Natural photographs contain dense, overlapping structures that can be weakly expressed in feature maps from whole-image vision encoders. We introduce forced focused attention (FFA) and its Python implementation, vividRGB. FFA is an encoder-agnostic inference-time procedure that applies a frozen encoder to enlarged, overlapping local crops, averages inverse-aligned reflections, and reconstructs a feature field in source coordinates. We evaluate 50 high-resolution photographs from grassland vegetation, fungal networks, airborne forests, coral reefs, and phytoplankton microscopy, using DINOv3 ViT-B/16. Across the five domains, FFA at 5% or 10% focus produces higher mean RGB-edge boundary F1 than direct DINOv3 encoding and automatic SAM3 masks. Domain means range from 0.244 to 0.596 for FFA, versus 0.028-0.076 for direct DINOv3 and 0.020-0.090 for SAM3. Qualitative feature maps also show finer appearance structure. These results concern RGB-edge preservation rather than verified species, traits, or biological segmentation. Narrower focus generally improves boundary preservation, while extraction time and GPU energy use increase sharply and gains diminish at finer settings. FFA may be useful for relatively small, high-resolution ecological image collections where dense visual cues matter. The alpha release vividRGB 0.1.0 and versioned research resources support reproducible cue inspection.
Figures & tables
Figure 1: Photographic examples across ecological scales. (A) Vegetation community in a project photograph, also used in the first row of Figures 2 and 3 ; (B) coral-reef scenery (central crop; source: [ 8 ] ); (C) fungal hyphae network (source: [ 9 ] ); (D) airborne forest imagery (source: [ 10 ] ); and (E) aquatic phytoplankton microscopy (source: [ 11 ] ). Panel sizes do not represent a common physical scale.
Domain
Source and content
Working dimensions (pixels)
Grassland
MOIN/KICS-Zert; mixed vegetation at four German sites
1312×1216 to 2048×2048 , including nonsquare windows
Coral reefs
Coralscapes; benthic community frames [ 8 ]
2048×1024
Phytoplankton
PMID2019; laboratory microscopy [ 11 ]
1360×1024 or 2048×2048
Table 1: The selected five-domain image collection. Dimensions are those actually supplied to this experiment, expressed as width × height; they are not nominal dimensions of every source image. Each domain contributes n=10 .
Figure 2: RGB inputs and PCA–varimax false-color displays for five illustrative images. The first row uses the same square crop of the vegetation photograph in Figure 1 A (photo 1781861008867 ) in every column, as an additional illustration outside the 50-image benchmark; the other four rows retain benchmark examples. Columns show RGB, DINOv3 original, DINOv3-FFA5%, and DINOv3-FFA10%. Each image has a centered PCA16 basis and orthogonal 16-dimensional varimax rotation shared across its DINOv3 treatments; the first three rotated axes supply display colors. The vegetation row pools equal score samples from the three displayed treatments for its shared 2nd–98th percentile limits. The other rows retain archived limits that also include FFA20%. Colors are feature coordinates, not biological classes. Other image sources: fungal Zenodo 5725751 [ 9 ] ; NEON/NeonTreeEvaluation [ 10 ] ; Coralscapes [ 8 ] ; PMID2019 [ 11 ] .
Figure 3: RGB and appearance partitions for five illustrative images. The vegetation row uses the same square crop of the photograph in Figure 1 A (photo 1781861008867 ) in every column, as an additional illustration outside the 50-image benchmark; the other four rows retain benchmark examples. Columns show RGB, DINOv3 original, automatic SAM3, DINOv3-FFA5%, and DINOv3-FFA10%. The three DINOv3 treatments use independent full-covariance GMMs on their image-specific 16-dimensional PCA fields; labels show the selected tested component counts under the λ=4 penalty. Automatic SAM3 retains its accepted masks and uncovered background. Colors are arbitrary partition labels and do not match across methods. Appearance partitions can reflect substrate, illumination, and optical artifacts; biological identities are not inferred. Other image sources: fungal Zenodo 5725751 [ 9 ] ; NEON/NeonTreeEvaluation [ 10 ] ; Coralscapes [ 8 ] ; PMID2019 [ 11 ] .
Domain
DINOv3 original
SAM3 automatic
DINOv3-FFA 5%
DINOv3-FFA 10%
Grassland (MOIN)
0.061 ± 0.026
0.020 ± 0.016
0.596 ± 0.106
0.441 ± 0.083
Fungal networks
0.076 ± 0.057
0.034 ± 0.027
0.290 ± 0.108
0.244 ± 0.093
Airborne forest (NEON)
0.034 ± 0.011
0.041 ± 0.044
0.549 ± 0.136
0.360 ± 0.135
Coral reefs (Coralscapes)
0.041 ± 0.010
0.036 ± 0.024
0.436 ± 0.136
0.266 ± 0.074
Phytoplankton (PMID2019)
0.028 ± 0.007
0.090 ± 0.032
0.456 ± 0.118
0.288 ± 0.138
Table 2: Primary RGB-edge boundary F1 (mean ± sample SD; ten images per domain). The highest mean in each domain is bold. DINO-derived partitions use independent K=6 K-means. SAM3 uses automatic masks with highest-predicted-IoU overlap priority.
Figure 4: Primary RGB-edge boundary F1 across crop sizes of f . The left panel shows reused automatic SAM3 and DINOv3 original. Black points are equal-five-domain means with bootstrap 95% intervals. The right panel shows DINOv3-FFA at different focus level of f , with the same F1 scale. 100% indicates almost equivalent to the whole image analysis as with DINOv3 original. Complete cohorts contain ten images per domain, 50 overall; curves end at 2%.
Method
Duration (s)
Board energy (kJ)
Efficiency ( 10−3J−1 )
SAM3 automatic
5.0 ± 0.3
0.32 ± 0.02
3.104 ± 0.181
DINOv3 original
0.317 ± 0.165
0.0210 ± 0.0109
70.072 ± 46.793
DINOv3-FFA100%
0.328 ± 0.060
0.0093 ± 0.0017
111.835 ± 26.628
DINOv3-FFA40%
3.1 ± 0.7
0.15 ± 0.05
7.053 ± 1.803
DINOv3-FFA20%
11.4 ± 2.8
0.72 ± 0.21
1.480 ± 0.318
DINOv3-FFA10%
43.8 ± 11.9
2.96 ± 0.83
0.3585 ± 0.0764
Table 3: Single-run implementation cost for the newly measured DINOv3 original control and complete DINOv3-FFA cohorts at 100%, 40%, 20%, 10%, 5%, 4%, 3%, and 2%, plus reused automatic SAM3 (mean ± sample SD; 50 images each). DINOv3 original/DINOv3-FFA costs include encoding and both PCA reconstructions; SAM3 costs cover its automatic generation/filtering/NMS pipeline. Calibration and scoring are excluded, and stages differ. Energy is GPU board energy, not wall-plug energy; efficiency is the mean of per-image 1/J in 10−3J−1 . Incomplete finer-focus results are omitted.
Figure 5: Performance–computation trade-off across crop-sizes, also showing DINOv3 original and automatic SAM3. (A) Primary RGB-edge boundary F1 versus extraction duration. (B) F1 versus GPU board-energy efficiency, 1/J . Horizontal axes are logarithmic; higher F1 with shorter duration or greater efficiency is preferable. Faint points show individual images, colored markers are domain means ( n=10 ), and black markers are equal-five-domain means ( n=50 ). Circles and percentage labels mark DINOv3-FFA; diamonds mark DINOv3 original and triangles SAM3.
Figure 6: Marginal RGB-edge F1 gain per additional 1000 J (left) and joules per +0.01 absolute F1 (right), calculated from paired differences with equal weight per dataset. Both horizontal axes show focus f on a descending logarithmic scale; the left vertical axis is linear and the right is logarithmic. Each shaded rectangle spans adjacent tested focus settings horizontally and the stratified paired-image bootstrap 95% interval vertically. Black points show observed step estimates at geometric midpoints for display; whiskers show the same intervals. Falling gain per joule and rising cost per gain describe diminishing energy efficiency, which is distinct from a zero-gain performance plateau.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Experiment or control
Recorded settings
Primary DINOv3 comparison
Native single-orientation control; FFA 5% and 10%; 384-pixel input, 75% overlap, four reflections; six-treatment shared PCA16; independent K=6 (12 sensitivity).
Reconstruction
Crop minimum 16; bilinear placement; per-axis Hann floor 0.05; positive coverage normalization; no token L2 normalization or feature whitening.
Automatic SAM3
1008-pixel input; internal 32×32 point grid; batch 4; no crop layers; IoU 0.88, stability 0.95, NMS 0.7; highest-IoU overlap priority.
Long-side-256 grids; PCA d=3,16 ; K-means K=3 –12; full-covariance GMM K=1 –12; post hoc λ=4 ; sensitivity λ=1 –64.
Crop-size analysis
Separate calibration; eight complete 50-image cohorts, 100%–2%; four reflections; two reconstruction branches in measured DINOv3 cost.
Appendix
Table S1: Essential controls by experiment. Encoder architecture is fixed by the checkpoint; crop geometry, calibration, partitioning, and evaluation are experimental choices. Selected methodological details are given in the surrounding supplementary sections.
Domain
Focus
Baseline
Δ
95% interval
Wins
Raw p
Holm p
Grassland
5%
DINOv3
0.535
[0.460, 0.604]
10/10
0.00195
0.03906
Grassland
5%
SAM3
0.575
[0.505, 0.642]
10/10
0.00195
0.03906
Grassland
10%
DINOv3
0.380
[0.326, 0.434]
10/10
0.00195
0.03906
Grassland
10%
SAM3
0.421
[0.367, 0.476]
10/10
0.00195
0.03906
Fungi
5%
DINOv3
0.214
[0.144, 0.289]
10/10
0.00195
0.03906
Fungi
5%
SAM3
0.257
[0.181, 0.327]
10/10
0.00195
0.03906
Appendix
Table S2: Complete paired differences in RGB-edge boundary F1: FFA minus baseline, n=10 per domain. DINOv3 denotes the native original control; SAM3 denotes automatic masks. Wins are strictly positive differences. Intervals are descriptive and unadjusted. Values are retained from the frozen per-image results.
DINOv3-FFA5%
DINOv3-FFA10%
λ
Mean K
Median
K=4 –6
Cap
Mean K
Median
K=4 –6
Cap
1
11.14
11
0%
48%
11.90
12
0%
96%
2
8.10
8
14%
8%
10.26
11
4%
36%
4
5.34
5
74%
0%
6.34
6
64%
0%
8
3.38
3
34%
0%
4.04
4
56%
0%
16
2.38
2
8%
0%
2.64
3
10%
0%
Appendix
Table S3: Effect of the strengthened BIC-style penalty at 16 dimensions. Frequencies are empirical percentages over the same 50 images; “cap” is selection of the maximum candidate K=12 . The retained exploratory setting is λ=4 .
d
Features
Mean K
Median
K=4 –6
K=12
3
DINOv3 original
10.02
11
10%
42%
3
DINOv3-FFA5%
5.60
5.5
60%
0%
3
DINOv3-FFA10%
6.58
6
46%
4%
16
DINOv3 original
11.90
12
0%
92%
16
DINOv3-FFA5%
5.34
5
74%
0%
16
DINOv3-FFA10%
6.34
6
64%
0%
Appendix
Table S4: Full-covariance GMM selection at λ=4 with 3 versus 16 retained PCA dimensions. Each row contains 50 images. Criterion values are not compared across dimensions.
Range
Coordinate
Knot
Slope change
Frequency
100%–2%
log10 mean energy
20%
Acceleration
100.0%
100%–2%
log10 inverse focus
20%
Acceleration
100.0%
100%–2%
Linear mean energy
5%
Deceleration
62.3%
40%–2%
log10 mean energy
4%
Deceleration
99.3%
20%–2%
log10 mean energy
4%
Deceleration
93.0%
Appendix
Table S5: Hinge sensitivity to coordinate and included focus range. Frequency is bootstrap reselection of the observed knot, not the probability of a biological optimum.
DINOv3
SAM3 image encoder
Domain
Original
FFA10%
FFA5%
Original
FFA10%
FFA5%
Grassland
0.0915
0.4837
0.5787
0.2487
0.4733
0.4799
Fungal networks
0.0314
0.2172
0.3346
0.1258
0.3040
0.3439
Airborne forest
0.0317
0.3979
0.6096
0.2997
0.5789
0.6062
Coral reefs
0.0511
0.2172
0.4144
0.2512
0.4522
0.4923
Phytoplankton
0.0240
0.0654
0.1637
0.1316
0.4004
0.4496
Appendix
Table S6: Primary RGB-edge boundary F1 for the five matched illustrations ( K=6 , 16 PCA coordinates). Each domain entry is one image; the final row is its descriptive five-example mean. SAM3 columns are image-encoder partitions, not automatic masks.
Method
Precision
Recall
F1
Density
3-axis F1
DINOv3 original
0.4853
0.0244
0.0459
0.0040
0.1792
DINOv3-FFA10%
0.4368
0.2184
0.2763
0.0420
0.2864
DINOv3-FFA5%
0.4764
0.3984
0.4202
0.0761
0.3843
SAM3 image encoder
0.3403
0.1559
0.2114
0.0445
0.2654
SAM3-FFA10%
0.3083
0.8012
0.4417
0.4064
0.5317
SAM3-FFA5%
0.3227
0.9230
0.4744
0.5895
0.5284
Appendix
Table S7: Descriptive means across the same five cases. Precision, recall, F1, and boundary density use the primary 16-coordinate partition. The last column uses a separately fitted K=6 partition of the first three varimax axes. Mean F1 is the mean of image-level F1, not the harmonic mean of the displayed mean precision and recall. Density is the fraction of eligible image pixels marked as predicted boundaries.
Localized image edits can change a photograph's meaning while leaving most of it authentic, so forensic analysis must identify where an edit occurred. We show that patch-level perturbation responses from frozen DINO encoders are themselves localization maps. Training-free Localization of AI-image Edits from patch-token Drift (TRAIL) applies one global Haar perturbation and maps cosine drift between corresponding patch tokens. On 80 source-disjoint CocoGlide test images, TRAIL reaches .903 patch AUROC versus .912 for the mask-supervised Detective SAM; fixed-threshold Dice is .619 versus .709, while an oracle threshold raises TRAIL to .790. Transferred unchanged to Poisson image interpolation, TRAIL reaches .855 AUROC versus .864, showing that the cue persists without a generator. Across sixteen DINO encoders, the best block lies at normalized depth .80-.94. Global context matters: AUROC falls from .903 globally to .857 for local-in-canvas perturbations and .735 for independently encoded crops. Frozen DINO patch tokens therefore contain a strong late-layer localization signal whose visibility depends on the perturbation and preserved context. Code: https://github.com/VishalJ99/trail-image-edit-localization.
Zane Kumar, Vishal Jain, Bernhard Kainz
Department of Computing St Paul’s School London London, United Kingdom · Department of Computing Imperial College London London, United Kingdom · Computing / AIBE Imperial College London / FAU London, UK / Erlangen, Germany
Brain encoder models predict cortical fMRI responses from the internal activations of pretrained vision and language networks, and are typically evaluated by held-out prediction accuracy. This is a useful signal for training but a poor one for interpretation: it tells us an encoder fits the data without telling us whether it has internalized the functional organization of the brain. We propose feature visualization -- gradient ascent on the encoder's predicted activation for a target region of interest (ROI) -- as a complementary interpretability technique, and apply it to TRIBE v2 composed with V-JEPA 2 (ViT-G, 40 layers), holding both frozen and synthesizing still images for seven regions spanning the ventral and dorsal visual hierarchies. Under identical hyperparameters, the probe recovers a visible progression of increasing spatial scale and feature complexity across V1 to V4, matching the ventral-stream hierarchy. It also produces three distinctive downstream regimes: radial "frozen-motion" streaks for the middle temporal area (MT) despite static-only optimization, face-like features for the fusiform face area (FFA), and consistent rectilinear line patterns for the parahippocampal place area (PPA). Optimized FFA stimuli drive the predicted region ~4x as much as a natural face photograph, consistent with feature visualization producing adversarial super-stimuli rather than canonical exemplars. The probe is simple, differentiable, and applicable to any brain encoder with a differentiable backbone, allowing for qualitative evaluation of brain encoders.
Production vision pipelines silently degrade on blurry input, wasting compute on downstream OCR, retrieval, and vision-language model (VLM) calls that cannot recover a usable output. We present MagikaDocumentFromPixel, a lightweight, CPU-friendly image quality gate that classifies a single image as sharp, blurred, or uncertain in roughly 7 ms on a single CPU core. The contributions are (i) a recipe selected from a 46-configuration, 8-sweep empirical search that isolates input resolution as the dominant lever and shows architecture capacity only pays off at >= 384 px; (ii) a confidence-aware routing formalism grounded in classical selective prediction; (iii) the Edge Prior Module (EPM), a Laplacian-magnitude auxiliary input channel that gives the network direct access to the spectral evidence that classical blur heuristics rely on and that lifts test F1 by +1.3 points in a matched-env comparison; and (iv) an observation that the gate is one instance of a recurring design pattern that appears independently in Magika content-type detection, risk-controlled OCR with VLMs, and DocVLM. The final recipe MobileNetV3-Large with the EPM trained at 384x384 on paired GoPro Large frames, evaluated with 5-scale test-time augmentation reaches F1 = 0.9803 (AUC 0.9989) with a 17 MB ONNX artifact, improving over our fixed-scale baseline on the same hardware (F1 = 0.9672) by +1.31 points. We are explicit about limitations: results are on a single motion-blur distribution, numbers are from a single seed, and calibration is qualitative rather than measured.
Duy Tran Thanh
OneMount Group, a Vietnam-based company in the finance and banking sector.