FOUNDING OFFERSign up to unlock 35% off your first plan.

How to Use Gemini Decision 1 for Image and Video QA Checks

SUMMARY

To use Gemini Decision 1 (gemini-decision-1) for image and video QA checks, submit your media and brief once through the v1beta Gemini Decisions API and evaluate bounded rules in a single pass. Configure noul checks for binary yes/no gates, choice for routing labels, and score for ordered quality rubrics inside the questions block. Each check returns a confidence value to route ambiguous files to human review under input-only pricing where shared media is billed once. Because probabilistic scores only audit inputs rather than producing finished creatives, Stillmore offers independent businesses a hands-off alternative: you upload existing product photos and offer details while Stillmore's team may evaluate tools like Gemini Decision 1 alongside editing and pre-delivery review to deliver finished images, videos, and captions you publish yourself.

Lan He

Stillmore Blog cover for Gemini Decision 1, featuring blue glass Gemini star sculpture.
Contents

How Gemini Decision 1 Runs Single-Pass Multimodal Checks

Gemini Decision 1 evaluates images and videos by scoring closed sets of candidate answers in one forward pass instead of generating text responses token by token. That setup lets a quality assurance pipeline score framing, label legibility, and policy rules across supplied photos, video clips, or rendered creative drafts in a single API request.

I reviewed Google DeepMind's official announcement to verify the model's documented limits, SDK entry point, and billing rules, summarized in the reference block below.

Documented Gemini Decision 1 Specs (Google DeepMind, Nov 11, 2026): gemini-decision-1 (google-genai v1beta client.decisions.create, Google AI Studio free prototyping tier) handles text, images, audio, video, or PDFs (256K token cap) across up to 1,000 noul, choice (2 to 255), or score (0-indexed) checks in 150 to 250 ms (16K context) at $0.10 per 1M input tokens ($0 output; shared input billed once).

As an illustrative audit, suppose a small beverage shop wants to screen one bottle photo alongside an offer note before designing a promotional post. Rather than running three separate generative prompts to check label visibility, pick a retouching queue, and grade counter clutter, one request evaluates all three rules against that shared photo and text pair.

Prerequisites and Input Preparation for Photos, Video, and Briefs

Before writing QA rules, configure your API key and SDK client, then assemble a structured input list that pairs each visual upload with the written business details your rules need to verify.

Per Google DeepMind's announcement, media uploads and text notes share a single Content array, with resolution controls trading fine detail against token cost.

For an illustrative pre-production check, pair a phone photo or short vertical clip with a plain text entry stating the exact product name, finish or flavor, and offer end date. A camera frame shows packaging and lighting, yet it cannot tell an automated rule that a weekend discount ends on Sunday, so keeping offer text beside the image prevents context mismatches early.

Asset Type Written Context to Bundle Common Pre-Check Flaw Acceptance Gate Before Upload
Product photo Exact SKU name, material, and offer wording Promotion dates or obscured back-label claims Zoom in beforehand to verify label legibility and full edges
Short video clip Scene goal, featured item, and target callout Spoken promo terms missing from visual frames Select clear visual detail when auditing small printed text
Apparel or soft goods Full garment view showing sleeves, seams, and tags Cut-off cuffs or warm-bulb color shifts Capture an unobstructed front frame in neutral lighting
Exported ad draft Approved headline copy, price, and brand spelling Subtle typos or clipped edges after resizing Run checks against the final exported aspect ratio

Step-by-Step Setup: Configuring noul, choice, and score Checks

Configuring Gemini Decision 1 takes five steps: bundle your visual asset and brief into the shared input list, add noul binary blockers, map choice options to production queues, build score rubrics for visual clarity, and enforce confidence floors for human review.

Following Google DeepMind's guide, a single request mixes those rules and returns candidate probabilities for downstream branching.

Below is an illustrative QA schema adapted strictly from the documented request keys (model, input, questions, type, instructions, criteria) to screen a coffee-bag photo before design work; inspect the live response object in your own environment before writing parser logic.

from google import genai

client = genai.Client(http_options={"api_version": "v1beta"})

result = client.decisions.create(
    model="gemini-decision-1",
    input=[
        {
            "type": "image",
            "uri": "https://example.com/assets/roast-bag.jpg",
            "mime_type": "image/jpeg",
        },
        {
            "type": "text",
            "text": "Product: House Espresso 250g. Offer: 2-for-1 through Sunday.",
        },
    ],
    questions={
        "label_blocked": {
            "type": "noul",
            "instructions": "Is glare, a hand, or framing hiding the front label?",
        },
        "intake_queue": {
            "type": "choice",
            "instructions": "Which production queue fits this photo?",
            "criteria": {
                "ready_for_design": "Full bag outline and label text are clearly visible.",
                "minor_cleanup": "Shape is complete, though minor dust needs retouching.",
                "request_reshoot": "Edges are cropped or printed text is unreadable.",
            },
        },
        "lighting_grade": {
            "type": "score",
            "instructions": "Rate subject separation from the counter background.",
            "criteria": [
                "Heavy shadow merges bag into cluttered counter.",
                "Usable light with minor competing props nearby.",
                "Clean separation with clear texture and outline.",
            ],
        },
    },
)
Pipeline Step QA Mechanism Merchant Check Objective Downstream Routing Action
1. Build shared context Shared input list Anchors every check to one visual file and brief Send combined asset and brief once
2. Set hard blockers noul binary gate Flags cut-off edges, blur, or glare over labels Reject upload or request another angle
3. Assign work queues choice routing buckets Sorts assets into retouch, edit, or reshoot lanes Route winning label to matching team queue
4. Grade visual clarity score ordered rubric Ranks subject lighting and background separation Prioritize cleaner uploads for hero spots
5. Gate low confidence Application confidence floor Catches borderline or ambiguous predictions Escalate uncertain files to human QA

Common Mistakes in Multimodal QA Prompts and Rubrics

Most noisy QA routing stems from four configuration mistakes: writing overlapping choice definitions, phrasing checks like open-ended critiques, omitting confidence fallback floors, and expecting automated scoring to recover product details never captured on camera.

As Google DeepMind's announcement outlines, each rule uses its own prompt text, so vague multi-part wording adds overhead and splits option odds; meanwhile, downstream retouching cannot reliably reconstruct a seam, cap texture, or label line missing from the source photo.

Consider an illustrative boutique coat check where one choice question asks, "Is this photo ready for Instagram?" That wording mixes sleeve framing, fabric tone, and background mess into one bucket. Replacing it with three atomic checks (one noul rule for full sleeves in frame, one noul rule for tag legibility, and one score rubric for background clutter) makes routing behavior far easier to audit against test uploads.

Configuration Mistake Routing Failure Symptom Checklist Fix Human Escalation Trigger
Overlapping choice rules Shared traits split probability across options Restrict each label to one distinct visual trait Route close label ties to a reviewer
Open-ended prompt wording Closed-set scorer cannot write editing notes Split into binary noul or ordered score checks Return specific reshoot steps on failure
Missing confidence floor Borderline glare or dim clips pass unchecked Enforce a minimum confidence cutoff per check Send sub-threshold uploads to manual review
Overusing fine detail settings Extra visual tokens raise cost without clearer text Match detail level to the smallest needed check Request a closer still shot for fine print

Checking Results and Separating Model QA Scores from Finished Creatives

Validating a Gemini Decision 1 workflow means testing borderline assets against your confidence cutoffs while keeping automated rule scores strictly separate from finished creative sign-off.

Automated checks can evaluate raw source photos or rendered exports against your rubric, yet a passing score only reflects the model's probability estimate and can still miss subtle defects. As detailed in Stillmore's pre-delivery creative workflow checks, finishing promotional content requires inspecting five creative dimensions: whether product shape matches the real item, how color and surface texture hold up under new lighting, whether product and offer wording match the brief, whether motion leaves key details readable, and whether composition holds a clear focal point on small screens.

For an illustrative example, an automated rule might flag a dim bottle photo where a finger covers the cap, or verify basic aspect ratio on an exported clip. Even after a passing check, a human creative reviewer still needs to confirm that a composited shadow falls naturally, amber glass keeps its authentic hue, and a weekend offer caption remains readable while the bottle turns on screen.

Review Dimension What Automated QA Evaluates What Finished Creative Review Requires Decision Owner
Product shape Estimates whether supplied outlines meet rubric rules Confirms silhouette and proportions stay true after compositing API builder (scoring) / Creative lead (sign-off)
Color and texture Grades lighting separation on the submitted asset Verifies tint and surface finish match the real item in-scene API builder (scoring) / Post-VFX reviewer (sign-off)
Promotional text Cross-checks supplied brief fields against visible frames Verifies headline spelling, layout hierarchy, and mobile legibility Merchant brief + Pre-delivery reviewer
Motion and pacing Checks supplied frames against predefined timing criteria Shapes edit rhythm, camera moves, and readable hold time Post-production and creative lead
Output delivered Returns probabilistic scores and labels for branching Ships finished images, videos, and social captions Creative team or internal designer

Next Steps: Wiring the Decisions API vs. Delegating Creative Production

Your next step depends on whether your team wants to maintain automated QA scripts plus creative editing in-house, or supply existing photos to a studio that delivers checked, publish-ready creatives.

Running Gemini Decision 1 still leaves rubric maintenance, retouching, motion editing, and caption writing on your team. For independent businesses that prefer skipping API pipelines and production software, how Stillmore turns existing photos into finished social content works as an advertising creative agency: you upload existing product, shop, or team photos once, choose styles in Inspiration or enable ongoing creation, and receive finished images, videos, and captions to publish yourself. Stillmore's photography and live-action, post-production and VFX, and AI production teams may evaluate and use models such as Gemini Decision 1 when they suit a brief, combining them with editing and creative review so you never need to learn production tools.

Consider an illustrative neighborhood café holding 15 everyday phone photos of seasonal drinks and pastries. A team with developer resources can wire Gemini Decision 1 to triage sharp uploads from blurry duplicates before passing files to an editor, whereas a busy café owner working with Stillmore uploads those photos once, selects a launch look or delegates ongoing creation, and publishes the delivered visuals and captions directly.

Conclusion

Gemini Decision 1 gives developers a fast, closed-set endpoint to run noul, choice, and score checks across visual uploads and briefs with confidence-based fallback routing. Whether screening raw phone photos at intake or checking exported drafts before release, probabilistic QA scores only estimate rule fit rather than delivering retouched composites, video pacing, or copy. Decide whether to embed these checks inside an internal production pipeline or hand your existing photos to a dedicated creative team so checked images, videos, and captions arrive ready for your channels.

Resources

Resource Direct Link What It Verifies
Google DeepMind Decisions API Announcement Gemini Decisions API: Fast, calibrated multimodal decisions Official gemini-decision-1 documentation, Python snippet, question types, limits, and billing model
Stillmore Creative Production Workflow Inside Stillmore: Our Creative Production Workflow Photography/live-action, post/VFX, and AI team roles, five pre-delivery checks, and photo detail boundaries
Stillmore Overview and FAQ What Is Stillmore? Content Creation for Small Businesses Reusable photo Library, manual vs. automatic creation modes, Credit rules, and merchant publishing control

Frequently Asked Questions

Can Gemini Decision 1 inspect both photos and short videos in one API call?

Yes, one request handles combined visual files and written briefs so your checks run against a single context. Pairing copy with product imagery lets automated rules compare visible labels against exact SKU names or promotion dates before design work starts. Still, structuring media payloads and maintaining QA logic requires engineering time. Independent businesses using Stillmore skip technical setup altogether: you supply existing product or shop photos and offer details while Stillmore's team handles multi-format image and video production plus pre-delivery review.

Do I need to set up Gemini Decision 1 myself to get pre-checked marketing visuals?

No, configuring the Decisions API is only necessary if your team is building custom automated triage or routing software. Business owners who want finished social posts do not need to write Python scripts, tune rubrics, or learn design software. At Stillmore, you supply existing product, space, or team photos along with your promotion details. Stillmore's creative specialists may evaluate and use models such as Gemini Decision 1 when suitable, while handling compositing, editing, and pre-delivery quality checks before delivering finished visuals and captions.

How do `noul`, `choice`, and `score` questions work together for content QA?

These three primitives separate one asset inspection into hard pass/fail blockers, workflow routing, and quality grading. Use noul checks to catch missing edges or obscured labels, choice labels to sort usable files into editing or reshoot queues, and score rubrics to rank lighting clarity across uploads. That structured triage parallels how creative teams verify deliverables before release, such as Stillmore specialists checking product shape, color and surface texture, accurate text wording, readable motion pacing, and small-screen composition prior to delivery.

What should happen when a Gemini Decision 1 confidence score is ambiguous?

Route any asset returning an uncertain confidence score to a fallback branch or human reviewer instead of auto-approving it. Borderline results commonly occur when glare sits near a label, lighting falls near a threshold, or two routing options overlap. Because probabilistic outputs only estimate fit against your instructions and can misclassify edge cases, manual review prevents flawed uploads from moving forward silently. Likewise, while Stillmore uses software and agentic workflows across production stages, human creative specialists retain final judgment on piece quality.

What photo issues cause intake QA checks to fail before video or poster production?

Intake checks commonly reject phone shots with cut-off sleeves, heavy glare masking printed labels, motion blur, or missing offer context such as end dates. Because retouching cannot reliably reconstruct product details never captured by the camera, zooming into each upload beforehand prevents wasted production rounds. For Stillmore clients, providing clear angles with separate product names and promotion wording helps production proceed smoothly; turning on automatic creation also requires at least 3 suitable photos in Library on Try it or 15 on Growth.

Does passing a Gemini Decision 1 check mean an asset is ready to publish?

No, a passing score only indicates the submitted file met your configured criteria as estimated by the model, which can still overlook subtle defects. Gemini Decision 1 outputs probabilities and routing labels rather than retouching composites, pacing video cuts, or writing promotional copy. Creative editing and human sign-off remain necessary before any marketing post goes live. With Stillmore, the creative team produces and checks finished images, videos, and matching captions before delivery, leaving publishing decisions entirely in your hands.

How is Gemini Decision 1 billed when running many checks on one photo or video?

Each request charges once for your uploaded media regardless of rule count, with extra checks adding only prompt text volume. Even at low token rates, forecasting an internal pipeline still means tracking clip lengths, media sizing, retries, and separate creative labor. Independent businesses wanting predictable deliverable budgets instead of variable token tracking can use Stillmore Plan Credits: one finished video uses 3,000 Credits and one finished image uses 500 Credits, with optional Add-on Credits available when active.

Can Stillmore users combine self-selected campaign styles with automatic creation?

Yes, Stillmore lets you manage and switch between manual Inspiration submissions and automatic creation inside My Creations at any time. Even while automatic mode runs, you can choose a specific Inspiration style and photos for a new launch while letting the team handle ongoing everyday posts. Both modes draw from your shared monthly Plan Credits. On Growth, for example, submitting 2 video requests yourself leaves enough remaining monthly Credits for 4 videos produced by the team, with merchants publishing deliverables themselves.

How long does Stillmore take to produce and check requested images or videos?

Full creative production and pre-delivery checks take days rather than millisecond API scoring. Self-submitted image and video requests are delivered within 5 days once payment, required materials, and your submission are complete. On Growth in automatic mode, your first video arrives within 5 days once at least 15 suitable Library photos, active payment, and sufficient Credits are ready, continuing on a normal 5-day cadence. Late-month uploads may not finish before month-end, though the team tries to catch up when materials arrive late.

Can uploaded product photos be reused across future Stillmore promotions?

Yes, product, space, and team photos stored in your Stillmore Library remain available for both self-selected Inspiration requests and team-arranged creation. Following Stillmore's "Upload once. Stand out forever." model, a clear set of everyday phone photos can power seasonal offers or varied video pacing without scheduling fresh shoots every month. Reuse applies only while stored photos accurately depict what you currently sell, so upload updated shots whenever packaging changes, new items launch, or a new creative layout requires an angle not yet shown.

Does Stillmore post finished content to social accounts or guarantee engagement?

No, Stillmore never posts directly to your social accounts and does not guarantee views, clicks, or sales. Once creative production and pre-delivery review finish, Stillmore delivers completed images, videos, and matching captions so you decide when and where to publish on Instagram, TikTok, or YouTube Shorts. Likewise, passing an automated Gemini Decision 1 QA threshold only reflects the model's estimate against your rubric rather than predicting marketing performance; real engagement still depends on your offer, audience, and timing.

About the author

Lan He

I'm Lan. Years of writing promotional and editorial work in the US, mostly for companies with big audiences and bigger expectations — where a single article might reach hundreds of thousands of readers, and where the good ones kept earning traffic long after the campaign ended. I learned what makes a page survive that: one clear audience, one real question, images that show the actual work, and a structure that keeps people reading. Now I write for small business owners who do everything themselves, answering the questions they actually ask: what's worth posting, which photos are worth taking, and how to tell one story across images, video, and words. I've seen good content carry a small company further than its ad budget ever did.

More about Lan He & all articles ↗

Give them a reasonto stop by.

Send us your phone photos. Get posts ready to share.

See what we can make
Stillmore fashion creative look
All stories ↗