Identity consistency
Compare the same eye, brow, mouth, jaw, and age cues across every result.
Batch workflow guide
Consistency is a batch workflow, not one setting
A photo set amplifies small input differences. Group similar targets, lock the identity reference, test a representative frame, and review the set with one acceptance standard.
Plan the set
Group targets by the visual problem they share
Do not treat a mixed set as one uniform input. Separate targets by pose, lighting, expression, resolution, or occlusion so one reference is not forced across conflicting conditions.
| Grouping signal | Keep together | Split into another test |
|---|---|---|
| Pose | Front-facing and slight turns | Extreme profiles or steep head tilt |
| Lighting | Similar soft or directional light | Night scenes, colored light, or hard backlight |
| Expression | Neutral and mildly expressive faces | Closed eyes, wide mouth shapes, or extreme expression |
| Resolution | Targets with comparable face size and detail | Tiny, compressed, or heavily filtered faces |
| Occlusion | Clear faces or similar accessories | Hands, hair, glasses glare, masks, or props over landmarks |
Pilot run
Test the middle and the edge case
- Choose one representative target: use the image that best reflects the majority of the set.
- Choose one difficult target: include the most important profile, shadow, expression, or occlusion case.
- Hold the reference constant: compare target conditions before swapping identity references.
- Review with one rubric: use the same identity, scene, and technical checks for both results.
- Split or scale: process the full group only when the pilot passes; move difficult images to a separate group.
The face swap quality guide explains which reference and target conflicts to fix first. If your goal is to replace several people inside one group photo rather than produce several output files, use the batch vs multiple face swap decision guide.
Ecommerce and apparel catalogs
Build each permitted catalog batch around SKU, angle, and review state
A catalog workflow needs more structure than a folder of mixed images. Confirm the model release, likeness permission, source-image rights, intended sales channel, and disclosure decision before upload. Then preserve the product record that connects each target, generated output, reviewer decision, and approved export.
| Control | Keep constant inside one batch | Review before approval |
|---|---|---|
| SKU or look | One product, colorway, styling set, and intended channel | Garment boundaries, labels, closures, jewelry, and accessories remain attributable to the target |
| Camera angle | Front, three-quarter, profile, back, or detail views in separate groups | Reference pose compatibility, jaw, ears, hairline, and occluded edges |
| Lighting and scene | One studio setup, location, exposure family, and crop pattern | Face color and shading remain coherent with the target scene |
| Identity reference | One approved reference per pose and lighting family | Identity cues remain consistent without changing target hair, body, clothing, or background |
| Export record | SKU, angle, source version, output version, reviewer, and decision | Every approved file maps back to the authorized target and generated task |
- Inventory the set. Assign SKU, angle, lighting group, rights record, and intended channel before generation.
- Pilot one typical and one difficult image. Include the most important profile, hand, hair, accessory, or hard-shadow case.
- Lock the approved reference for that group. Changing identity and scene conditions at the same time makes failures hard to diagnose.
- Generate within current batch limits. Use no more than 20 images, 30 MB each, and 95 MB combined per request.
- Review every output, not a sample average. Use the quality scorecard and keep product, rights, and disclosure checks outside the numerical score.
- Export only approved versions. Keep rejected and retry files out of the publishing folder and retain the authorization and review record.
Evidence boundary: face replacement does not verify garment accuracy, fit, product truth, marketplace compliance, or conversion performance. Human review remains required for every published catalog image.
Quality control
Review the set as a sequence
Scene preservation
Check that hair, accessories, hands, lighting, and framing still belong to each target.
Outlier detection
Sort failures by pose, light, expression, resolution, or occlusion instead of retrying the whole set blindly.
Retry discipline
Change one input variable at a time and keep approved outputs separate from the retry group.
Operational rule: a batch is ready when its weakest important image meets the intended use, not when the average result looks acceptable.
Modes, formats, limits, and cost
Know the output count before you submit
Before uploading, use the free local batch manifest planner to validate the mapping direction, file counts, declared formats, sizes, planned outputs, and exact credits without sending the selected files. It does not detect faces, verify identity or consent, judge scene compatibility, or predict output quality.
The current batch face swap workflow supports one face applied to multiple target photos or multiple source faces applied to one target photo. It accepts JPEG, PNG, WebP, AVIF, HEIC, and HEIF images, allows up to 20 images per batch, limits each image to 30 MB and the combined batch to 95 MB, and charges 3 credits per generated output. The exact output count and credit total are shown before submission.
The selected files are checked locally for readiness before upload. Uploaded and generated media is removed from DeepSwapAI servers within 24 hours. Trial image exports from an account with no completed credit purchase visibly contain DeepSwapAI.com and the account's 100% complete email address. Every character is visible: the address is not masked, shortened, hashed, or omitted, and the interface shows that exact full address before generation. After any completed one-time credit purchase, future image exports are watermark-free. Review the privacy and consent checklist before uploading or sharing another person's likeness.
How this guide was made
A repeatable production checklist, not a volume or conversion claim
The DeepSwapAI Product Team organized the guide around the current batch workflow and variables a user can observe and change. The ecommerce section is an editorial control reference, not evidence of garment accuracy, marketplace acceptance, throughput, conversion lift, or business results. Mode, format, cost, limits, browser-local manifest boundaries, privacy, watermark, and catalog-review facts were reviewed on July 28, 2026; see the verification methodology.
Pilot two images before the full set
Use one representative target and one difficult edge case. Scale only after both meet the same review standard.