What it’s not<\/strong> a replacement for: final brand assets that need pixel-perfect logo placement, legally vetted product photography, or any image destined for print. It’s a drafting and concepting tool, and it works best when you treat it that way.<\/p>\n\n\n\n GPT Image 2 is live in ChatGPT for Plus, Team, and Enterprise subscribers. For API access \u2014 which you’ll want if you’re generating variants at volume \u2014 it’s available through OpenAI’s API. Worth checking the OpenAI image generation docs<\/a> before you start, because the size parameters changed with this release.<\/p>\n\n\n\n The model supports flexible aspect ratios now, which is a genuine upgrade. For ad work, you’re mainly working with:<\/p>\n\n\n\n You can specify these directly in the prompt or via the size parameter in the API. Get this right before you generate anything \u2014 resizing after the fact doesn’t recompose the image, it just crops it.<\/p>\n\n\n\n Don’t skip this step. GPT Image 2 follows complex instructions well, but “well” is relative to the clarity of what you give it.<\/p>\n\n\n\n Before you touch the prompt field, have these ready:<\/p>\n\n\n\n If you have a reference image \u2014 a previous ad that performed well, a mood board screenshot \u2014 you can upload it. GPT Image 2’s instruction-following for image editing is noticeably better than its predecessor.<\/p>\n\n\n\n Prompt structure matters more than prompt length. The GPT image generation prompting guide<\/a> recommends going: background\/scene \u2192 subject \u2192 key details \u2192 constraints \u2014 and adding the intended use so the model calibrates its “mode” of output.<\/p>\n\n\n\n Here’s what that looks like in practice for a UGC-style skincare ad:<\/p>\n\n\n\n “Lifestyle product photo, warm natural light bathroom counter, a woman in her 30s applying serum, mid-shot. Product bottle front and center in foreground, label legible. Text overlay at bottom: ‘First Month FREE \u2014 Use Code GLOW25’. Clean sans-serif font, white text on soft dark gradient bar. Intended use: Instagram Stories ad, 9:16 ratio, UGC-style, warm and relatable, not polished or stock-photo-looking.”<\/em><\/p>\n<\/blockquote>\n\n\n\n A few things that tripped me up early:<\/p>\n\n\n\n This is where the workflow starts to pay off. Once you have a first concept you like, you can chat your way to variants.<\/p>\n\n\n\n “Recreate this at 1:1 for Instagram feed, keeping the same composition and copy.”<\/p>\n\n\n\n “Give me a version where the CTA says ‘Try It Free’ instead of ‘First Month FREE’.”<\/p>\n\n\n\n “Make a version with a darker background \u2014 the product is getting lost.”<\/p>\n\n\n\n Each turn preserves context from the previous generation, so you’re not starting cold. Honestly, this is the part I didn’t expect to work as well as it did. The composition held up across format changes better than I thought it would \u2014 it wasn’t just cropping, it was actually recomposing for the new aspect ratio most of the time.<\/p>\n\n\n\n That said: always compare variants side by side before committing. Sometimes an element shifts in ways that aren’t immediately obvious \u2014 the text might end up in a different position, or a prop disappears.<\/p>\n\n\n\n Go through the checklist you’d use for any ad creative before it leaves your hands:<\/p>\n\n\n\n Text accuracy first.<\/strong> Read every word in the image. GPT Image 2 is much better than previous models, but it’s not infallible \u2014 especially with longer strings, stylized fonts, or text that bleeds over a busy background. If the CTA is wrong, the ad is wrong.<\/p>\n\n\n\n Brand elements.<\/strong> If you fed the model a color palette, verify it landed. If you have a logo, you added it separately in post \u2014 GPT Image 2 will approximate, not replicate.<\/p>\n\n\n\n Faces and people.<\/strong> If there’s a person in the image, look at their hands. Still a weak point for most image models. Zoom in before you move on.<\/p>\n\n\n\n Platform compliance.<\/strong> Different ad platforms have different rules about text coverage ratios, prohibited categories, and disclosure requirements. The model doesn’t know your platform’s current ad policies \u2014 that’s your job to verify.<\/p>\n\n\n\n If something’s off, describe the fix conversationally. “The price text is cut off on the right \u2014 move it to be fully within the safe zone.” That usually works in one or two follow-ups.<\/p>\n\n\n\n My honest output after testing this for a full campaign brief: GPT Image 2 gets you to a presentable first draft faster than any other tool I’ve used. Not a final asset \u2014 a first draft. The gap between “AI draft” and “actually published” still exists, and it varies by campaign.<\/p>\n\n\n\n Where it genuinely helped:<\/p>\n\n\n\n Where I still needed human cleanup:<\/p>\n\n\n\n On that last point: don’t skip it. OpenAI’s usage policies<\/a> are clear that compliance with applicable laws is on you, not the platform. For ad work, this means you’re responsible for ensuring the creative doesn’t infringe third-party trademarks or copyrights \u2014 and as one legal analysis of AI-generated image copyright risks points out, the AI tool granting you access to generate something is not the same as that output being cleared for commercial use. Run a reverse image check on anything high-visibility before it goes live.<\/p>\n\n\n\n Platform ad policies are a separate layer \u2014 Meta, TikTok, and Google all have their own content review systems that may flag or reject AI-generated creative. This is not a reason to avoid the tool; it’s a reason to build a review step into your workflow.<\/p>\n\n\n\n![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/04\/image-277-1024x464.png\")
<\/figure>\n\n\n\nWhat You Need Before You Start<\/h2>\n\n\n\n
Account, Access, and Aspect Ratio Setup<\/h3>\n\n\n\n
\n
Creative Brief and Brand Inputs<\/h3>\n\n\n\n
\n
![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/04\/image-276-1024x347.png\")
<\/figure>\n<\/figure>\n\n\n\nStep-by-Step Workflow for Ad Creative Generation<\/h2>\n\n\n\n
Generate the First Concept<\/h3>\n\n\n\n
\n
\n
![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/04\/image-275-1024x552.png\")
<\/figure>\n<\/figure>\n\n\n\nCreate Size and Copy Variants<\/h3>\n\n\n\n
Review Outputs and Fix Misses<\/h3>\n\n\n\n
![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/04\/image-274-1024x362.png\")
<\/figure>\n<\/figure>\n\n\n\nResults, Limits, and Compliance Risks<\/h2>\n\n\n\n
\n
\n
GPT Image 2 vs Other Ad Creative Tools<\/h2>\n\n\n\n
![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/04\/image-273-1024x572.png\")
<\/figure>\n<\/figure>\n\n\n\n