Topic

AI

10 posts on AI.

Three crops of the smallest text on the Inter sheet with the fields Gemini read: the PNG (228 KB) 32 of 32, JPEG quality 40 (172 KB) 32 of 32, JPEG quality 5 (76 KB) 9 to 14 of 32, where the digits are visibly smeared.

Does compressing an image change what Gemini reads? Full-colour JPEG quality 95 to 40 didn't, shrinking to 600 pixels did

Gemini 3.8 Flash read the same number of fields from a full-colour jpegli JPEG saved at quality 40 as from one saved at quality 95, on both of the text sheets we tested. A copy shrunk to 600 pixels wide, a PNG with no JPEG loss in it, read 15 fields fewer than the quality 95 file on both, comparing medians. A field is one code, price, date or 12-digit number, and each sheet holds 192 of them as random strings, so there's nothing to guess from. Anthropic's vision page says "heavy JPEG compression can make text difficult to read" and gives no number. Google's image page advises "Use clear, non-blurry images" and doesn't mention compression. We read both pages on 8 November 2026. These numbers are for Gemini 3.8 Flash, through the API, on 8 November 2026.

7 min read
Bar chart of file sizes for a photo at four print sizes, 8x10 to 24x36. PNG bars run from 11.3 to 70.6 MB and JPEG bars from 1.4 to 6.5 MB, with a dashed line at Etsy's 20 MB limit.

How many pixels does an AI print need? 300 per inch, and at 16x20 our PNG is over Etsy's download limit

The usual answer is 300 pixels per inch of print: 2400 x 3000 for an 8x10, 4800 x 6000 for a 16x20, 7200 x 10800 for a 24x36. A square Midjourney 8.2 HD image is 2048 x 2048 and a square Gemini 4K image is 4096 x 4096, so neither reaches a 16x20 at that rate without more pixels. (Shops write DPI. Here it means pixels per inch of print.) From a 4K Gemini picture we made 16x20 files of 4800 x 6000 pixels: the PNG is 35.8 MB, over Etsy's 20 MB per file for digital downloads, and the JPEG is 3.6 MB.

7 min read
Ten squares: 4 filled for tools that kept the IPTC AI tag at their defaults, 6 hollow for tools that dropped it, beside the title Does compression remove the AI tag?

Does compression remove the AI tag Merchant Center asks for? 10 tools, WordPress and Iminify tested

Google's Merchant Center help says: "Don't remove embedded metadata tags such as trainedAlgorithmicMedia from such images. All AI-generated images must contain the IPTC DigitalSourceType trainedAlgorithmicMedia tag." (Merchant Center's help page is dated February 2024; we read it on 26 October 2026.) So we put that tag on two AI images, a harbour scene from Nano Banana 2 (a JPEG) and a bakery illustration from the original Nano Banana (a PNG), ran them through 42 command-line runs and Iminify, added three larger JPEGs for a WordPress 7.1.3 site, and read the tag back with ExifTool each time. Each tool's default decides whether the tag survives, and neither lossy nor lossless predicts it. Of the ten tools in our table, four keep the tag at their defaults and six drop it, and for nine of them we found an option that flips that. Iminify, which is our product, lost it at every level that compresses: it survived in 6 of 51 uploads, and those six came back byte for byte unchanged. One ExifTool command puts the tag back for 514 to 568 bytes and, on the nine files we checked, leaves every pixel alone.

8 min read
Two cards under the title "Do AI images hurt SEO? The file can.": a Gemini 2K picture as delivered, 3.95 MB, largest paint 20.3 s; the same picture resized to 720 wide, 104 KB, 1.4 s, in Lighthouse's mobile test.

Do AI images hurt SEO? What Google's pages say, and what a 2K Gemini picture did to a page

We put Gemini pictures on a plain article page, one size at a time, and ran Lighthouse's mobile test on each. Straight from the generator, the 1K file (1.06 MB) took 6.2 seconds to become the page's largest paint (LCP), the 2K file (3.95 MB) took 20.3 seconds and the 4K file (10.51 MB) took 52.4 seconds. Google's target for that metric is 2.5 seconds. Do AI images hurt SEO? In the Google pages we read, nothing marks an image down for being AI-made. The file is another matter: the same 2K picture, resized and compressed to 104 KB, took 1.4 seconds. Iminify's compressor also drops the C2PA manifest that marks a Gemini file as AI-made, and Google's Merchant Center asks sellers to keep an IPTC tag that says the same.

8 min read
Grid of results for one iPhone photo sent in seven formats to three AI APIs: Claude's and OpenAI's read JPEG, PNG, WebP and GIF and refused AVIF, TIFF and HEIC; Gemini 3.8 Flash read all seven.

Can ChatGPT read HEIC? What five AI tools list, and what three APIs did with an iPhone photo

On 11 October 2026 we sent the same iPhone photo, a 1.7 MB HEIC of two greyhounds looking into a café, to three AI APIs (the interfaces developers use, not the chat apps). Gemini 3.8 Flash answered "Two leashed dogs looking inside a stylish, cozy restaurant." OpenAI's API replied "The image data you provided does not represent a valid image." Claude's API returned an error naming the four types it takes: JPEG, PNG, GIF and WebP. So can ChatGPT read HEIC? OpenAI's help page for ChatGPT lists three image types: PNG, JPEG and non-animated GIF. HEIC isn't one of them. Convert the photo to JPG before you attach it: JPEG, the same format as JPG, is on every format list we read from OpenAI, Anthropic, Google, Microsoft and xAI. Iminify's HEIC to JPG page converts it without an account. Two of the lists we read name HEIC: Google's image page for the Gemini API and xAI's FAQ for the Grok app.

7 min read
Three c2patool verdicts on a Google-signed Gemini JPEG: the 1,037,497-byte original is Trusted, a libjpeg-turbo jpegtran -copy all -optimize copy with the same pixels is Invalid (dataHash.mismatch), a cjpegli quality 90 copy has no claim.

Does compression remove Content Credentials? Two signed Gemini images through 13 tools

We took an image made by Nano Banana 2, one of Google's Gemini image models, and asked c2patool whether its Content Credentials hold. It answered "Trusted": a C2PA manifest signed by Google LLC, stating "Created by Google Generative AI.", with every hash matching. Then we ran that JPEG, and a PNG from the original Nano Banana, through 13 tools, WordPress 7.1.2 and Iminify among them, and asked again. Every file those tools wrote failed if it differed from its source by even a byte. Most tools dropped the manifest. Five kept a copy of it that no longer verified, in some cases without changing a single pixel. If a compressed file has to carry a working credential, something has to sign it again after the last byte changes, and that adds bytes of its own.

8 min read
ComfyUI's Flux Dev example render of a fox-eared girl in a maid dress holding a cake, captioned: 1,215,064-byte PNG to 125,684-byte WebP, and the workflow still loads in ComfyUI.

ComfyUI PNG to JPG loses the workflow. WebP and AVIF can keep it

ComfyUI's Flux Dev example image is a PNG of 1,215,064 bytes, and the two text chunks holding its prompt and workflow take 12,240 of them. The rest is 1024x1024 pixels stored without loss. Across the 92 PNGs in ComfyUI's examples repository that carry a workflow, the text was no more than 3.32 percent of any file, with a median of 0.62 percent. Converting to JPG shrinks the file, but ComfyUI's own loader can't read a workflow back out of a JPG. A WebP or AVIF can keep it, at about a tenth of the size. We took the text-to-image examples of ten models from ComfyUI's examples repository, made smaller copies several ways, and dropped the copies back onto ComfyUI 0.37.0's canvas to see which ones still held the workflow.

6 min read
Bar chart for one 4080x3072 phone photo, measured with each vendor's token counter: 1,102 tokens on Gemini 3.8 Flash, 4,743 on Claude Opus 5.5 and 14,746 on GPT-5.6 at default detail.

How many tokens is an image? Claude, GPT and Gemini, measured on seven real files

A 12.5-megapixel phone photo counts 4,743 input tokens on Claude Opus 5.5, 1,102 on Gemini 3.8 Flash and 14,746 on GPT-5.6 at its default detail setting, over 13 times Gemini's count. We measured all three with the vendors' own token counters on 23 September 2026. And none of the three depends on how many bytes the file has. Claude counts pixels up to a cap, GPT-5.6 counts every pixel unless you ask for less, and Gemini 3 barely looks at the size.

7 min read
Bar chart: a 12.5 MP phone photo at JPEG quality 90 is 4,309,090 bytes as a file and 5,745,456 bytes as base64, past a dashed line at the 5,242,880-byte limit on Bedrock and Google Cloud.

Image exceeds 5 MB maximum? Claude counts the base64, not your file

Our 12.5-megapixel test photo, re-saved with jpegli at quality 90, is 4,309,090 bytes. That's under 5 MB, and it's still too big for Claude on Amazon Bedrock or Google Cloud. Anthropic's limit there is "5 MB (base64-encoded)", and base64 turns those 4,309,090 bytes into 5,745,456. The error compares the encoded size with 5,242,880 bytes, as in this one from a Claude Code issue: image exceeds 5 MB maximum: 5281812 bytes > 5242880 bytes. So the ceiling is 3,932,160 bytes of actual file on Bedrock and Google Cloud, and roughly 7.5 MB on the direct Claude API, whose limit is 10 MB. Separately, no side may pass 8000 pixels, or 2000 once a request carries more than 20 images. For photos, one step clears all of it: resize to 2000 pixels on the long edge and save it as JPEG.

9 min read
The harbour picture Nano Banana 2 generated at 4K, 5504x3072: 10.51 MB as sent, a quality-100 JPEG, and 1.58 MB re-encoded as a quality-82 JPEG holding 40 dB.

How big is a Nano Banana image? 52 files measured, from 512 px to 4K

We asked Nano Banana 2 for a 4K, 16:9 picture of a fishing harbour through the Gemini API on 23 September 2026. It came back as a 5504x3072 JPEG of 10.51 MB. The same request again gave 11.10 MB. Both files were saved at JPEG quality 100, the top of the scale. So, how big is a Nano Banana image? From Nano Banana 2 or Pro, a detailed picture is roughly 1 MB at 1K, 4 MB at 2K and 10.45 to 12.13 MB at 4K, and a flat illustration less. The original Nano Banana sends a PNG instead, about 2.1 MB for the same prompt at 1024x1024. Here are the 52 images of our test set, from all four Nano Banana models, and what five of them shrink to.

9 min read

Try it on your own images

Drop a photograph in, pick Smart, and compare the result with the original before you download. Free, no account needed.