Benchmark Playground

Image Area Comparator

Compare selected areas across images, tune speed and scoring, then run a full benchmark suite over many saved pairs.

Source: WaitingTarget: WaitingActive models: 2Engine: idle

Source Reference

Upload source image and mark the area to compare

Accepted: PNG, JPEG, WEBP

Source Area Box JSON

left/top/right/bottom

Unlocked: uploading a new image resets area to default.

Target Candidate

Upload target image and mark the matching area

Accepted: PNG, JPEG, WEBP

Target Area Box JSON

left/top/right/bottom

Unlocked: uploading a new image resets area to default.

Control Center

Tune decision strictness, speed, and models here. Then run one compare or add the current pair to your benchmark suite.

Speed Presets

Mode: Custom

Presets are one-click starting points. You can still edit all controls manually at any time.

Decision Settings

Threshold 0.85

Pass/Fail cutoff0.85

Semantic (model) weight0.35

Pixel weight auto-calculates to 0.65.

Speed Settings

Minimum area size24px

Compare size224px

Quick modeRun only the first loaded model for faster responses.

Model Selection

2 selected

Maximum selected models: 12.

SigLIP Base Patch16

Newer than CLIP; strong default semantic signal at moderate cost.

idleunknown

SigLIP2 Base Patch16 (224)

Next-gen SigLIP2 variant; usually stronger but slower and larger.

idleunknown

DINOv2 Small

Modern self-supervised vision encoder; useful for structure-focused matches.

idleunknown

CLIP Base Patch32

Legacy baseline; very fast but often less discriminative than newer models.

idleunknown

CLIP Base Patch16

Legacy baseline with finer patches; still useful as a fallback check.

idleunknown

CLIP Large Patch14

Legacy larger CLIP backbone; slower but occasionally useful as a tie-breaker.

idleunknown

CLIP Large Patch14 (336)

Higher-resolution CLIP variant for fine-detail matching; highest cost.

idleunknown

Score Formula

Hybrid score = (0.35 x model score) + (0.65 x pixel score)

Runtime: Transformers.js in Web Worker | Aggregation: minimum-across-models

Run all saved pairs and review quality, speed, and consistency across cases.

No benchmark cases yet

Add source-target pairs to enable benchmark analytics.

Threshold 0.85

Progress: 0%

View: Live compareRun idleScored: 0

Pass rate

0.0%

0 pass / 0 fail

Avg hybrid score

0.0000

range 0.0000 -> 0.0000

Avg latency

Running: 0 | Errors: 0

Best / worst case

Best: -

Worst: -

Add current source/target pairs here, then run all to benchmark many image comparisons quickly.

Results

view: liveengine: idle

Compare selected areas to see conservative multi-model scores and per-model breakdown.

Source crop

Target crop