Images IA pour Google Ads : Veo3, Flux, Midjourney

AI adoption in Google Ads creatives exploded in 2025-2026 — across the accounts observed in public Google Ads benchmarks, roughly 35 to 55% of Display, PMax, and Demand Gen image creatives are partially or fully AI-generated in Q1 2026, vs 12-18% in 2024. On the video side, Veo3 (Google, launched Q2 2024 and broadly deployed in 2025) and Sora 2 (OpenAI) run at 18-30% of Google Ads video assets on short 6-15 second formats. The progression is fast but uneven: fashion, beauty, food, and gaming e-commerce often pass 60% AI, while health, finance, premium B2B remain under 25% for compliance reasons.

The real 2026 question is no longer "should we use AI to generate Google Ads creatives" but "how do we industrialize without degrading quality or violating policy." This article compares the 4 dominant tools (Veo3, Flux, Midjourney v7, Imagen 3) across 8 dimensions, delivers prompts by format (square, landscape, portrait, video), details the late-2025 updated Google Ads policy compliance, the 50-variation batch workflow in 1h, and the AI-vs-real-photo A/B measurement. For the broader context of automated formats, see our complete Performance Max 2026 guide. For the RSA AI text pillar, our article on RSA + AI test rotation. To visualize the volume × CTR impact on your account, our CTR calculator with vertical benchmarks returns the result in 2 inputs.

AI images for Google Ads in 2026: where do we stand?

Generative AI for advertising images went from experimental bonus to standard creative workflow brick in 2026. Frontier models (Flux 1.1 Pro, Veo3, Imagen 3, Midjourney v7) now produce magazine-quality photorealistic images in under 30 seconds, at a marginal cost of a few cents per image. What was unthinkable in 2022 (re-imagining the same scene in 50 variations in 1h) has become the new production baseline.

Three structural 2024-2026 changes that flipped adoption:

• Photorealistic quality — end of crude artifacts (6-fingered hands, illegible texts). Flux 1.1 Pro and Imagen 3 produce images indistinguishable from professional photos on 80%+ of simple prompts.
• API costs in free fall — went from ~$0.16 per image in 2023 to $0.03-$0.05 in 2026 on Flux. Allows generating 50-100 variations at negligible marginal cost.
• Operational AI video — Veo3 (Google) and Sora 2 (OpenAI) have been producing 6-15s videos of commercial quality from simple text prompts since 2024-2025. It's the most visible disruption.

Adoption by vertical on aggregated Google Ads benchmarks Q1 2026:

The most effective AI image use cases on Google Ads:

• Rapid variations of a hero visual — tested across 20-30 variations in 2h, you identify the best-performing version.
• Custom backgrounds for e-commerce products — instead of a neutral studio backdrop, you generate varied use contexts.
• Multiple formats from 1 visual — square, landscape, portrait generated in parallel.
• Fast seasonality — weekly visual refresh with seasonal context (Black Friday, Christmas, summer).
• Multi-language / multi-country — local adaptations (clothing, food, landscapes) without per-market shoots.
• Concept boards — explore 50 creative directions in 1h before producing.

Use cases where AI is still limited:

• Depicting physical products with precise specific features (faithful material textures).
• Identifiable personalities (banned for most without consent).
• Complex multi-character scenes with natural interactions.
• Ultra-distinctive brand voice (luxury, high-end lifestyle).
• Nuanced cultural contexts (religious rituals, precise regional dress codes).

Official Google Ads reference on advertising image policies: the Google Ads Image Requirements guide. For Flux 1.1 Pro documentation on the Black Forest Labs side: the official Flux 1 announcement. Both sources should be reviewed before scaling an AI workflow into Google Ads production.

Tool comparison: Veo3, Flux, Midjourney, Imagen 3

The 4 dominant 2026 tools have specific strengths that suit different use cases. No "best tool in absolute terms" — the right approach is multi-tool by need of the moment.

Practical reading of the comparison:

Veo3 (Google) — for Google Ads video. Launched in 2024, generally available mid-2025, it's the leading AI video tool for the Google Ads context. Photorealistic quality, typical 6-15 second duration (up to 60s in beta), 720p-1080p resolutions, native audio included since late 2025. Native YouTube and PMax integration via Google Cloud Vertex AI. High cost ($0.10-$0.30/second), but unbeatable ROI if you broadcast heavily on YouTube + Demand Gen. Limitation: no ultra-complex camera motion, no 100% fidelity to real products (always with ~5-10% drift on details).

Flux 1.1 Pro (Black Forest Labs) — for industrial batch production. Launched late 2024 by the ex-Stable Diffusion team, became the 2026 standard for automated workflows. Excellent photorealistic quality, best-in-class complex prompt adherence, very competitive API cost ($0.03-$0.05/image). Available via direct BFL API or via replicate.com. No video support. Recommended for: rapid hero product variations, custom backgrounds, multiple square/landscape/portrait formats.

Midjourney v7 — for stylized creativity and moodboarding. Still the reference on creative quality and 2026 style, but its Discord-based nature (no reliable public API) limits industrialization. Plans $26-$105/month for manual usage. Excellent for: pre-campaign strategic moodboards, exploring non-photorealistic creative directions, specific styles (illustration, anime, painting). Limited for: automated batch production, pipeline integration.

Imagen 3 (Google) — for production via Google Cloud. The image counterpart to Veo3 on the Google side. Solid photorealistic quality, native Vertex AI and Google Cloud Platform integration. Competitive cost ($0.04-$0.08/image). Recommended if you're already deployed on Google Cloud (BigQuery data warehouse, other services). Less flexible than Flux on ultra-complex prompts or atypical styles, but easier to integrate in a Google stack.

Recommendation by profile:

• Multi-account agency / industrialization — Flux 1.1 Pro for batch + Veo3 for video + Midjourney for moodboards.
• Direct mid-market e-commerce advertiser — Flux 1.1 Pro for production + Midjourney for exploration.
• Direct advertiser on Google Cloud stack — Imagen 3 + Veo3 (native integration).
• Premium creative studio — Midjourney for moodboards + Flux or Imagen 3 for final production + Veo3 for video.
• Small advertiser / SMB — Flux 1.1 Pro via replicate.com (pay-per-use, no commitment).

Prompts by format: square, landscape, portrait, video

Each Google Ads format (square 1:1, landscape 16:9, portrait 9:16, video 6s) calls for a different prompt structure. The prompt that works in square doesn't work in portrait — composition, focal point, breathing room are radically different.

Prompt 1 — Square 1:1 e-commerce hero product (Flux 1.1 Pro)

{
  "model": "flux-1.1-pro",
  "prompt": "Ultra realistic product photography of a white leather sneaker, side angle 3/4 view, soft natural daylight from upper left, neutral cream background with subtle gradient, slight depth of field, premium magazine quality, centered composition with product occupying 65% of frame, breathing room top and bottom",
  "aspect_ratio": "1:1",
  "resolution": "1024x1024",
  "negative_prompt": "blurry, distorted, low quality, watermark, text overlay, hands, people, cartoon, illustration",
  "seed": "random",
  "guidance_scale": 4.5
}

Prompt 2 — Landscape 16:9 desktop Display banner (Flux 1.1 Pro)

{
  "model": "flux-1.1-pro",
  "prompt": "Cinematic interior office scene, modern professional workspace with laptop showing dashboard analytics, soft morning light through window, warm wood and beige tones, shallow depth of field with background blur, focus on left third of frame leaving right two-thirds for ad copy overlay, lifestyle business photography style, no people visible, premium B2B SaaS aesthetic",
  "aspect_ratio": "16:9",
  "resolution": "1920x1080",
  "negative_prompt": "people, faces, text, logos, cluttered, dark, oversaturated",
  "seed": "random",
  "guidance_scale": 4.0
}

Prompt 3 — Portrait 9:16 vertical mobile Stories (Flux 1.1 Pro)

{
  "model": "flux-1.1-pro",
  "prompt": "Vertical mobile-first composition, smartphone screen showing food delivery app interface in foreground bottom third, hands holding phone partially visible, blurred restaurant ambiance background top two-thirds with warm lighting, golden hour quality, lifestyle photography, casual urban setting, optimized for vertical 9:16 mobile viewing with main subject in lower 40% of frame",
  "aspect_ratio": "9:16",
  "resolution": "1080x1920",
  "negative_prompt": "horizontal composition, faces clearly visible, text, watermark, cluttered",
  "seed": "random",
  "guidance_scale": 4.5
}

Prompt 4 — 6-second video Demand Gen (Veo3)

{
  "model": "veo-3",
  "prompt": "6-second cinematic shot, slow camera push-in toward a modern coffee cup on a wooden cafe table, steam gently rising, warm morning light streaming from window left, shallow depth of field with bokeh background of blurred cafe interior, photoreal quality, no people, calm atmosphere, smooth camera motion no jerks",
  "duration_seconds": 6,
  "aspect_ratio": "16:9",
  "resolution": "1080p",
  "audio": "ambient cafe sounds, low volume, no music",
  "motion_intensity": "subtle"
}

The 6 AI photo prompt rules for Google Ads (validated in-account):

1. Always specify the format right in the prompt — square / landscape / portrait / video. Don't let the model guess.
2. State the composition focal point — where the subject sits in the frame, where overlay text space should be.
3. Describe the light explicitly — natural daylight, golden hour, studio softbox. Avoids inconsistent lighting.
4. Use the negative prompt aggressively — eliminate artifacts upstream (blurry, distorted, watermark, text, hands, faces).
5. Reference photography style — magazine quality, lifestyle photography, product photography. Frames the rendering.
6. No photorealistic people unless necessary — compliance risk + quality drift (hands, expressions).

Technical generation workflow (Python pseudocode):

# Pseudo-code batch image generation workflow Flux 1.1 Pro
import requests
import os
from concurrent.futures import ThreadPoolExecutor

BFL_API_KEY = os.environ["BFL_API_KEY"]
ENDPOINT = "https://api.bfl.ml/v1/flux-1.1-pro/generate"

def generate_image(prompt_config):
    response = requests.post(
        ENDPOINT,
        headers={"Authorization": f"Bearer {BFL_API_KEY}"},
        json=prompt_config
    )
    return response.json()["image_url"]

def batch_generate(prompts_list, max_parallel=10):
    with ThreadPoolExecutor(max_workers=max_parallel) as executor:
        urls = list(executor.map(generate_image, prompts_list))
    return urls

# Generate 50 variations of hero product
variations = [
    {**base_prompt, "seed": i, "guidance_scale": 4.5}
    for i in range(50)
]
image_urls = batch_generate(variations)
# Total time : ~3-5 min for 50 images
# Total cost : ~$1.65-$2.75 for 50 images

Google Ads compliance: what's banned in AI imagery

Google Ads accepts AI visuals in 2026 but enforces several strict restrictions updated in late 2025. Violating these rules = accounts suspended without notice, which we see regularly in audits. Compliance is not optional.

Strict 2026 bans (immediate account suspension):

• Photorealistic human faces representing specific people without consent — what the community calls "deepfake-adjacent." Banned even if the person isn't named. If the generated face resembles a celebrity or an identifiable real person, the asset is rejected.
• Photorealistic depictions of children — near-total ban in 2026, except very specific business cases (toys, education) with explicit disclaimer.
• Misleading deceptive images — product depicted differently from reality (e.g., showing a product with features it doesn't have, non-realistic before/after results in health/beauty).
• Sensitive imagery in regulated verticals — health, finance, gambling, elections — without explicit disclaimer and without prior Google Ads approval.
• Violent, shocking, sexualized imagery — standard Google Ads rules, applied strictly to AI as well.
• Religious or ethnic symbols used commercially without context — rule expanded in 2025.

Practical 2026 recommendations:

• Tag AI images in metadata — Google Ads Asset description fields, "Generated with AI" mention or internal code.
• Light visual disclaimer on certain formats — not universally mandatory, but recommended in health/finance.
• Non-visible watermark (C2PA metadata) — Adobe and Google have been pushing the C2PA standard since 2024 to trace AI provenance. Flux 1.1 Pro and Imagen 3 include C2PA metadata by default.
• Quarterly audit of AI assets in production — verify no asset has become non-compliant after a policy update.

Additional precautions for industrializing agencies:

• No photorealistic people generation by default unless legally validated business need.
• Legal/compliance review process on all AI assets before production, especially in health/finance.
• Internal documentation of prompts and models used, for audit traceability.
• Explicit consent if the campaign depicts employees, clients, partners (even if AI-generated and non-photorealistic).
• Monthly policy watch on Google Ads + IAB updates + C2PA standards.

For advertisers who want a complete Google Ads account audit including AI asset compliance, our Google Ads audit checklist details the 47 control points, including the AI compliance section since late 2025.

Batch workflow: generate 50 variations in 1h

The batch workflow is the AI image superpower for Google Ads — producing 50 variations of a hero visual in 1h, which would have taken 1-2 weeks in photographer + studio + post-prod. The clean method requires structural discipline, otherwise you accumulate unusable AI waste.

6-step batch workflow (aggregated Google Ads benchmarks, observed agencies):

# Pseudo-code RSA Google Ads visual batch workflow
def batch_workflow_pmax_assets(brand_brief, target_count=50):
    # Step 1: prompt template
    base_prompt = build_prompt_template(
        vertical=brand_brief["vertical"],
        product=brand_brief["product"],
        brand_voice=brand_brief["voice"],
        format_target="multi"  # square + landscape + portrait
    )

    # Step 2: generate 30 variant prompts
    prompts = generate_prompt_variants(
        base=base_prompt,
        variations_axes=["composition", "lighting", "background", "mood"],
        count=30
    )

    # Step 3: parallel batch image generation
    raw_images = batch_generate_parallel(prompts, max_parallel=10)
    # ~30 images in 3-5 min

    # Step 4: algorithmic scoring (technical quality)
    scored_images = score_images_algorithmic(raw_images, criteria=[
        "no_artifacts_detected",
        "composition_centered",
        "no_text_overlay_existing",
        "color_palette_match_brand",
        "resolution_target_met"
    ])
    # ~22 images pass the algorithmic filter

    # Step 5: human review (subjective quality)
    human_validated = human_review(scored_images, criteria=[
        "brand_voice_match",
        "message_market_fit",
        "no_compliance_issue"
    ])
    # ~15-18 images validated

    # Step 6: format derivation
    final_assets = expand_to_formats(human_validated, formats=["1:1", "16:9", "9:16"])
    # ~45-54 final multi-format assets

    return final_assets

Time + cost breakdown (real workflow on a fashion e-commerce account):

For reference — equivalent traditional photographer commission:

• $2,000-$5,000 for 50 product visuals + retouching.
• 5-10 day lead time (studio + post-prod).
• 1-2 formats max per visual (need reshoot to adapt).
• Premium brand quality guaranteed (but not always needed in digital paid).

The cost-time ratio is around 1:50 to 1:200 in favor of AI for digital paid use cases. But beware: this ratio applies only to cases where AI quality is sufficient (Display, Demand Gen, PMax). For premium brand visuals that will end up in print + retail, the photographer remains relevant.

Industrial batch workflow recommendations:

• Prompt templating — create 5-10 reusable prompt templates per vertical (fashion e-com, food, B2B, gaming, etc.).
• Automated pipeline — n8n, Zapier, or custom Python script to orchestrate steps (cf. n8n Google Ads).
• Validated asset library — cloud storage with metadata (vertical, format, observed perf) for reuse.
• Systematic human review — never direct AI upload without human review. It's the most frequent mistake.
• A/B test all new series — see section 6.

A/B test: AI quality vs real photo

AI vs real-photo A/B testing is the only serious way to measure if AI quality is sufficient for your specific campaigns. No universal answer — it depends on vertical, format, brand maturity.

AI vs photo image A/B test methodology (Google Ads data, n=42 tests 2025-2026):

• Setup — 2 isolated ad groups in the same campaign, same keywords, same budget. One with AI assets, the other with photo assets.
• Duration — minimum 21 days, minimum 5,000 impressions per ad group.
• Metrics — CTR, conversion rate, CPA, ROAS if applicable.
• Variable preservation — identical bid strategy, identical geo, identical audience signals.

Average results observed on aggregated Google Ads benchmarks Q1 2026:

Reading the results:

• AI wins clearly in verticals where creativity and visual freshness rule (gaming, food, mass-market e-com).
• AI is equivalent in mass-market verticals with low-differentiation brand voice (mass B2B SaaS, lead gen, industrial).
• AI loses in verticals where visual trust and premium brand voice rule (health, finance, luxury, premium B2B).
• The productivity gain is universal — even when AI performs slightly less well, production time -50 to -80% frees up budget for other optimizations.

Practical 2026 recommendation:

• AI-friendly verticals (mass e-com, food, gaming, lead gen) — AI by default on 60-80% of assets, quarterly A/B test to validate.
• Mixed verticals (B2B SaaS, industrial) — AI for fast variations + photo for hero brand. 50/50 ratio.
• Trust-critical verticals (health, finance, luxury) — photo by default, AI as exploration complement only.
• Always A/B test — no generalization without measurement on your specific account.

Typical mistakes and mitigation strategies

On AI image Google Ads workflows referenced in 2025-2026, here are the 7 recurring mistakes — each one reduces real AI ROI and explains why some advertisers wrongly conclude that "AI doesn't work for our vertical." Systematic mitigation on each.

Mistake 1 — AI upload without human review. The most frequent and most expensive mistake. AI regularly produces technically correct but weird outputs (subtly 6-fingered hands, strange expressions, incorrect product details). Mitigation: systematic human review before each upload, score 0-3 on 4 criteria (technical quality, compliance, brand voice, message-market).

Mistake 2 — No format respect right in the prompt. Generating in 1024x1024 then cropping to portrait 9:16 degrades composition. Mitigation: generate in target format from the start via explicit aspect_ratio, or use format-specific prompt templates.

Mistake 3 — Photorealistic faces without checking compliance. Account suspension risk. Mitigation: no photorealistic faces by default unless legally validated business need + consent + C2PA watermark.

Mistake 4 — AI over-use in trust-critical verticals. Health, finance, luxury, premium B2B: real photo remains superior for visual trust. Mitigation: systematic A/B test before industrialization, AI/photo ratios adapted by vertical.

Mistake 5 — No A/B test between AI series. Industrializing on the basis of the first series without measuring its performance vs alternatives. Mitigation: quarterly A/B test at minimum, ideally each new 50-visual series.

Mistake 6 — Non-industrialized workflow (ad-hoc manual generation). Multiplies time by 3-5x vs automated batch workflow. Mitigation: n8n / Zapier / custom Python pipeline to orchestrate generation, scoring, format derivation. Invest 2-3 dev days once to save hundreds of hours over 12 months.

Mistake 7 — No validated asset library. Re-generating per campaign instead of reusing validated assets. Mitigation: cloud storage organized by vertical / format / observed perf, metadata on each asset, easy search for reuse.

Bonus — 3 mistakes specific to Veo3 / Sora 2 AI video:

• Camera motion too complex in the prompt (e.g., drone shot with 3 angle changes in 6 seconds). Veo3 produces artifacts. Mitigation: simple prompts, 1 camera motion per video.
• No audio coherent with the scene. Veo3 sometimes adds random audio if not specified. Mitigation: explicitly specify desired audio or request silence.
• 15s duration on overly dynamic scenes. Frame-to-frame coherence degrades after 10s. Mitigation: prefer 6-8s for dynamic scenes, keep 15s for calm scenes.

For advertisers who want to industrialize the AI image workflow without building the batch / scoring / A/B infrastructure yourself, our SteerAds audit integrates the workflow above and proposes an AI industrialization plan segmented by vertical and asset criticality, with a pilot A/B test on 1-2 ad groups before rollout. To go further on the AI Google Ads pillar, see our complementary articles 30 JSON Google Ads prompts and RSA + AI test rotation. AI image generation for Google Ads in 2026 is neither magical nor useless — it's a new productivity lever that transforms the frequency and quantity of visual tests possible. Well industrialized with compliance and A/B test, it frees 50-80% of creative production time while maintaining or improving performance in AI-friendly verticals. Poorly industrialized, it's a trap of apparent productivity that causes account suspensions and brand voice degradations. Methodological discipline makes all the difference — see also the official Google Ads documentation for more details.

To go further, see also our guides on AI negative keywords discovery clustering, Python API automation, Zapier Make Google Ads.

Sources

Official sources consulted for this guide:

• modelcontextprotocol.io
• schema.org
• blog.google
• anthropic.com

FAQ