Reference-context engagement scorer

Tester access. Enter the shared password to continue to the scorer, model mathematics, demo video, and contact sheets.

Sessions expire after 30 minutes. Password attempts are rate limited.

Tester materials

Mid-training material

Produced partway through training, from an earlier model than the one scoring on this page. Development diagnostics, not finished output.

Model mathematics

The scorer maps frozen DINOv3 patch representations and structured posting context to a scalar estimate of log engagement. The following defines the 512-pixel configuration used by each member of the demonstrated ensemble.

Patch representation

After channel normalization, the image is passed through a frozen DINOv3 encoder. The classification and register tokens are discarded. The scoring head receives only the 1,024 spatial patch embeddings:

X = {DINOv3}_{patch} (I_{norm}) \in ℝ^{B \times 1024 \times 768} .

The cached classification token is not used by the scoring head.

Context representation

Six scalar covariates are log-transformed, standardized, and assigned learned feature-specific values when missing:

s_{log} = \log (1 + \max (s, 0)), s_{z} = \frac{s_{log} - μ_{s}}{σ_{s}}, s_{imp, j} = (1 - m_{j}) s_{z, j} + m_{j} φ_{j} .

The transformed scalars are encoded by a 6–32–32 MLP. Its output is concatenated with learned embeddings for page, gender, carousel position, carousel length, and person count, together with source-quality covariates and missingness indicators. A final affine projection produces the conditioning vector

c = W_{c} [s_{MLP}; e; q; m] + b_{c} \in ℝ^{B \times 128} .

Adaptive layer normalization

Each pooling block contains separate adaptive normalization modules for its query, patch sequence, and feed-forward input. For $Z \in ℝ^{B \times T \times 768}$ , LayerNorm is evaluated independently at every sequence position over the 768 feature coordinates:

LN (Z)_{b, t, j} = \frac{Z_{b, t, j} - μ_{b, t}}{\sqrt{σ_{b, t}^{2} + ε}} .

Context is converted to a featurewise scale and offset and broadcast across sequence positions:

[γ (c), β (c)] = W_{mod} c + b_{mod}, AdaLN (Z, c) = LN (Z) ⊙ (1 + γ (c)) + β (c) .

Each modulation is a 128-to-1,536 affine map. Its parameters are zero-initialized, so AdaLN initially reduces exactly to non-affine LayerNorm.

Singleton-query cross-attention

A learned query $q_{0} \in ℝ^{1 \times 1 \times 768}$ is expanded across the batch. In each of two blocks, 12 heads of width 64 compare this singleton query with all patch positions. For head h,

Q_{h} = \tilde{Q} W_{h}^{Q}, K_{h} = \tilde{X} W_{h}^{K}, V_{h} = \tilde{X} W_{h}^{V}, A_{h} = softmax (\frac{Q_{h} K_{h}^{T}}{\sqrt{64}}), O_{h} = A_{h} V_{h} .

Here $\tilde{Q} = {AdaLN}_{q} (Q, c)$ and $\tilde{X} = {AdaLN}_{kv} (X, c)$ . Softmax is taken over the 1,024 patch positions.

The 12 head outputs are concatenated and affinely projected. Attention-weight dropout is applied during training. The resulting vector is added to the unmodulated residual query:

O_{attn} = Concat (O_{1}, \dots, O_{12}) W^{O} + b^{O}, Q_{ℓ}^{'} = Q_{ℓ - 1} + O_{attn} .

Feed-forward update and block recurrence

The attention residual is independently modulated and passed through a 768–3,072–768 feed-forward network with GELU and hidden-state dropout:

M_{ℓ} = W_{ℓ, 2} Dropout (GELU (W_{ℓ, 1} {AdaLN}_{mlp} (Q_{ℓ}^{'}, c) + b_{ℓ, 1})) + b_{ℓ, 2}, Q_{ℓ} = Q_{ℓ}^{'} + M_{ℓ}, ℓ \in {1, 2} .

Both blocks attend to the original frozen patch matrix X. The second block therefore revisits every patch using the query state produced by the first block.

Scalar prediction

After the second block, the length-one query axis is removed. No additional normalization or activation precedes the image-path output:

r = Q_{2} [:, 0, :] \in ℝ^{B \times 768}, y_{img} = W_{out} r + b_{out}, y_{ctx} = W_{a 2} GELU (W_{a 1} c + b_{a 1}) + b_{a 2}, \hat{y} = y_{img} + y_{ctx}, y = \log (1 + like count) .

The context-only branch contains no image information. Both outputs are unrestricted real-valued scalars.

Reference-context ensemble

At inference, nuisance covariates are fixed to a shared reference context. Three fold checkpoints evaluate the upright and horizontally reflected image. Each of the six scores is standardized against its checkpoint-specific held-out distribution before aggregation:

z_{f, a} = \frac{s_{f, a} - μ_{f}}{σ_{f}}, f \in {2, 3, 4}, a \in {upright, flip}, \bar{z} = \frac{1}{6} \sum_{f}^{} \sum_{a}^{} z_{f, a}, percentile = 100 \cdot {\hat{F}}_{g} (\bar{z}) .

The empirical distribution ${\hat{F}}_{g}$ is the held-out calibration population for gender reference g. The displayed ensemble spread is the range of the six member percentiles; it is not a confidence interval.

Full-data distilled student (optional head)

One image-only head may be selected in place of the ensemble. It receives the same frozen patch matrix X and a learned gender embedding at the teacher's context width, and nothing else: no page, no follower count, no carousel slot. The reference context is not removed but inherited, because the quantity it was trained to predict is a five-fold teacher's reference-context score. This successor was trained on 879,868 images, compared with 63,882 for the earlier student.

Writing $\bar{z}$ for that teacher score, training standardizes it within gender and the student predicts the standardized target under a Gaussian likelihood whose variance floor is the teacher's own inter-checkpoint spread:

t = \frac{\bar{z} - μ_{g}}{σ_{g}}, (μ, \log s) = Student (X, g), v^{2} = e^{2 \log s} + σ_{teacher}^{2}, L = \frac{1}{2} (\log v^{2} + \frac{{(t - μ)}^{2}}{v^{2}}) .

A label the five teacher checkpoints disagree about is a label the student is not penalised for missing. The predicted uncertainty is an internal quantity and is never reported as a statement about the photograph.

At inference the target standardization is inverted, which places the student back on the ensemble's axis, and the same empirical distribution supplies the percentile:

\hat{z} = μ \cdot σ_{g} + μ_{g}, percentile = 100 \cdot {\hat{F}}_{g} (\hat{z}) .

The student has no calibration population of its own, and an estimate is less dispersed than the quantity it estimates, so its percentiles sit closer to the median than the ensemble's for the same photograph. That compression is measured on held-out first-image single-person posts and reported with every student score. It is an experimental comparison, not the deployed model.

Live scorer walkthrough

What this number is

Predicted engagement for this photograph posted solo, as the thumbnail, by a standardized account — with page reach, follower count, posting history, carousel position and source compression all held fixed.

That fixing is the whole point. Left free, a score mostly measures how many followers an account has. Pinning them to the same value for every image means what varies between two scores is the photograph.

It predicts how a population responded to an image. It is not a measure of attractiveness or of a person's worth. And one number is not the whole answer: the six ensemble members typically span 39 percentile points, which is why the band is shown at the same weight as the estimate.

Reference-context engagement scorer

Reference-context engagement scorer

Validated photo-region examples

Score an image

The live scorer is not answering

Result

Scored by the full-data student

Which region drove this score

College context comparison

Tester materials

Mid-training material

Patch representation

Context representation

Adaptive layer normalization

Singleton-query cross-attention

Feed-forward update and block recurrence

Scalar prediction

Reference-context ensemble

Full-data distilled student (optional head)

Live scorer walkthrough

What this number is