Your selection is explicit
Keep, Review and Exclude are choices you make. The workbench never overwrites a decision because of a model prediction.
Let AI assist.
Keep your eye in charge.
Photo selection, subject recognition and image editing are different jobs. Compare four documented photography workflows, then use a local workbench that preserves your decisions and makes its limits visible.
Choose the task
before you choose the tool.
A practical shortlist
For a culling first pass, compare Imagen, Narrative Select and Aftershoot against your actual selection workflow. For local adjustment masks after selection, Lightroom Classic serves a different job. The recommendations here are editorial inferences from official feature documentation; no commercial product test, measured time saving or accuracy ranking is claimed.
| Product | Documented role | What to evaluate | Workflow distinction |
|---|---|---|---|
| Imagen | Culling methods can select preferred images within groups or a target image count; Culling Studio supports manual review and changes. | Test whether grouped candidates and the requested count retain the moments you need, then inspect the handoff to your editing project. | A requested output count is a workflow preference, not proof that every selected frame is artistically right. |
| Narrative Select | First Pass organizes scenes into five ranked tiers with explanations; the photographer makes the final selection. | Inspect the explanation and nearby frames rather than accepting a tier as your final artistic judgement. | A ranked first pass supports review. It does not remove the need to check expression, story and delivery requirements. |
| Aftershoot | Automated culling preferences include duplicate grouping, blurry-photo detection and closed-eyes detection. | Compare strictness settings and revisit grouped alternatives. Check intentional movement and meaningful expressions yourself. | Preferences change the candidate set; a flagged image can still be useful in its context. |
| Lightroom Classic | AI-supported masks can target subjects, sky, backgrounds, objects and people for nondestructive Develop adjustments. | Inspect mask boundaries and local adjustment results on the photos you intend to deliver. | Develop masking is an editing capability. It is not a replacement for a culling first pass. |
Official documentation checked October 5, 2026. Pricing, hardware performance and file-format coverage are not compared here; verify the relevant product's current requirements before buying. This is a focused workflow comparison, not an exhaustive list of every AI photography product.
Too many similar frames?
Compare the culling products' documented grouping and review interfaces. Use a representative sequence and examine the rejected alternatives, not just the selected folder.
Need a consistent edit?
Selection and editing are separate stages. Evaluate the editing workflow against your own references after deciding which frames deserve work.
Need useful local organization?
The workbench below adds a human review manifest, optional subject-label predictions and pixel measurements. It does not emulate a professional face or scene-ranking model.
Assistance with visible boundaries.
Real inference, a narrow task
A pretrained MobileNet model suggests ImageNet subject classes. It cannot rank artistic quality, detect closed eyes or recognize a named person.
Useful review exports
Download a contact-sheet PNG, CSV or JSON containing your notes and decisions. Original images remain untouched.
Your local culling workbench.
Choose JPEG, PNG or WebP images. RAW and camera-specific formats require a compatible image editor first. The workbench stores resized previews in this tab; files are not uploaded.
Load files or try the demo. Model weights load only when you request inference.
The demo contains original generated patterns, not real photographs or preset recognition answers. The model will produce real predictions on those patterns, which may be unhelpful. Use your own photos to evaluate subject recognition. No quality score or automatic rejection is produced.
No images loaded.
Exports include all images and their decisions even when the screen is filtered. These are review manifests, not Lightroom flags or sidecar files. No image is deleted by choosing Exclude.
What does the model actually do?
The page loads TensorFlow.js 4.22.0 and a pretrained MobileNet v1 0.25-width model from site-hosted files, then evaluates the image pixels. The top three ImageNet classes and model probabilities are shown as suggestions. Probabilities are not calibrated accuracy, image quality or artistic scores. Subject classes can be wrong, especially for unfamiliar, abstract or heavily edited images.
Resized RGB pixel measurements report mean luminance, near-black/near-white pixel proportions and edge variation. These are simple computations, not machine learning. They depend on scene content, resizing and processing; intentional silhouettes and blurred backgrounds are not automatically defects.
The model and runtime add a few megabytes of downloads; exact assets are listed in the source section. Inference can be slower on older devices. If WebGL fails, the tool attempts TensorFlow's CPU backend. If the model is unavailable, manual selection and exports remain useful; no simulated predictions are substituted.
The strongest frame may be the one a classifier misunderstands. Use a prediction to organize your review; use expression, light and the story to make your selection.
A review process
you can account for.
Keep the source intact
Use copied or exported previews for browser review and retain the camera originals. A resized contact sheet helps navigation, but it cannot establish critical focus at full resolution.
Separate technical review from the story
Inspect focus, expression and framing in the appropriate editor. Then review whether a frame contributes to the brief. A generic subject label cannot tell you whether a gesture matters to the client.
Compare the alternatives
When a culling product groups similar frames, check the neighboring candidates and the reason for the recommendation. If you keep a technically imperfect photograph for its narrative value, record that intention.
Make the handoff explicit
Export your review manifest with filenames and notes. Reconcile it with the original folder in your editor. The workbench's CSV does not modify catalog ratings, move files or apply edits.
Questions for your own edit.
“Does this frame add a different moment, or just another version of the same one?”
Review prompts, not customer quotations.
Useful distinctions.
Is the model an AI quality assessor?
No. MobileNet predicts subject classes from ImageNet. Decisions remain yours. The pixel measurements are separately labelled computations and do not prove focus accuracy or exposure quality.
Can the workbench detect duplicate shots or closed eyes?
No. It does not implement burst similarity, face analysis, face identity or closed-eye detection. The commercial products' documented culling capabilities are discussed above; this local workbench has a narrower role.
How private is local inference?
Image decoding, metrics and model inference run in your browser. The tool loads its runtime and weights from this site and does not upload images to an inference service. Keep the export you need; clearing the set or reloading removes this tab's review state.
Can I recover the originals from the contact sheet?
No. The contact sheet contains resized review images. Keep the original photo files separately and use filenames to reconnect your manifest with the source folder.
Sources and implementation.
Checked October 5, 2026. No independent test of the compared commercial products, paid-plan price or measured workflow saving is claimed.
Culling workflows
Imagen culling documentation, Narrative First Pass and Aftershoot's getting-started guide support the feature descriptions in the comparison.
Editing workflow
Adobe's Lightroom Classic masking documentation supports the distinction between AI mask creation and final local adjustments.
Real browser inference
TensorFlow's MobileNet model documentation describes ImageNet classification and its limitations. The local model uses the official MobileNet v1 0.25 model manifest and TensorFlow.js.
Local pixel processing
Canvas pixel access supplies the decoded RGBA values used for the explicitly labelled numerical measurements. Pinned dependency and model source information records the shipped assets.