Slides Framework

Quick primer

Before we start

The words you will hear in the next slides, in plain language.

  • Text AI

    LLM

    Large language model. Trained on huge amounts of text, it reads and writes text: answers, summaries, judgments. It only knows what you show it.

  • The engine

    Model

    The trained file that does one job. Cutout, ghost mannequin and try-on each use their own image model. Different job, different model.

  • Where it runs

    Local vs cloud

    Local: our own hardware. Files stay in the studio, size is limited by the machine. Cloud: a provider’s servers. Bigger models, pay per use, files leave our building.

  • Doing the work

    Agent vs LLM

    An LLM answers one prompt. An agent chases a goal: it calls tools, checks the result and retries until done. Every agent uses an LLM; an LLM alone is not an agent.

Today: image generation and QC run on our own local server. The steps around them are scripted, not agents yet.

Who we are

Living Colors Studio

Product-image production: cutouts, clipping paths, ghost mannequin, colour and retouching.

  • How it works today

    A client sends files and instructions. A designer edits, and we reply. It works, and clients like talking to a person.

  • The problem

    More orders means more people. We want volume to grow without headcount growing alongside it.

  • What we’re building

    AI for the repeatable pieces, so volume stops growing with headcount. Taste, judgment and client conversations stay human.

What we'll cover

AI Capabilities

  • 01

    Photoshop-Enabled AI

    We use Adobe Photoshop’s built-in AI features in our day-to-day production workflows to support tasks such as generative image extension, background expansion, and selected image manipulations. These capabilities enhance our existing editing process while keeping our production team in control.

  • 02

    AI-Assisted Workflows

    We are building dedicated AI-assisted workflows for specific image-production tasks. They are working prototypes, not a finished service. Cutout keeps the original pixels; ghost mannequin, try-on and extract generate new ones, so each result is checked by QC and a reviewer. In the following slides, we’ll demonstrate them live, review the generated results, and show where quality control fits.

AI Assisted Workflows

Work in progress · All local, no cloud

  • Cutout
    Cutout
  • Ghost mannequin
    Ghost mannequin
  • Extract dress
    Extract dress
  • Virtual try-on
    Virtual try-on

What we’re building

Meet ALOalo

A pipeline executes one step on one image, well. It doesn't know whose image it is. Once the pipelines got fast, the bottleneck moved to coordination — reading instructions, remembering each client's rules, routing, checking, delivering. ALO is the coordination layer we’re building for that. It isn’t an editing tool, and it’s still learning and improving.

  • Account Agent

    Knows each client’s rules — white value, shadow policy, naming — so every job starts with them attached.

  • QC Agent

    Checks outputs against the client’s standard — edges, color, shadows. In testing today.

  • Dispatch Agent

    Knows every designer’s skills and sends each task to the right person: AI, a designer, or a senior reviewer.

Runs on local LLMs wherever possible — images and instructions stay on systems we can audit, not black boxes. Human in the loop, always.

The aim

From order to delivery

How ALO is meant to fit around the people and tools already in the process.

  1. 1Order arrivesEmail with a link and instructions
  2. 2UnderstandClient, rules, where the files are
  3. 3RouteAI pipeline or a designer, only if the client agrees to AI edits
  4. 4EditPipeline or designer
  5. 5QCChecked against that client’s standard
  6. 6DeliverUpload, reply, close the loop
  • Fits what you already use

    Connects to email, project tools and your CRM, scoped to how each studio works. It reads and routes; it doesn’t replace them.

  • People stay in front

    Everything the client sees and hears stays human. A reviewer checks the work, and designers keep the judgment calls.

  • Cutout & background removal

    Clean, studio-grade cutouts straight off the product shot — hair, fabric edges, and jewelry kept intact. Runs on our own server, nothing leaves the building.

    • Before
    • After
    $ ./ImageStudio/run.sh "in/models/ZA716222LEOMOD.tif"
    • Original photo
    • Background removed
  • Hair matting

    Cutout & background removal

    Soft edges only where hair actually is — strand by strand, no halo and no chopped-off flyaways.

    • Before
    • After
    • Original photo
    • Background removed
  • No deletion

    Cutout & background removal

    Thin parts are kept, gaps are cut out — nothing gets deleted just because it's fine or fiddly.

    • Before
    • After
    • Original photo
    • Background removed
  • Clean export

    Cutout & background removal

    Ships as a layered PSD with a real layer mask and a vector clipping path, not just a flattened PNG.

    • Layers
    • Channels
    • Paths
    • Photoshop Layers panel
    • Photoshop Channels panel
    • Photoshop Paths panel
  • Guided by the back photo

    Ghost mannequin

    Two photos in: the front shot and a neck photo the model uses as a reference for the garment’s own lining — not a flat patch.

    • Raw
    • Neck
    • Ghost
    D="in/batch-04/ghost mannequin/29-9-2026/Blazer/"
    $ python workflow/run_ghost_mannequin_v3_fabrics.py --raw "$D/Raw file.jpg" --neck "$D/Neck files.jpg" --note "$D/Ghost Brief.txt"
    • Raw photo
    • Neck photo
    • Ghost mannequin result
  • Works on any garment

    Ghost mannequin

    Sheer sleeves, a fitted waist, a collar that has to sit right — the same pipeline, no per-garment tuning.

    • Before
    • After
    • Raw photo
    • Ghost mannequin result
  • One photo, any outfit

    Virtual try-on

    Swap the garment, not the shoot — one source photo, as many dresses as you have.

    • Original
    • Navy
    • Red
    • Black
    • Original photo
    • Navy dress try-on
    • Red dress try-on
    • Black dress try-on
  • Two photos, one swap

    Virtual try-on

    A person photo and a flat garment shot in — the same person, same pose, wearing it. A --notes line fixes the hand.

    • Person
    • Garment
    • Try-on
    • Fix
    $ python workflow/vto_workflow_klein_tryon_QC.py --person "in/Masking/Raw images/4.jpg" \
        --dress in/dress/dress_red.png --notes "dont put hand in the pocket as the dress dont have pocket"
    • Person photo
    • Garment
    • Try-on result
    • Try-on result, hand fixed with --notes
  • Clean extraction

    Extract dress

    The model, the pose, the room — gone. Just the garment, flat, on white.

    • Before
    • After
    • Model photo
    • Extracted garment
  • No invented labels

    Extract dress

    --no-label stops the model inventing a neck label that was never there.

    • Before
    • After
    $ python workflow/extract_clothes_qwen.py --input "in/Masking/3-10-2026/Raw Files/4.jpg" --item top --no-label
    • Default extraction
    • With --no-label
  • From image to motion

    Image to video

    The try-on result becomes a 5-second posing clip. Static camera, small movement, so the dress holds.

    • Raw
    • Try-on
    • Video
    $ python workflow/run_hunyuan_video.py vto3-person-d1.png
    • Raw model photo
    • Try-on result
  • How we keep control

    AI Quality Control

    A vision model grades every result against its source photos, in a fixed order, so a real problem can’t hide behind a minor one.

    1. 1Mannequin or model gone?Auto-fail
    2. 2Same garment? Colour, fabric, logos and cut match the source.Auto-fail if changed
    3. 3Hidden parts filled? Neck and openings show the real inside.
    4. 4Natural shape? Worn 3D form, centred on white.

    Real logged verdicts from our own test runs. We haven’t yet measured how often QC is wrong; the next step is measuring it against human reviewers.

  • Fail ornith-1.5:9b · ghost mannequin
    Source blazer photo
    Source
    Generated result that failed QC
    Generated
    The garment colour differs: the original blazer is navy blue, the generated version is a lighter slate/grey-blue. The cuff detail differs: the original shows white cuffs, the generated image shows dark/grey cuff openings.
    Verdict log — real runs, mixed jobs
    VTOFailWhite trousers never applied — model shown bare-legged.
    ExtractPassGarment colour, cut and fabric finish match the source photo.
    VTOFailFace noticeably different from the reference. Automatic fail.

Use case · e-commerce

One shoot, many outputs.

Reuse the photos you already have, instead of setting up another shoot.

Next SKU, next model: start again from photos you already haveOn-model photoone shootModel already shotearlier shootMannequin photoplus neck photoExtractgarment, on whiteGhost mannequinclean product formTry-ongarment on a modelVideo5 s posing clipProduct pageQC + reviewer check

Purple steps are AI. Useful when the product-shot setup was forgotten or there was no time for it.

Data & security

All local. No cloud.

Generation and quality control both run on a local server. A local vision model reviews every result before a person does. No images go to a third-party service.

  • Data

    Your photos are processed on our own servers. Nothing is uploaded to cloud services.

  • Security

    The QC check runs on a local vision model (Ollama) on the same server. No cloud API sees your images, and no third-party vendor sits in your data path.

  • Fully Local

    Open models on our own hardware, with specialised LoRAs for ghost mannequin and extraction.

    SAM 3.1FLUX.2-kleinQwen-EditOrnith-1.5

Our own hardware, with our retouchers reviewing every result.

Production challenges & limitations

Where AI still needs people

What a retoucher pushes back on every week, and how we handle it.

  • 01

    Per-image instructions

  • 02

    Reference-driven edits

  • 03

    Multi-step
    pipelines

  • 04

    Studio quality
    bar

  • 05

    Client
    standards

01

Per-image instructions

Every garment has its own brief. A sequin top and a sheer blouse can’t share one prompt, so each image gets its own written note and the model is re-anchored on it every time. Notes steer details like hands, jewelry and shoes; fine details like hems still need a human check.

← →

Known limits: colour — brightness is matched, but hue shifts on some fabrics need a manual check. Fine and sheer fabrics — lace, mesh and sheer sleeves need more review than solid garments. Volume — throughput depends on the GPU box; we size it to your daily volume in the pilot.

Where brands are heading

How brands are thinking

One example: a recent hiring post from a global sportswear brand.

“We’re building out a team defining how [the brand] uses generative AI to create content at scale — across eCommerce, social, retail, and more. Think: prompt architecture, evaluation frameworks, workflow automation, and a whole lot of creative + technical judgment.”
  • They want

    Content at scale

    Generative AI across eCommerce, social and retail, not one-off experiments.

  • They name

    Evaluation frameworks

    How AI output is judged is a job in its own right. Quality control is part of the plan from the start.

  • They need

    Production-ready systems

    Turning experiments (Midjourney, Runway, ComfyUI) into workflow automation that holds up every day.

The same direction we’re taking: automate the repeatable work, and measure quality.

How we think about it

Possibilities

Automation works where work repeats. Quality control is the natural first step, brand by brand.

  • Start here

    QC per brand

    Guideline, approved references and past verdicts.

  • Next

    Automate proven steps

    Only where QC is reliable.

  • Later

    Brand-specific looks

    Train a brand’s look into generation.

GuidelineReference imagesReviewer verdicts
→
Brand profile
→
Any vision modelswap as models improve
Why we think it works
  • It doesn’t edit images, so it can’t break a delivery.
  • Better vision models drop in; the brand profile carries over.
  • Every verdict can be checked against a human reviewer.

Collaborative workflow discussion

Rethinking the Current Workflow

Let's identify the opportunity together.

  • Understand

    How does work currently move from intake to delivery?

  • Break down

    What are the individual editing and quality-control steps?

  • Prioritize

    Which steps are repetitive, time-consuming, or difficult to scale?

  • Optimize

    Where could AI assistance, workflow automation, or human expertise improve the process?

Questions to start the conversation

  • Which part of your workflow would you most like to improve?
  • Is the priority lower cost, faster turnaround, greater capacity, or more consistent quality?
  • Are there services that could be separated into simpler, repeatable tasks?
  • What are the rules per brand — naming, white value, shadow policy, turnaround, QC handoff?

Let's Build it Together

Your production expertise. Our editing experience. Practical AI workflows.

What would you like to see more of?

Thank you.

Rashed Al Mamoon

rashed@livingcolors.studio

The Living Colors Studio Ltd.

  • A second opinion, before yours

    AI Quality Control

    A vision model looks at every generated image next to its source photos and grades it the way a reviewer would — in a fixed order, so a real problem can never hide behind a minor one.

    1. 1Mannequin or model fully gone? Any visible stand, neck or arm stump.Auto-fail
    2. 2Same garment? Colour, fabric, pattern, logos and cut match the source photo.Auto-fail if changed
    3. 3Hidden parts filled correctly? Neck and openings show the real inside, not a hole.
    4. 4Natural shape? Worn 3D form, centred on clean white.

    Stitching, fold placement, label text — those are logged as notes, never a fail. The checklist changes per task: ghost mannequin, try-on and extract each get their own priority order.

  • QC agent running against a job

    The QC pass running against a live job, right after generation

  • QC in testing

    A real fail, not a hypothetical

    This blazer run looked fine at a glance. The QC model compared it to the source photo and called it — here's the actual verdict it wrote.

    Fail ornith-1.5:9b · ghost mannequin
    Source blazer photo
    Source
    Generated result that failed QC
    Generated
    The garment colour differs: the original blazer is navy blue, the generated version is a lighter slate/grey-blue. The cuff detail differs: the original shows white cuffs, the generated image shows dark/grey cuff openings. Fabric texture appears subtly different between the matte original and the generated piece.
  • Verdict log — real runs, mixed jobs
    VTOFailWhite trousers never applied — model shown bare-legged. Automatic fail under the leftovers rule.
    ExtractPassGarment colour, cut and fabric finish match the source photo; no person or mannequin visible.
    VTOFailFace noticeably different from the reference — eyes and nose altered. Automatic fail.
    GhostPassMannequin fully removed, garment shape natural. Note: small brand tag absent, expected for ghost shots.
    VTOPassPerson, scene and garment all match; no leftover clothing at neck, wrists or legs.

    Every line above is a real logged verdict from our own test runs — not a mockup. Generation quality is still improving run to run. QC has caught the real misses in these tests; we haven’t yet measured how often it is wrong, and checking it against your reviewers is the first thing we’d do in a pilot.

    And it isn't tied to our generator: the same check can read any image — including a human retoucher's finished file — against a reference photo and a written brief.

    D="in/batch-04/ghost mannequin/29-9-2026/Blazer"
    $ python workflow/qc_workflow.py ghost --refs "$D/Raw file.jpg" "$D/Neck files.jpg" --result "out/wf-ghost/Raw file_ghost/ghost_run008.png"