CastFaces

Computer vision, web search automation, and production QA

CastFaces

An identity-first pipeline for actor and public-figure profile cards: find candidates in open web sources, prove the image belongs to the right person, transform it into a strict portrait format, and ship every decision with a reviewable audit trail.

See the transformations Explore the workflow
Profile Cards
actor and public-figure portraits
Identity First
source evidence before image cleanup
400x400 PNG
strict output format and QA
Find Search public web results, profile pages, official sites, media databases, and configurable source adapters.
Verify Verify identity with names, dates, roles, aliases, source authority, snippets, and page context.
Normalize Align faces, clean backgrounds, standardize crops, upscale when useful, and validate the final portrait.
Deliver Package final images, source references, metadata, debug views, and operator reports.

What the system does

From a list of actors to verified, CMS-ready profile cards.

CastFaces is built for media catalogs, streaming content operations, talent databases, editorial archives, and public-figure directories where the hard part is not only finding a face. The system must prove whose face it is, decide whether the source is usable, and produce a portrait that matches a strict publishing format.

The architecture is modular: source discovery, proxy-aware collection, image processing, quality audits, retry logic, and operator review can run together or as independent services.

Open-source discovery

Collect candidate images using person name, birth date, role, aliases, and public credits from configurable sources.

Identity verification

Rank candidates with Identity Match, source authority, contextual evidence, disambiguation checks, and reject reasons.

Strict portrait format

Create consistent 400x400 PNGs with face size control, eye-line alignment, centered framing, and clean background.

Traceable QA

Generate Portrait Quality and Frame QA reports, debug overlays, retry history, and delivery summaries for review.

Scene-to-card transformation

Messy web candidates become compliant profile-card portraits.

Source images are rarely studio headshots. They can be conference photos, backstage images, screenshots, press stills, or phone-like crops. The engine detects the face, estimates geometry, evaluates whether the image is usable, removes distracting background, aligns the eyes, controls face size, and validates the result against the output spec.

Synthetic examples of messy public source images converted into clean profile-card portraits
Demo board: open-source scene candidates converted into clean portrait-card outputs.
DEV Pass 99%

Conference scene extraction

Angled face, off-center framing, people in the background, and mixed lighting. The output is straightened and normalized.

Fictional profile
Evan Hartley
DOB
Apr 18, 1987
Roles
Actor, producer, festival guest
Known history
2 credited titles, 1 panel appearance
Verification
High confidence, 96%
  • Official bio: exact name, matching headshot context, and role listed as actor.
  • Festival page: conference appearance matched to a credited panel photo.
  • Credits: linked to fictional titles Northline and Harbor Night.
Distorted web source
Compliant portrait

Real DEV debug: standard overlay plus projected person mask, eye shapes, mouth shape, and source-edge markers.

Blocked by QA 40%

Backstage candidate cleanup

Blur, crowd context, poor background, and partial motion. The engine preserves identity evidence, then blocks the output when source-crop risk remains visible.

Fictional profile
Maya Sterling
DOB
Nov 03, 1991
Roles
Actor, voice performer
Known history
Series launch, cast page, press caption
Verification
Strong match, 91%
  • Cast listing: exact name appears near role and production title.
  • Event caption: backstage image context matches fictional series launch.
  • Alias check: alternate spelling resolved without date or role conflict.
Distorted web source
Blocked output

Real DEV overlay shows the same landmarks and crop gates used to block visible source-cut risk.

DEV Pass 98%

Phone-like crop correction

Tilted close-up, edge crop, perspective distortion, and tight framing. The result is centered and sized for profile use.

Fictional profile
Daniel Cho
DOB
Jan 22, 1989
Roles
Actor, host, public speaker
Known history
Profile source, event page, host listing
Verification
Accepted, 88%
  • Profile source: matching name, profession, and public appearance metadata.
  • Image context: nearby page text confirms the same fictional person.
  • Conflict check: same-name candidate rejected because profession did not match.
Distorted web source
Compliant portrait

Frame QA verifies the square crop; mask and feature overlays expose eyes, mouth, face contour, and transition risk.

Visual examples on this prototype use synthetic, non-celebrity portraits generated for demonstration.

DEV engine assessment

Every demo output is checked with the same QA language the engine uses in production.

The numbers below come from CastFaces DEV runs on July 4, 2026 using the synthetic source and result files shown on this page. Final portraits are processed through the engine, rendered with real debug overlays, and scored against the same visible-cut, geometry, background, and format gates used by the delivery workflow.

Run ID CFDEV-20260704-SYN

Local DEV engine, synthetic demo assets, public-safe labels.

Result QA 2 / 3 pass

Two final portraits pass; one is blocked for visible source-crop risk.

Source risk 3 flagged

Blur and off-angle evidence are preserved before transformation.

Release package Auditable

Image, real debug overlay, source reference, and validation data.

Result QA 99%

Evan Hartley

result-1.png
Source risk 95%

Blurry; bad angle, yaw -31 deg, sharpness 42.

Final pass 99%

Face 75.0%, yaw -10 deg, eyes 0.335, sharpness 161.

Format
400x400 PNG
Watermark gate
100%
Text/logo gate
100%
Background transition
98%
Circle final 99 green / OK
Square frame 68 yellow / right source crop chord
Base score 99.86 weighted component subtotal
Modifier x1.00 no adjustment applied
Face size
75.0%
Yaw / pitch
-10 deg / 0 deg
Eye openness
0.335
Sharpness
161
Background std
6.0
Watermark probability
0.1127
Face integrity100 / 35.00 pts Body framing100 / 20.00 pts Face scale100 / 15.00 pts Background cleanliness100 / 8.00 pts Background transition98 / 6.86 pts Eyes100 / 6.00 pts Sharpness100 / 6.00 pts Pose100 / 3.00 pts
Result QA 40%

Maya Sterling

result-2.png
Source risk 95%

Blurry source, sharpness 15, yaw 25 deg.

Final pass 40%

Blocked: visible top source crop. Face 75.5%, yaw 3 deg, sharpness 240.

Format
400x400 PNG
Visible-cut gate
Fail
Watermark gate
100%
Text/logo gate
100%
Circle final 40 red / visible top source crop
Square frame 40 red / same visible-cut defect
Base score 91.11 before visible-cut penalty
Modifier x0.45 visible-cut penalty
Face size
75.5%
Yaw / pitch
3 deg / 0 deg
Eye openness
0.367
Sharpness
240
Background std
7.5
Watermark probability
0.1278
Face integrity75 / 26.25 pts Body framing100 / 20.00 pts Face scale100 / 15.00 pts Background cleanliness100 / 8.00 pts Background transition98 / 6.86 pts Eyes100 / 6.00 pts Sharpness100 / 6.00 pts Pose100 / 3.00 pts
Result QA 98%

Daniel Cho

result-3.png
Source risk 95%

Blurry source, sharpness 22, tight crop risk.

Final pass 98%

Face 76.0%, yaw 1 deg, eyes 0.309, sharpness 208.

Format
400x400 PNG
Watermark gate
100%
Text/logo gate
100%
Background transition
81%
Circle final 98 green / cutout edge noted
Square frame 99 green / OK
Base score 98.67 weighted component subtotal
Modifier x1.00 no adjustment applied
Face size
76.0%
Yaw / pitch
1 deg / 0 deg
Eye openness
0.309
Sharpness
208
Background std
6.8
Watermark probability
0.1833
Face integrity100 / 35.00 pts Body framing100 / 20.00 pts Face scale100 / 15.00 pts Background cleanliness100 / 8.00 pts Background transition81 / 5.67 pts Eyes100 / 6.00 pts Sharpness100 / 6.00 pts Pose100 / 3.00 pts

Output requirements

Every accepted image is transformed to the publishing spec.

The goal is not a pretty thumbnail by chance. The output must match a repeatable profile-card contract: fixed size, face proportion, eye-line, centering, clean background, no text or logos, and a traceable validation report.

Real DEV mask overlay: person mask, eye shapes, mouth shape, source edges, and circular-safe composition.
400x400 PNG Fixed square output for predictable CMS and card rendering.
Face 70%+ Face/head scale is controlled so the portrait fills the card without feeling cropped.
Eye-line alignment Landmarks drive rotation and vertical positioning; off-angle sources are corrected when safe.
Centered identity Front-facing portraits are centered; three-quarter faces allow controlled artistic offset.
Clean background Background is segmented, smoothed, color-balanced, and checked for artifacts.
No text or logos Watermarks, signage, poster text, and UI remnants are detected before release.
Single-person focus Multi-person and ambiguous crops are downgraded or rejected before delivery.
Debug + JSON report Each output can ship with landmarks, masks, geometry, quality signals, and decision history.

Quality control

The system knows when not to ship a portrait.

CastFaces separates identity confidence from visual quality. Identity Match decides whether the image likely belongs to the right person. Portrait Quality and Frame QA decide whether the processed result is usable as a circular or square asset.

  • Watermarks, text, logos, and poster-like backgrounds can hard-cap quality.
  • Visible source cuts, floating heads, bad transitions, and dirty backgrounds are scored explicitly.
  • Auto Fix retries safe adjustments, compares the new score, and rolls back regressions.
Blocked 40%

Blocked with a readable reason

The real DEV overlay and validation report flag visible top source crop. The operator can rerun search, adjust processing parameters, or reject the candidate before delivery.

Workflow

Built as services, not a one-off script.

The same project can be used through an operator UI, direct REST API calls, or a mixed review-and-automation mode.

  1. 01

    Load people

    CSV, JSONL, JSON, or TXT with profile IDs, names, dates, roles, and optional public credits.

  2. 02

    Fetch candidates

    /fetch_photos collects images, URLs, snippets, dimensions, and source errors.

  3. 03

    Score identity

    Identity Match applies source authority, name/date/context matching, caps, and rank adjustment.

  4. 04

    Process portraits

    /process_photo returns result PNGs, debug overlays, metadata, and validation.

  5. 05

    Improve automatically

    /smart_fix/recommend suggests safe next parameters for geometry, background, and detail issues.

  6. 06

    Export delivery

    Bundles include final images, source references, reports, JSON traces, and summary CSV files.

Operator experience

A reviewable pipeline for real media operations.

Operators can monitor search progress, inspect candidates, launch full-auto processing, switch to manual controls, and export a delivery bundle with every decision preserved.

Search Queue Online
Search gallery and candidate review
Processing Workbench Auto Fix
Quality
91%
Automatic retries with manual override controls
Delivery Bundle Ready
  • result_circle.png
  • result_square.png
  • debug_overlay.png
  • validation_report.json
  • source_reference.json
Auditable export for CMS or downstream review

Image rights and compliance

Built to support lawful image workflows, not bypass them.

This website prototype uses synthetic demonstration portraits. They are not real public figures, clients, or production search results.

CastFaces does not grant ownership, copyright, publicity, privacy, or trademark rights in any third-party image found online. Production use should be limited to images that are owned, licensed, public-domain, consented, contractually authorized, or otherwise cleared for the intended commercial use.

Operators remain responsible for reviewing source terms, image licenses, model releases, right-of-publicity rules, privacy requirements, takedown requests, and customer-specific compliance policies before publishing. This notice is practical product guidance, not legal advice.

Engineering profile

Useful for teams that need practical ML, backend, and automation work.

This project demonstrates production-minded engineering: service boundaries, GPU-aware processing, explainable audits, operator tools, cache management, deployment scripts, and customer-facing documentation.

FastAPI Streamlit Docker CUDA 12.1 Redis BiRefNet Real-ESRGAN YOLO Face landmarks REST APIs JSON reports Batch exports

For jobs and clients

Need a system that finds, cleans, verifies, or packages visual data?

CastFaces is a strong starting point for photo enrichment, media catalog automation, identity-aware scraping, content QA, and custom CV pipelines.

Discuss a project Review capabilities

Project inquiries route to the CastFaces contact inbox.