns) and automation (why AI pilots fail to reach production) generalises to creative ops with these specific overlays. Run on your specific brand at kratt ad factory for DTC ecommerce or book a review.
Last updated: 4 May 2026.