Voyager Ignacio’s research

Not all layers
are equal.

Learning to select layers for reliable out-of-distribution detection.

Voyager learns which frozen CLIP layers to trust for each image, instead of using one fixed fusion rule.

Read the paper Explore examples
Ignacio M. De la Jara1,3Cristian Rodriguez-Opazo2Damith Ranasinghe1,3

1 Adelaide University · 2 Australian National University
3 Crossing Lab

Recorded flowers example and selected CLIP layers Router 1 of 5 selects layers 1, 2, 7, 8 and 12. Positions follow layer order; all CLIP layers are computed. Flowers · iNaturalist 1 2 3 4 5 6 7 8 9 10 11 12 Shallow to deep · layer 12 always included
Catch a missed shift

Router 1 of 5 selects layers 1, 2, 7, 8, 12.

Ensemble decision: OOD-like

Reference: OOD

Choose a recorded example

Recorded with frozen CLIP ViT-B/16. All layers run; highlighted scores form one expert. The ensemble averages five expert scores. Inspect the full evidence ↗

18.86%Average FPR@95 ↓ lower is better

94.49%Average AUROC ↑ higher is better

8.8 pointslower average FPR@95
than four-shot Local-Prompt (ID-derived context)

ImageNet-1K
CLIP ViT-B/16 · 4-shot ID + WebContEX

One fixed route
misses evidence.

The final representation is only part of the story.

Intermediate CLIP layers retain different visual cues. The useful combination changes with the input and the distribution shift, so averaging the same layers for every image can leave useful evidence untapped.

Voyager keeps the representations and confidence scores frozen. It learns a smaller decision: which sparse layer expert should supply this image’s score?

It builds on Mysteries of the Deep, moving from fixed intermediate-layer fusion to learned, input-conditioned selection.

Textured leaves, a DTD example from the paper.
Textures · DTD
A flowering plant, an iNaturalist example from the paper.
Nature · iNaturalist
A coastal landscape, a Places example from the paper.
Places · Places365
A wooded lake, a SUN example from the paper.
Scenes · SUN

Four OOD shifts from the study. Images are excerpts from the paper’s recorded routing examples.

Same model.
A different route.

Each expert averages a subset of layer scores. Every candidate includes the final layer as a semantic anchor; the router selects one expert per image.

Follow the computation

Select a component, flame, snowflake or distribution to see what it does.

Pan the diagram horizontally, or choose a part below.

Inference and layer selection with Voyager Expert creation How Voyager learns Penguin · ID Puffin · OOD ID label prompts Text encoder T(t₁) T(t₂) T(tₙ) Visual encoder Layer 1 Layer 2 Layer N ⋮ OOD scores S₁ S₂ Sₙ ⋮ Voyager Inside Voyager · enlarged view E₁ E₂ E₃ Eₖ … Router Gθ Mean layer features 🚀 Voyager ID Penguin OOD Puffin S₁ Sₙ S₂ S₃ Sₙ ⋮ Scores Experts S₁ S₂ S₃ Sₙ E₁ E₂ E₃ Eₖ ⋮ ⋮ + + + Fixed subset means Final layer included Expert distributions Schematic scores OOD context ID images CLIP OOD method Voyager ID / OOD score distributions OOD ID ID boundary L = CE(q, Gθ) + 0.01 Lbal Soft top-32 expert targets

Part 1 of 9 · Inference

An image enters a fixed detector

Penguin and puffin illustrate ID and OOD inputs in the figure. At inference, Voyager receives an image without an ID/OOD label. Its route is predicted from frozen visual features.

Input image → frozen CLIP features

The birds and flows are a schematic, not an inference result.

Inspect real model traces ↗
Interactive reproduction of the paper architecture. Snowflakes indicate frozen weights; flames identify router learning. Moving curves and signals are conceptual illustrations, not measurements.
Open the original figure ↗

Same backbone. Different wiring.

Change the image. See which connections can change.

Teaching example

The penguin is the ID example in the paper illustration.

Final layer only

One terminal score

Penguin input Same frozen CLIP Backbone Scores Layer 1 S1 Layer 2 S2 Layer 3 S3 Layer 4 S4 Layer 5 S5 Layer 6 S6 Layer 7 S7 Layer 8 S8 Layer 9 S9 Layer 10 S10 Layer 11 S11 Layer 12 S12 No fusion One score is read Read S12 OOD score Layer 12 only

Read the final layer’s score directly. Intermediate layer scores do not enter the output.

Layer 12.

Fixed fusion

One shared recipe

Penguin input Same frozen CLIP Backbone Scores Layer 1 S1 Layer 2 S2 Layer 3 S3 Layer 4 S4 Layer 5 S5 Layer 6 S6 Layer 7 S7 Layer 8 S8 Layer 9 S9 Layer 10 S10 Layer 11 S11 Layer 12 S12 One fixed subset Same for both birds Fixed mean OOD score The same five layers

Average the same chosen subset for every image. Here the five-layer subset stays fixed when you change the bird.

Layers 3, 4, 6, 8, 12.

Voyager routing

An image-conditioned recipe

Penguin input Same frozen CLIP Backbone Scores Layer 1 S1 Layer 2 S2 Layer 3 S3 Layer 4 S4 Layer 5 S5 Layer 6 S6 Layer 7 S7 Layer 8 S8 Layer 9 S9 Layer 10 S10 Layer 11 S11 Layer 12 S12 Mean features Router Gθ Choose Expert mean OOD score A route for this image

Mean-pooled frozen features enter the router. It chooses the expert whose selected layer scores are averaged.

Layers 3, 5, 6, 9, 12.

Penguin: final-layer scoring still reads layer 12, fixed fusion still combines 3, 4, 6, 8 and 12, and the illustrative Voyager route uses 3, 5, 6, 9 and 12.

Solid arrows carry features or scores. The dashed arrow is the router’s subset choice. One router is drawn; the deployed detector averages five.

These routes illustrate the three structures; they are not model predictions. Inspect recorded layer evidence ↗

WebContEX: retrieval-based augmentation

Add weak OOD context to router training using the ID vocabulary.

Construction in the paper ↗

The images are retrieved from Wikimedia Commons after semantic filtering. This expands the router’s training context; CLIP stays frozen and the context is used only during training.

penguin other ID names ID labels candidate nouns WordNet filter Retrieve & clean CLIP features layer scores fixed cache Freeze & cache ID + context Router Gθ Train router

Step 1 of 5

Start from the ID class names

The ID label set defines which concepts belong to the task. WebContEX uses those names to establish a semantic exclusion set, without needing target-OOD examples.

Few-shot ID examples have domain label y = 0; retrieved context images later receive y = 1.

The deployed filter checks the full ID vocabulary, excludes ID synsets and their ancestors/descendants, and requires at least two WordNet hops. It also applies vocabulary depth, lexical-category and lemma-frequency rules. Semantic filtering does not guarantee image-level disjointness from every evaluation dataset.

Retrieved context from the paper

Examples from Figure C.1, used for router supervision.

A floor-cleaning device in a WebContEX retrieval.A bell in a WebContEX retrieval.A rocky shoreline in a WebContEX retrieval.A woven hut in a WebContEX retrieval.
What the deployed construction checks

Balanced retrieval. Up to five query variants per synset; target 20 images per synset, with per-query and aspect-ratio caps.

Image quality. At least 256 × 256 pixels; valid decoding, orientation normalization and photographic-quality checks.

Within-shard duplicates. URL/page-ID checks, exact JPEG SHA-256 checks and an 8 × 8 perceptual hash with a Hamming-distance cutoff of 8.

Training-only cache. The deployed pool has 5,078 entries. Frozen features and layer scores supervise soft expert targets; query names are not classification targets, and context images are not required at inference.

Frozen evidence

CLIP image features, text prototypes, and layer-wise OOD scores stay unchanged. An expert is an unweighted mean over a final-layer-anchored subset.

Lightweight learning

A small MLP router learns from few-shot in-distribution examples and WebContEX: weak web-retrieved context constructed from the ID label set. Training uses no target-OOD images or labels. The deployed context comes from Wikimedia Commons; concepts are filtered to be at least two WordNet hops from the ID taxonomy. This semantic filter is not a guarantee of complete image-level disjointness; the paper details the construction and deduplication scope.

Simple inference

Each router chooses one expert. The deployed five-router ensemble averages their outputs while sharing one frozen CLIP forward pass. WebContEX is needed only during training. The paper reports 4.18 ms per image for the deployed model versus 4.08 ms for fixed 12-layer GL-MCM and 3.21 ms for MCM on its measured single-GPU setup; these are hardware-specific timings.

Results, in context.

ImageNet-1K as the in-distribution dataset, CLIP ViT-B/16, and four OOD shifts. FPR@95 measures the OOD fraction accepted when ID recall is 95%.

FPR@95 (%) ↓ · lower is better

Four-shot results; MCM, GL-MCM and MoD are training-free references.
Method iNaturalist SUN Places DTD Average
MCM zero-shot 30.91 37.59 44.69 57.77 42.74
GL-MCM zero-shot 17.42 30.75 37.62 55.20 35.25
MoD fixed fusion 15.98 45.58 35.71 25.51 30.70
CoOp 14.60 28.48 36.49 43.13 30.67
LoCoOp 16.05 23.44 32.87 42.28 28.66
ID-Like 8.98 42.03 44.00 25.27 30.07
Local-Prompt 9.65 20.40 29.39 51.20 27.66
Voyager 3.93 21.65 21.88 27.96 18.86

Bold marks the lowest FPR@95 among methods with the selected shot count. Training-free references are shown for context. Baselines follow the paper’s reported comparisons; one-shot Local-Prompt is the authors’ rerun. Voyager additionally uses WebContEX, while Local-Prompt uses ID-derived local context. Voyager has the lowest average here; Local-Prompt is lower on SUN and ID-Like on DTD.

The gain comes from both the training context and the routing decision.

With the router recipe held fixed, replacing WebContEX with ID-derived context increases average FPR@95 from 18.86 to 26.71. With WebContEX and expert scores fixed, learned routing improves on the strongest score-based selector by 10.1 points.

These controls measure different contributions. The headline comparison is to a four-shot baseline; Voyager additionally uses weak web-retrieved context.

Look inside the routes.

Explore the paper’s diagnostics. Layer use and routed examples help explain the mechanism; they do not certify that every OOD input is detected.

Paper figure showing Voyager’s distribution of selected layers.
Layer-selection distributions from the paper. These summarize routed behavior; the schematic above uses illustrative selections.

Where the evidence stops.

Voyager requires a few-shot ID set and a weak web-retrieved training context. Results vary across datasets and backbones; the lowest average does not mean the lowest FPR on every shift.

The wider study includes Pascal-VOC, ImageNet-100, near-OOD subsets, and additional backbones. See the paper for their protocols, full comparisons, training costs, and limitations.

Take a closer look.

Not All Layers Are Created Equal: Learning to Select Layers for Reliable Out-of-Distribution Detection

The paper, implementation, and the work that motivated input-conditioned layer selection.

Read the paper Explore recorded Voyager examples

Code: DynAMO repository · public release pending

Read Mysteries of the Deep

Camera-ready paper, updated 2 November 2026. Voyager is implemented in DynAMO. The repository is currently private; a public code release is pending. MIT code licence · Contact.

Cite this work

@misc{delajara2027voyager,
  title = {Not All Layers Are Created Equal:
    Learning to Select Layers for Reliable
    Out-of-Distribution Detection},
  author = {De la Jara, Ignacio M. and
    Rodriguez-Opazo, Cristian and
    Ranasinghe, Damith},
  year = {2027},
  note = {WACV 2027, camera-ready paper},
  url = {https://mezosky.github.io/voyager/}
}

Camera-ready citation, updated 2 November 2026. Final proceedings metadata will be added when available.

Two birds. Different destinations.

The penguin stays in familiar territory. The puffin heads beyond it.

In-distribution Out-of-distribution Beyond the ID vocabulary ID OOD

Penguin · ID. Puffin · OOD.

A playful illustration of the paper’s penguin/puffin example.