Regression in handling of images larger than around 1MB

[AlexandreBouffard]

I have a script that iterates over images to perform an inline prompt using defImages().

• Images smaller than around 1MB process as expected. Debugging shows that their content attribute is populated.
• Images larger than around 1MB throw error Unsupported buffer-like object in resolveBufferLike(). Debugging shows that their content attribute is missing.

This script was working fine a few days/weeks ago, with same input files.

[pelikhan]

Looking

[pelikhan]

I may have broken things when working with my 2gb videos.

[pelikhan]

Try 101.0

[AlexandreBouffard]

Now working with 1.101.1.

However initializing env with files (in my case 30 images each between 500kB and 2MB), before first prompt is sent, takes around 7 minutes with node process at 100% of one CPU core. During that time, in the GenAIScript tab in VS Code, I see files appearing one by one every 15 seconds or so. I do not recall any preprocessing delay, maybe at most a few seconds, last time I ran this script with same input files.

[pelikhan]

Let me test this scenario.

[pelikhan]

do you use defImages(env.files) in your script?

[AlexandreBouffard]

I do a for loop on env.files and use _.defImages(file) in the inline prompt, but without any prompt at all, just initializing env.files takes around 7 minutes.

For example with test folder containing 30 images each between 500kB and 2MB:

script({
  files: ["test/*"],
});

for (const file of env.files) {
}

[pelikhan]

Perfect thanks for the repro. I've started loading binary files and this might be the issue. Also it's possible that these images get processed by Jimp which can be slow as well. looking

[pelikhan]

Oopsies, we were tokenizing the binary data in the trace. Back to instantly fast (until it hits the resizing step)

[pelikhan]

@AlexandreBouffard try v1.101.2

[AlexandreBouffard]

@pelikhan fixed!