Google's New AI Image Models Generate 4K Files. Here's What to Do With Them.

Google launched its new Imagen 4 family and updated Gemini native image models this week, and the output quality is genuinely significant. Imagen 4 Ultra at 2K produces compositions with better text rendering accuracy, more coherent fine detail, and more convincing lighting than its predecessors. The Gemini 3.1 Flash Image model (sometimes referred to by its internal name Nano Banana 2) takes 4K output to a broader audience through the API.

The images are impressive. They're also, by default, enormous.

A single Imagen 4 Ultra output at 2K resolution is typically 8-15MB in PNG format. A Gemini 3 native 4K output can reach 25MB or more. These are files designed for quality, not for web delivery — and the default assumption in Google's documentation is that you'll decide what to do with them afterward. Most people don't decide. They download and upload directly, and wonder why their landing page LCP is measured in seconds instead of milliseconds.

What's Different About Compressing AI-Generated Images

If you've read the general guide to compressing AI art from Midjourney and DALL-E, the Gemini and Imagen workflow shares the same principles with a few model-specific differences.

The core challenge with all diffusion model output is the noise structure. Diffusion models generate images by starting from random noise and iteratively denoising toward the target image. Even in the final output, a faint residual pattern from that process remains — a fine-grained, high-frequency structure distributed across the image that doesn't correspond to real-world texture. It looks like mild grain or slight texture even in areas that should be perfectly smooth.

That noise structure matters for compression because:

JPEG treats it as image content. JPEG's discrete cosine transform encodes high-frequency variation, and diffusion noise is high-frequency by nature. JPEG spends compression budget on it the same way it would on real texture, which means you need a higher quality setting to avoid JPEG introducing block artifacts that compete visually with the diffusion noise. Standard photo quality settings (JPEG Q75) are too aggressive for AI images; Q80-85 is the practical minimum to avoid visible degradation.

WebP handles it more gracefully than JPEG. WebP's in-loop deblocking filter smooths out the block-boundary artifacts that JPEG introduces around the diffusion noise without blurring the underlying sharp detail. For equivalent file sizes, WebP produces visibly better results on AI-generated content.

Compression Settings Specifically for Gemini and Imagen Output

Here's what actually works for these models' output, based on their specific noise characteristics:

For Gemini 3 (Nano Banana 2) 4K output: Gemini 3's native images have a relatively fine diffusion noise structure — less pronounced than older diffusion models but present. The 4K resolution means you're almost certainly downscaling before publishing, which also helps (downscaling averages out the noise).

Start by downscaling to your actual display resolution: 1200-1600px for a website hero image, 800-1000px for a blog header. Then compress:

• WebP quality 78 → typically 250-450KB for a 1200px hero image
• AVIF quality 68 → typically 150-280KB for the same image
• JPEG quality 82 → typically 400-700KB, with more visible artifacts

For Imagen 4 Ultra 2K output: Imagen 4's output tends to have smoother transitions and less diffusion noise than Gemini native — the Imagen family has been trained specifically for image quality rather than multimodal reasoning. This means you can compress slightly more aggressively.

• WebP quality 75 → typically 200-380KB for a 1200px hero image
• AVIF quality 65 → typically 120-220KB
• JPEG quality 80 → acceptable but noticeably worse at edges

For Imagen 4 Fast (the low-cost option at $0.02/image): Imagen 4 Fast has a more pronounced noise structure than Ultra — it's the model you use when you need speed and cost, not maximum quality. For web use, compress conservatively: WebP quality 80-82, or JPEG quality 85+. Going more aggressive on this model introduces visible artifacts that weren't present in the Imagen 4 Ultra equivalent.

The Metadata You're Probably Distributing Unknowingly

Google's Imagen and Gemini image generation APIs embed metadata in the output files. At a minimum, this typically includes:

• The text prompt used to generate the image
• The model and version (e.g., imagen-4.0-ultra-generate-001)
• The API request timestamp
• In some cases, safety evaluation results from the model's content filtering

If you're publishing these images on your website, in email newsletters, or anywhere publicly accessible, that metadata travels with the file unless you explicitly strip it. Your competitors can see your prompts. Your clients can see your workflow. Your privacy in terms of creative process is exposed by default.

Run any Imagen or Gemini output through Optimage's metadata-stripping step before publishing. It removes EXIF data, XMP fields, and any embedded text blocks while leaving the image pixel data intact. For a 15MB PNG that becomes a 300KB WebP, the metadata is irrelevant to the quality of the published result, but it's not irrelevant to your business.

The Practical Workflow for Marketing and Web Use

One thing worth noting about the 4K resolution support in Gemini 3: for most web applications, generating at 4K is actually counterproductive. The higher resolution means longer generation time, higher API cost, and a larger source file to compress down — but the published result at 1200px looks the same whether you compressed from 4K or 2K, because you're throwing away the same amount of resolution either way.

Generate at the resolution closest to your target display size, or at 2K if you want comfortable headroom. The 4K option is useful for print, large-format displays, and cases where you're genuinely going to use the full resolution. For a website, the LCP hero image at 1600px display width needs at most a 1600-2400px source.

At $0.06-0.13 per image, Imagen 4 Ultra and Gemini 3 outputs are worth treating as production assets rather than disposable sketches. Compressing them properly for web delivery is the last step that makes them actually useful in production — not an optional cleanup.

Related reading:

• AI-Generated Images Are Beautiful and Enormous: Here Is How to Compress Them — the general guide covering Midjourney and DALL-E workflows
• AVIF vs WebP vs JPEG 2026 Benchmark — format comparison with real performance numbers
• What Is EXIF Metadata and Why Strip It — why stripping metadata matters beyond just file size