6 controlled samples
Reference, reduced resolution, soft focus, low exposure, high exposure, and heavy JPEG compression.
Original input research
Face swap input readiness benchmark
One owned source image, six controlled variants, five transparent pixel measurements, and zero generated-output claims. Use the data to understand what a preflight checker can measure before upload and what it cannot predict.
Direct answer
This is an input stress test, not an AI accuracy leaderboard
The experiment changes one variable family at a time around the same DeepSwapAI-owned source creative. It publishes decoded pixel measurements and exact transformations. It does not run a face swap, compare models, estimate identity similarity, or declare a quality winner.
5 pixel measurements
Mean luminance, contrast standard deviation, dark clipping, highlight clipping, and adjacent-edge difference.
0 output claims
No generated image, success rate, realism score, identity score, speed result, or product ranking.
Controlled variants
See the exact image set behind the measurements
Every plate derives from the same source asset. Visual inspection remains necessary because aggregate metrics cannot show every local artifact.
Measured data
Decoded-pixel measurements at a maximum 512-pixel analysis side
Luminance uses Rec. 709 coefficients. Dark clipping counts luminance at or below 10; highlight clipping counts luminance at or above 245. Edge difference is the mean absolute luminance difference across horizontal and vertical neighboring pixels.
| Variant | Dimensions | Mean luminance | Contrast SD | Dark clipped | Light clipped | Edge difference |
|---|---|---|---|---|---|---|
| Reference encode | 960 x 540 | 125.82 | 58.95 | 1.22% | 0% | 7.74 |
| Reduced resolution | 320 x 180 | 126.22 | 58.81 | 1.1% | 0.01% | 10.55 |
| Soft focus | 960 x 540 | 124.8 | 55.19 | 0.16% | 0% | 2.68 |
| Low exposure | 960 x 540 | 57.18 | 30.37 | 9.22% | 0% | 3.87 |
| High exposure | 960 x 540 | 197.15 | 62.53 | 0% | 36.64% | 8.41 |
| Heavy JPEG compression | 960 x 540 | 125.58 | 58.95 | 1.29% | 0.01% | 7.45 |
Important: a value is a measurement, not a universal pass mark. The compressed JPEG shows why: its global measurements remain close to the reference even though visual block and ringing artifacts can still be present.
Reproducible method
Exact transformations and immutable file fingerprints
The source SHA-256 is 2f293e48f7d5d6a08ce9ab759a124e00ce92f2a204cd530284ce7d1f0ecc7c3d. Each output is decoded, composited over white if needed, resized proportionally to no more than 512 pixels on its longest side, and sampled in row order.
Reference encode
Resize the owned source to 960 x 540 and encode as WebP at quality 82.
SHA-256: 8c852071b02e03d91eb264655fef14654026fa46dc32e736a2b25c2ab0a1186a
Reduced resolution
Resize the same source to 320 x 180 and encode as WebP at quality 82.
SHA-256: 955c0e4e97a7af8564b39733b9654d3294abd07fa014c530ca733d8708000cc0
Soft focus
Resize to 960 x 540, apply a Gaussian blur with sigma 4.5, and encode as WebP at quality 82.
SHA-256: e8edf51e0ef2b38916c858fe4aff932b14081d60ebd55552fbfc360586350886
Low exposure
Resize to 960 x 540, apply y = 0.52x - 8 to RGB channels, clamp to 8-bit range, and encode as WebP at quality 82.
SHA-256: 727172796e3cf8efdefd601d21177321c1bfcfbc458d629a689897aecf442769
High exposure
Resize to 960 x 540, apply y = 1.35x + 45 to RGB channels, clamp to 8-bit range, and encode as WebP at quality 82.
SHA-256: f2b01d592108fea3bb82e888733d89fb7356c307cda00d6d55d587b422da957b
Heavy JPEG compression
Resize to 960 x 540 and encode as JPEG at quality 15 with 4:2:0 chroma subsampling.
SHA-256: cd952c4c50fb78e941a62492e7ef6c3cc2fae96abac7df87aa1f32d4106b97ae
Download the complete dataset
The JSON and CSV contain the formulas, scope, limitations, transformations, dimensions, byte sizes, metrics, and SHA-256 fingerprints. The BibTeX file provides a stable citation.
Research data catalog
Six reusable assets with explicit evidence boundaries
This catalog keeps input measurements, single-output human review, metric selection, official-source product review, API-contract review, and illustration-pixel analysis separate. Cite the specific asset you used, its version, and its canonical URL rather than transferring a claim from one evidence type to another. The website remains canonical; the public research repository and immutable, checksum-verified v1.6.2 release provide a versioned external mirror. GitHub's signed release attestation binds commit c1d268d4a140
Controlled input-readiness benchmark v1.0.0
Six controlled image variants, five decoded-pixel measurements, exact transformations, file fingerprints, JSON, CSV, and BibTeX. Canonical dataset: deepswapai.com/research.
Face swap output quality scorecard v1.0.0
A separate seven-criterion human-review rubric with local calculation, reusable JSON and CSV, and explicit limits on biometric or provider-level claims. Open the scorecard.
Face swap evaluation metric decision map v1.1.0
9 source-cited evaluation families plus a 3-protocol research crosswalk, each with explicit inputs, prerequisites, and interpretation boundaries. No metric is run and no provider is scored. Open the metric map.
Source-checked face swap tool landscape v1.0.0
13 normalized product records backed by 66 current official-source URLs, with visible unknowns and no output-quality ranking. Open the comparison dataset.
Face swap API contract landscape v1.0.0
6 normalized API records backed by 21 current official-source URLs. It preserves incompatible price units and official-source conflicts without running a generation request or ranking providers. Open the API contract dataset.
Official demonstration panel audit v1.0.0
Three aligned decoded-pixel regions, exact crop coordinates, formulas, source fingerprint, JSON, CSV, and BibTeX. It measures one illustration and does not score a generated output. Open the before-and-after audit.
Download the canonical research catalog
The JSON-LD describes all six datasets, their evidence boundaries, canonical records, licenses, citations, and distributions. The CSV provides the same inventory for tabular tools.
Citation and reuse: cite “DeepSwapAI Controlled Image Input-Readiness Benchmark v1, https://deepswapai.com/research”. The JSON, CSV, formulas, transformations, and measurement metadata use CC BY 4.0. Referenced sample images remain DeepSwapAI-owned and are not relicensed for unrelated reuse.
Interpretation boundary
What the experiment establishes and what remains unknown
Resolution is explicit
Pixel dimensions and file size are exact properties. Upscaling a small image changes dimensions but cannot reconstruct missing identity detail.
Exposure leaves measurable clipping
The controlled bright variant moves 36.64% of sampled pixels to the highlight-clipping range. That establishes lost input range, not a predicted output defect.
Blur lowers local edge change
The soft-focus variant reduces adjacent-edge difference from 7.74 to 2.68. The metric is scene-wide and does not isolate a face.
Compression needs visual review
Heavy JPEG compression can preserve global averages while creating local artifacts. A technical preflight must not turn a few metrics into a false quality score.
Limitations: one source scene is not a population. The analyzer does not detect faces, identity, consent, pose, expression, gaze, hair, occlusion, or model behavior. It cannot predict realism, accuracy, processing success, preference, or final export quality.
From input evidence to output review
Use a separate protocol after generation
Input measurements and output observations answer different questions. After generating a representative result, use the face swap quality scorecard to document identity, blending, pose, lighting, occlusion, technical integrity, and temporal stability without turning one output into a model-wide claim.
Research questions
Answers without overstating the evidence
Does this benchmark measure face swap output quality?
No. It measures controlled input-image properties only. No face swap output was generated, scored, or compared.
Can it predict a realistic result?
No. The measurements do not observe the generation model, identity, pose, expression, occlusion, or output.
Why publish a compressed sample that looks similar in the table?
Because it exposes a real limitation: global aggregate measurements can miss local JPEG artifacts.
Where can I test my own image?
Use the local readiness checker. It applies the same five measurement definitions in your browser without uploading the selected image.
How can I cite or reuse the data?
Use the downloadable BibTeX record or the visible attribution statement. Measurement data and metadata use CC BY 4.0; referenced sample images remain DeepSwapAI-owned.
Measure your own input without uploading it
Open the local checker, inspect each metric and warning, then make the smallest representative generation test.