next-image template: Request Entity Too Large

[sragss]

When I was working on templates/next-image I decided to make all inputs / outputs of the image models for the two edit endpoints and two generation endpoints base64 data URLs.

https://github.com/Merit-Systems/echo/tree/master/templates/next-image

Unfortunately we quickly exceed the datasize supported by either Next Client <-> Next Server or Next Server <-> gpt-image-1 / nano-banana.

Generation failed
HTTP 413: Request Entity Too Large FUNCTION_PAYLOAD_TOO_LARGE iad1::55kxp-1760452325185-1866997c3cd4

I believe we need to switch from base64 everywhere to hosted image URLS. I'm not sure if we can do this directly through the OpenAI / Gemini API or if we should rehost each image in a Vercel bucket and use those. Either way lots of plumbing needs to be adjusted.

[sragss]

@claude how about now?

[sragss]

@claude can you review this issue in depth and propose a solution? It will require searching for how the Gemini, gpt-image-1 and Vercel AI SDK work. If you're ready, open a PR.

[sragss]

https://vercel.com/guides/how-to-bypass-vercel-body-size-limit-serverless-functions

It looks like there are four options:

1. Increase Vercel body parse limit from 1mb -> 4.5mb (max)
2. Upload all images to blob store and pass around URLs
3. Compress the images on client
4. Steam all the data between Next Client <-> Next Server

2 is the obvious long term solution, but it decreases the ability to rapidly self-host. Many people using the template will not have Vercel nor Vercel blob set up and it will dramatically increase friction.

3 is kind of rancid. How much to compress? Fixed compression or iterative compression? Which compression algo? Compression as a function of the number and size of the attached input images?

4 is a hack.

[Otaiki1]

Hi i would like to solve this problem

[Otaiki1]

I would be using the 2nd option as a first attempt before going for option 4

[sachigoyal]

@sragss I have researched about this issue. The best way to implement it is using Vercel Blob. Should I work on it and make a PR.