I've spent the last year shipping AI-generated ads at scale. Here's what actually works.
I'm a Software Engineer at KeyValue — we provide engineering support for product companies. For the past year, I've been working with Pencil, an AI-powered ad generation platform that helps brands create static social ads at scale. My focus has been on the image generation pipeline and agentic AI systems — the intelligence layer that lets users interact with image and video tools through natural conversation.
Our users upload brand assets, describe what they want, and expect production-ready creatives back. Not "impressive for AI" — actually ready to run as paid media.
This post recaps the major image generation models I used this year, with honest opinions forged in production, not playground demos.
A year ago, our workflow was duct tape and prayer. We ran ComfyUI pipelines stitching together Imagen, Bria, and self-hosted Stable Diffusion. Every job required careful prompt engineering — if we wanted brand-appropriate output, we had to manually encode brand guidelines into the prompt. Relighting was hit-or-miss. Complex edits meant complex workflows. And users kept filing the same tickets:
The shift to multimodal models changed everything. We can now provide context directly — reference images, brand guidelines, previous iterations — and the model understands intent rather than just pattern-matching keywords. Quality and precision editing have both improved dramatically. Images that pass the "doesn't look AI-generated" test are now the norm rather than the exception.
Working with ad creatives over the past year, the same requests kept surfacing. Users didn't ask for "better diffusion models" — they asked for specific things:
So that's what we tested. Six scenarios, each tied to a real workflow need:
| Test | What It Measures | Why it Matters |
|---|---|---|
| Text Rendering | Accuracy of headlines, taglines | Menus, product labels, CTAs — text errors kill ads with a lot of postprocess |
| Surgical Edit | Change one thing without breaking the rest | Iterative refinement is the core of ad creative work |
| Character Consistency | Same person across multiple scenes | Campaign continuity, brand mascots, storytelling |
| Localization | Translate text while preserving layout | Global campaigns need localized assets fast |
| Real-Time Grounding | Factually accurate data in infographics | Can't ship hallucinated statistics |
| Photorealism | Does it look like a photograph or a render? | "Looks AI-generated" is a death sentence for ads |
Below is a hands-on comparison of leading image generation models for product teams, evaluated for editing precision, text rendering, character consistency, and real-world deployment fit.
Imagen 4 is our volume play. When users need to explore ten directions before committing to one — testing concepts, iterating on layouts, generating options for client review — waiting 30 seconds per image kills momentum. Imagen 4 Fast delivers up to 10x faster generation, which means we can show options quickly and refine the winners with more precise tools.
The speed gets appreciation, but the limits surface fast.
Prompt: A Viking longship cutting through rough, dark ocean waves, shot from a low, slightly off-center angle near the waterline. Jagged mountains loom in the distance under a stormy sky. Forked lightning illuminates the clouds behind the ship, creating dramatic rim light along the sails and hull. Cold rain and sea mist hang in the air, with volumetric lighting visible through the fog. The scene is photorealistic and high-resolution, captured in the style of a Hasselblad X2D 100C, shallow depth of field with crisp foreground detail and softly fading background. HDR processing enhances contrast and texture while preserving a moody, atmospheric look.
Result: Imagen 4 produced quality output with clean rendering, though the aesthetic leaned more to "digital illustration" than true photograph. Good for concept art and rapid ideation.
Prompt: Create a vintage coffee shop menu poster with the heading "MORNING RITUALS" in decorative serif font. Include 6 drink items with names and prices: Espresso $3.50, Cortado $4.25, Flat White $5.00, Pour Over $4.75, Cold Brew $5.50, Matcha Latte $6.00. Add a tagline at the bottom: "Crafted with care since 1987." Use warm brown tones and a weathered paper texture.
Result: Near-perfect accuracy — only a minor punctuation omission (missing period). Production-usable with minimal cleanup.
Prompt: Create an infographic showing the current top 5 tallest buildings in the world with their names, locations, heights in meters, and year completed. Use a clean modern design with building silhouettes arranged by height. Include a title "REACHING THE SKY" at the top.
Result: Failed entirely. Generated fictional building names and physically impossible heights. Diffusion-only models don't have real-world knowledge access.
Since Imagen 4 doesn't support image editing, we couldn't test surgical edits, character consistency, or localization.
Best for: Fast photorealistic generation, texture-heavy imagery, rapid creative exploration.
Nano Banana Pro solved three problems that kept hitting our queue: garbled multilingual text, fake-looking complex scenes, and hallucinated data. It's become our go-to for anything that needs to feel like a real photograph — and for any content where facts matter.
The multilingual accuracy changed what's possible for global teams.
But iteration has limits.
Prompt: Create a vintage coffee shop menu poster with the heading "MORNING RITUALS" in decorative serif font. Include 6 drink items with names and prices: Espresso $3.50, Cortado $4.25, Flat White $5.00, Pour Over $4.75, Cold Brew $5.50, Matcha Latte $6.00. Add a tagline at the bottom: "Crafted with care since 1987." Use warm brown tones and a weathered paper texture.
Result: 100% accuracy. Every drink name, price, and tagline rendered exactly as specified. Production-ready.
Prompt: Create an infographic showing the current top 5 tallest buildings in the world with their names, locations, heights in meters, and year completed. Use a clean modern design with building silhouettes arranged by height. Include a title "REACHING THE SKY" at the top.
Result: The only model to produce a factually accurate result. Correctly identified all five buildings with precise heights, locations, and completion dates — including Merdeka 118, which only became the world's second-tallest in 2023. Google Search grounding working exactly as intended.
Base character
Result: Excellent. Same woman recognizable across Paris, Tokyo, and library scenes. Supports up to 14 reference images and maintains consistency across up to 5 different people simultaneously.
Base image
Result: Near-perfect Japanese. Serviceable Hindi — some transliteration issues but readable and layout preserved perfectly.
Prompt: A Viking longship cutting through rough, dark ocean waves, shot from a low, slightly off-center angle near the waterline. Jagged mountains loom in the distance under a stormy sky. Forked lightning illuminates the clouds behind the ship, creating dramatic rim light along the sails and hull. Cold rain and sea mist hang in the air, with volumetric lighting visible through the fog. The scene is photorealistic and high-resolution, captured in the style of a Hasselblad X2D 100C, shallow depth of field with crisp foreground detail and softly fading background. HDR processing enhances contrast and texture while preserving a moody, atmospheric look.
Result: Most authentically "photographic" result. Volumetric god rays, muted color palette, atmospheric layering closest to actual Hasselblad medium format output.
Result: Successfully changed the strap color but introduced minor collateral changes to lighting and strap curvature. Good, not pixel-perfect.
You can read more about our hands-on experience with the Nano Banana Pro here.
Best for: Photorealism, multilingual assets, factual infographics, complex scenes.
FLUX entered our toolkit to solve two problems other models couldn't: lighting precision and product consistency. I've been experimenting with it primarily for relighting workflows — earlier attempts with open-source models had limited precision; shadows fell wrong, light color didn't match the scene. FLUX fixed that. And with FLUX.1 Kontext and quick LoRA training, we can lock product identity across dozens of ad variants.
The precision is what users notice first.
But the compute requirements mean it's not for every use case.
Result: FLUX 2 Max delivered the most precise edit — changed only the strap color while preserving marble veining, shadow direction, and exact watch positioning. Pixel-level precision ideal for e-commerce variants.
Base character
Result: Exceptional. Unmistakably the same person across all three scenes — hair, glasses, coat texture, even earring placement matched. FLUX supports up to 10 reference images and officially claims "state-of-the-art character consistency."
Prompt: A Viking longship cutting through rough, dark ocean waves, shot from a low, slightly off-center angle near the waterline. Jagged mountains loom in the distance under a stormy sky. Forked lightning illuminates the clouds behind the ship, creating dramatic rim light along the sails and hull. Cold rain and sea mist hang in the air, with volumetric lighting visible through the fog. The scene is photorealistic and high-resolution, captured in the style of a Hasselblad X2D 100C, shallow depth of field with crisp foreground detail and softly fading background. HDR processing enhances contrast and texture while preserving a moody, atmospheric look.
Result: Most technically complete image. Every prompt element rendered with high fidelity — water spray physics, wood grain texture, atmospheric depth.
Base image
Result: Japanese was adequate. Hindi failed — produced gibberish despite preserving layout. Stronger training on CJK scripts than Indic scripts.
Prompt: Create a vintage coffee shop menu poster with the heading "MORNING RITUALS" in decorative serif font. Include 6 drink items with names and prices: Espresso $3.50, Cortado $4.25, Flat White $5.00, Pour Over $4.75, Cold Brew $5.50, Matcha Latte $6.00. Add a tagline at the bottom: "Crafted with care since 1987." Use warm brown tones and a weathered paper texture.
Result: 100% accuracy. Production-ready.
Best for: Relighting, self-hosted deployment, multi-reference generation, brand-exact color matching.
GPT Image 1.5 finally gave us surgical precision. When users ask to change a product color, only the color changes. Adjust a headline, and the product stays put. The iterative editing loop — describe what's wrong, get a fix, refine further — finally works the way users expect.
The precision is what users notice immediately.
But multilingual work remains a gap.
Instruction following — Superior understanding of intent; natural iterative workflow
Text rendering — Strong on English headlines, CTAs, product labels
Result: Excellent. Executed the color swap cleanly with only subtle recomposition — may or may not matter depending on use case, but the edit itself was precise.
Base image
Result: Best performance of all models. Grammatically correct, culturally appropriate translations in both languages while perfectly preserving original layout, characters, and visual hierarchy. The model understood "replace text" not "regenerate image."
Prompt: Create a vintage coffee shop menu poster with the heading "MORNING RITUALS" in decorative serif font. Include 6 drink items with names and prices: Espresso $3.50, Cortado $4.25, Flat White $5.00, Pour Over $4.75, Cold Brew $5.50, Matcha Latte $6.00. Add a tagline at the bottom: "Crafted with care since 1987." Use warm brown tones and a weathered paper texture.
Result: 100% accuracy. Production-ready.
Base character
Result: Very good. Recognizable as the same character with minor expression variations that don't break identity. Supports up to 16 input images with the first 5 preserved at higher fidelity.
Prompt: A Viking longship cutting through rough, dark ocean waves, shot from a low, slightly off-center angle near the waterline. Jagged mountains loom in the distance under a stormy sky. Forked lightning illuminates the clouds behind the ship, creating dramatic rim light along the sails and hull. Cold rain and sea mist hang in the air, with volumetric lighting visible through the fog. The scene is photorealistic and high-resolution, captured in the style of a Hasselblad X2D 100C, shallow depth of field with crisp foreground detail and softly fading background. HDR processing enhances contrast and texture while preserving a moody, atmospheric look.
Result: Matched FLUX in detail while adding narrative energy — visible crew, dramatic action composition. High quality.
Prompt: Create an infographic showing the current top 5 tallest buildings in the world with their names, locations, heights in meters, and year completed. Use a clean modern design with building silhouettes arranged by height. Include a title "REACHING THE SKY" at the top.
Result: Clean visuals, mostly accurate data, but relied on pre-2023 training knowledge — missed Merdeka 118. No real-time grounding capability.
Best for: Iterative editing, precise instruction following, localization workflows.
This capability matrix compares modern image generation models across resolution, text rendering, editing precision, character consistency, and ecosystem maturity.
| Capability | GPT Image 1.5 | Nano Banana Pro | Imagen 4 | FLUX.2 [max] |
|---|---|---|---|---|
| Max Resolution | 1536×1024 (landscape) or 1024×1536 (portrait) | 2K native / 4K upscaled | 2K | 4MP |
| Text Rendering | Strong | Best-in-class | Strong | Excellent |
| Multi-Reference Input | Upto 16 (First 5 input images preserved with higher fidelity) | Up to 14 images | Limited | Up to 10 images |
| Real-Time Grounding | No | Yes (Google Search) | No | Yes (web context) |
| Inpainting / Fill | Excellent | Strong | No, check predecessors for this feature | Excellent |
| Character Consistency | Strong | Strong (up to 5 people) | Moderate | Strongest |
| Multilingual Text | Good | Excellent | Good | Good |
| Edit Precision | Best-in-class | Strong | Moderate | Strong |
| Speed | 4x faster than predecessor | Standard | Up to 10x (Fast mode) | Standard |
| Open Weights Available | No | No | No | Yes (FLUX.2 [dev]) |
| Ecosystem | ChatGPT, API | Gemini, Workspace, Adobe, Figma | Vertex AI, Gemini API, Workspace | API, Replicate, fal.ai, self-hosted |
There's no single 'best' image generation model in 2025 — only the best fit for your specific workflow. If your priority is iterative editing precision, GPT Image 1.5 delivers. If you need multilingual marketing assets with accurate text, Nano Banana Pro is unmatched. For raw photorealism and speed, Imagen 4 Fast outpaces the field. When brand-exact colors and character consistency matter most, FLUX.2 [max] offers the control. Pick based on what you're building, not on leaderboard rankings alone.
Models like GPT Image 1.5, Nano Banana Pro, and FLUX.2 max are production-viable because they support reference images, consistent edits, and predictable outputs. Diffusion-only models like Imagen 4 remain better suited for exploration rather than shipping workflows.
Nano Banana Pro stands out because it grounds outputs in Google Search data and renders multilingual text accurately. Models without real-time grounding often hallucinate statistics or introduce spelling errors, making them risky for factual visuals.
Multi-reference input is essential when the same person or product appears across multiple scenes. Models like FLUX.2 max, Nano Banana Pro, and GPT Image 1.5 are architected to preserve identity, while pure diffusion models regenerate identity every time.
Not entirely, but they reduce reliance on fragile prompts. Multimodal models understand context through reference images, prior iterations, and brand inputs, shifting effort from prompt hacking to creative decision-making.
