EXTENDED / SEMINAR DECK (~12–15 min): full version incl. related work + research model + future work. The lean 8-min flash version is flash_talk.md. Presenter notes are the HTML comments (Marp presenter view = press `p`).
SPEAKER (~20s): Influencer marketing has a tension — brands want control of their look, creators stay credible by keeping THEIR look. I built an automatic, image-based measure of how well those two looks match: visual-style congruence. Quick tour of the method and that it works.
SPEAKER (~45s): Creators build a recognizable visual signature — it's a trust heuristic. Sponsorship pushes the brand's aesthetic onto them, which can erode it. The construct I measure is visual-STYLE congruence: the aesthetic alignment between brand and creator. Today: the measure + evidence it works. (The effect on engagement/conversion is the next step, not today.)
SPEAKER (~50s): Crucial distinction. Style is what's left when you remove the objects (Gombrich) — here, the same apple across centuries, content fixed, style changing. We measure HOW an image looks, independent of WHAT it shows. The closest prior paper (Argyris 2020) also says "visual congruence" but classifies content/themes; we measure style. That's the wedge. TECH: be ready for "isn't this Argyris?" — theirs = supervised theme classification between influencer & followers; ours = unsupervised style embedding, brand→influencer.
SPEAKER (~70s): The engine. Push each frame through a frozen ImageNet CNN — no training. At several layers we take the GRAM MATRIX: the correlation between feature channels, summed over every spatial position. That summation is the key trick — it throws away WHERE features occur and keeps how they co-occur. Mathematically that's why it captures texture and colour, i.e. style, and not the objects. This is the style representation from neural style transfer (Gatys 2016). Each image becomes a style vector. TECH: 5 layers block1–block5_conv1 (low→high texture scales); per-layer upper-triangle, mean- centre/standardise/L2-norm; ~305k dims. "Just colour?" No — multi-scale texture co-occurrence; colour is one component.
SPEAKER (~45s): We average a creator's frames into a style SIGNATURE, and do the same for each brand. Congruence is just the cosine similarity between two signatures — and we mean-centre across entities so it reflects DISTINCTIVE style, not the average look everyone shares. That's the whole measure. Now: does it actually work? TECH: cosine is scale-invariant direction in style space (vectors L2-normed); mean-centring removes the common component → distinctiveness.
SPEAKER (~45s): First validation. Pairwise style similarity across 18 brands. With no labels, interpretable clusters fall out — fast food together, athletic/outdoor together, and beauty sitting opposite masculine/outdoor. The measure recovers structure a human would agree with. TECH: mean-centred cosine of centroids; colour auto-scaled; clusters are a sanity check, not a fitted result.
SPEAKER (~45s): Now match each creator to the brand whose style they're closest to. Cyan = best fit. It's not collapsing onto one brand, and the top hits are intuitive — which we'll see on real frames next. TECH: cell = mean-centred cosine(creator, brand); descriptive validation, not a causal claim.
SPEAKER (~30s): The cleanest example. Top row: the creator's frames most like the brand; bottom: the brand's frames most like the creator. Both alpine — but remember, the model never detects "mountain"; it's matching colour, light, and composition statistics.
SPEAKER (~30s): Another — a ski/outdoor creator and Red Bull, our strongest match. Same big-sky, snow, earthy palette on both sides. (Beauty and food creators match Maybelline and Chipotle too.)
SPEAKER (~40s): The check that matters most. For one creator, similarity to all 18 brands: top-3 are all sport/outdoor, bottom-3 are all beauty — and negative. So this isn't a measure where everyone is close to everyone; it's specific and directional. That's the validation I'd hang the talk on. TECH: negatives arise from mean-centring (below-average style); rules out "high cosine everywhere".
SPEAKER (~40s): Where this sits. Style is computable — neural style transfer and stylometry. Marketing has lots of work on image CONTENT, almost none on style. The closest prior "visual congruence" (Argyris 2020) is content/theme overlap between an influencer and their followers. The match-up hypothesis (Kamins 1990) says endorser–brand fit matters — we give it a visual, continuous measure. Arsel et al. 2025 in the Journal of Marketing theorizes the brand–creator authenticity tension in sponsored content; Mummalaneni et al. 2026 in Marketing Science studies how creators respond to engagement — both connect to our forthcoming engagement model. Our gap: STYLE congruence, brand–influencer, nonlinear. TECH: all citations verified — three in the lit/ folder (Argyris 2020 CHB; Arsel et al. 2025 J. Marketing 89(6):57–76; Mummalaneni et al. 2026 Marketing Science); rest from the ISMS deck + Kamins 1990 (J. Advertising 19(1):4–13).
SPEAKER (~35s): So what. We now have a scalable, content-agnostic measure of brand–influencer style fit — different from content-based congruence and from survey measures of fit. The natural next step, with data in hand, is to relate congruence to engagement and conversion on sponsored posts, and to help firms select influencers by style fit rather than follower count. TECH (if asked): identification via within-creator organic-vs-sponsored variation, creator/brand fixed effects, controls. Also extensible from frames to VIDEO (editing pace, motion, colour rhythm).
SPEAKER (~35s): With the measure in hand plus engagement/conversion data, the marketing questions. Does style congruence move engagement and conversion on sponsored posts — and is there an AUTHENTICITY BOUNDARY, a threshold where imposing too much brand look backfires on conversion? How does that boundary shift by brand status/price, product type, and brand–creator fit? The payoff is a rule for when to enforce brand aesthetics vs preserve authenticity. TECH: curve is hypothesised (to be estimated); identification via within-creator organic-vs-sponsored variation + creator/brand fixed effects + controls.
SPEAKER (~30s): And the authenticity boundary isn't fixed. For luxury / high-status brands the tolerance is narrow — high congruence is essential; for mass-market it relaxes. It also shifts with price, hedonic vs utilitarian products, and brand–creator fit. TECH: operationalised as interaction terms (congruence × moderator); hypothesised, to be estimated.
SPEAKER (~30s): Second direction — extend from frames to video. Right now a video is just the average of its frames, so we keep the look but lose the temporal style: pacing, motion, rhythm. We'd add interpretable temporal features for a per-video style vector — which also aligns the measure with the unit our outcomes live on. TECH: cut rate via shot-boundary detection (PySceneDetect); motion energy via optical flow; style drift = spread of frame style vectors within a clip. Optional: gram-pool a spatiotemporal backbone (VideoMAE/SlowFast) — but those are action-recognition pretrained → more content leakage.
SPEAKER (~20s): To close: visual style is measurable, brand-specific, and matchable at scale — an objective handle on the brand-control-vs-authenticity trade-off. Thank you — happy to take questions. LEAVE ~2 MIN FOR Q&A. Top likely questions: just colour? / content leakage? / why beauty-heavy? / does it affect sales? — answers in presentation.md backup notes.