Campaigns · Self-initiated concepts

Static ads

Thirty-four ads for Pip, LIDO, Nula and JACI, made from approved stills. Each set starts from a brief, with one idea per frame. In all but one ad, the type is set in code rather than drawn by an image model, and the product photography is placed, not regenerated.

At a glance

Role

Art direction, concepts, copy and final selection. Built the ad renderers and checks with Claude Code and Codex.

Tools

Node.js
SVG
sharp
Python
OpenCV
Playwright

Outcomes

Thirty-four ads across four campaigns, in 4:5, 1:1 and 9:16.
Copy sticks to confirmed product facts. LIDO's "3 for $10" sticker is marked as a placeholder, and JACI's Sale versions are only for use during a real sale.
Masters stay editable as SVG or HTML, and the LIDO notes record a repeat render matching its SVG and PNG checksums exactly.

Links

Pip ad with the model holding PRESS and the headline "Tone. Texture. PRESS." — One ad from each set: Pip, LIDO, Nula and JACI, all at 1080 × 1350.

LIDO ad: "Vanilla. Salt. Sun." beside a raised Salted Vanilla can and the LIDO cap. — One ad from each set: Pip, LIDO, Nula and JACI, all at 1080 × 1350.

How a set is made

01
Brief

Placement, format, the product facts that can be used and the claims that cannot. LIDO and JACI were briefed for the Meta feed at 1080 × 1350.
02
One idea per frame

Each ad gives a different reason to look. The first five LIDO ads test taste, an occasion, the brand voice, a café ticket and the range.
03
Copy against the facts

Headlines are checked against the confirmed product facts before layout. For LIDO and JACI, several lines per ad were compared and the reason for each choice written down.
Human approval
04
Source stills

Chosen from approved finals. No new photography was generated for the LIDO or JACI sets.
05
Render in code

Layouts are SVG or HTML, rendered to exact-size PNGs. Photographs are placed, scaled and cropped, never repainted.
06
Feed-size review

LIDO and JACI were reviewed on contact sheets at feed-preview size. On LIDO this removed small support lines and a receipt footer, and enlarged the range CTA.
Human approval
07
Checks

Pixel sizes read back from the files, type kept clear of faces and packaging, and for Nula, a script that measures contrast on the pixels behind every line of text.
08
Editable masters

SVG or HTML masters are kept beside the PNG exports, so any ad can be corrected and rendered again.

Human approval

Five ideas for LIDO

The LIDO set was planned as five different reasons to care, not five layouts of one idea.

Ad	What it tests	Headline	Built from
BLACK product hero	Taste	Deep cocoa. Clean finish.	The studio pack shot, cut out
Long lunch	The occasion	Let lunch run long.	Lifestyle still, full bleed
Type only	Brand voice	Coffee without the rush.	Type and the three flavour colours
Café ticket	A native social format	Stay for coffee.	A vector ticket, no photography
The range	Choice	Three ways to take it cold.	The three-can family still

The ads set their headlines in capitals. Two more lifestyle ads were added later from the same stills.

LIDO Black ad on cream: "Deep cocoa. Clean finish." with an intro offer sticker.

LIDO ad: "Let lunch run long." above the seaside lunch image.

Type-only LIDO ad on navy: "Coffee without the rush." above a three-colour band.

LIDO ad showing a café ticket on blue that reads "Stay for coffee. No rush."

LIDO ad: "Three ways to take it cold." over the couple and the three cans.

Choosing a headline

Options for the BLACK product hero, scored out of 5 in the copy notes. As the notes put it, the scores are "directional creative judgement, not performance predictions".

Option	Comprehension	Product connection	Pull	Decision
Deep cocoa. Clean finish.	5	5	5	Selected. Exact product language and a two-beat hierarchy.
Cold brew, cut clean.	4	4	4	Strong, but "cut clean" is less natural customer language.
Black. Slow-steeped.	4	5	3	Accurate, but reads like the label and is less distinctive.

The rules each set followed

Pip
No text over packaging
One dominant idea per frame, generous negative space, and only approved product facts: no medical, clinical, review or dermatologist claims. Production labels such as "01 / WASH" never appear on finished ads.
LIDO
Readable at thumbnail size
Headlines of three to seven words at 86–132 px on the 1080 × 1350 canvas, a 64 px safe edge, and a hierarchy that still reads at 324 × 405.
JACI SS26
Few words, all factual
Only the brand and SS26 or Sale. Dense portraits, movement and detail shots sit in a pink frame so the type stays off the photograph; frames with open wall, sky or road run full bleed.
Nula
Quiet in a loud category
Avenir Next for copy and a Futura wordmark that matches the mark on the case lid. Colours sampled from the product, and no shadows on cut-outs, no badges and no buttons.

One photograph, three versions

The same approved JACI frame across three rounds of the ad set.

Version 1
Version 2
Version 3

Version 1 carried a campaign line, an editorial serif and decorative rules. Version 2 removed every rule, panel, caption and micro-label, leaving one image and at most two text elements. Version 3 added SS26 and Sale markers and a pink frame for dense images, across 13 assets.

One system, three formats

The Nula set keeps one palette, one type pairing and one wordmark across story, square and feed placements.

Nula story ad: "All of it. None of the noise." and "Two tones. Four sizes. One case." above the case, two earplugs and four spare tips on a warm grey ground. — Story, 9:16

Nula ad: "Quiet, in your colour." above brown and pale earplugs. — Square, 1:1

Nula ad: "Quiet, to go." above the case on a tan background. — Feed, 4:5

Cutting out a can without repainting it

scripts/extract-lido-black-cutout.py (excerpt)

# The source is an approved exact packshot on a near-white sweep. Give
# GrabCut a narrow probable-foreground region and a conservative sure-
# foreground core, preserving source pixels and changing alpha only.
cv2.rectangle(mask, (245, 165), (780, 1385), cv2.GC_PR_FGD, -1)
cv2.rectangle(mask, (345, 285), (680, 1250), cv2.GC_FGD, -1)
…
cv2.grabCut(image, mask, None, background, foreground, 8, cv2.GC_INIT_WITH_MASK)
…
alpha = cv2.GaussianBlur(silhouette, (0, 0), 0.75)
alpha[alpha < 8] = 0
…
rgba = cv2.cvtColor(image, cv2.COLOR_BGR2BGRA)
rgba[:, :, 3] = alpha

The approved pack shot goes in and only the alpha channel is written, so the can's colour, label and proportions come out exactly as they went in.

Measuring contrast behind the type

tools/qa_ads.py (excerpt)

pg.add_style_tag(content="h1,p,.mark{visibility:hidden!important}")
pg.wait_for_timeout(120)
bg_png=pg.screenshot()
…
for bx in boxes:
    …
    patch=bg[y0:y1,x0:x1].reshape(-1,3)
    …
    c=contrast(tl,pl)
    # worst 2% of the box, so a small bright blowout is not averaged away
    worst=float(np.percentile(c,2)); med=float(np.median(c))
    need=3.0 if bx["size"]>=24 else 4.5

Each Nula ad is rendered once with its text hidden. The pixels behind every line are sampled, and the worst 2% is compared with WCAG contrast thresholds: 3:1 for text of 24 px and above, 4.5:1 below that. Anything under is flagged LOW.

Every set

Campaign	Ads	Formats	Rendered with	Checks
Pip	8	4:5	SVG masters, rendered with sharp	Hierarchy, mobile readability, brand cohesion, unsupported claims and product accuracy
LIDO	7	4:5	SVG masters with embedded fonts, rendered with sharp	Exact dimensions, a 288 × 360 contact sheet, text clear of labels, matching checksums on a repeat render
Nula	11	4:5, 1:1, 9:16	HTML rendered in Chromium with Playwright; one Notes-style ad generated with GPT Image	A contrast script run against the pixels behind each line of text
JACI SS26	8	4:5	SVG masters, rendered with sharp	PNG dimensions, a contact sheet of every asset, headlines clear of faces and garment details

JACI's third version rendered 13 assets; eight made the final set.

Static ads

At a glance

Role

Tools

Outcomes

Links

How a set is made

Brief

One idea per frame

Copy against the facts

Source stills

Render in code

Feed-size review

Checks

Editable masters

Five ideas for LIDO

Choosing a headline

The rules each set followed

No text over packaging

Readable at thumbnail size

Few words, all factual

Quiet in a loud category

One photograph, three versions

One system, three formats

Cutting out a can without repainting it

Measuring contrast behind the type

Every set