Do you know what the limit is? Specifically for the image AIs
Do you know what the limit is? Specifically for the image AIs
KNot sure! Working on figuring that out today for ya
MAwesome thanks! Also let me know if there is a way to increase this limit 
PIf anyone else needs an AI mock service, we've created one! You can check it out on GitHub: https://github.com/openshieldai/mockai or access it directly here: https://mockai.openshield.ai. Currently, OpenAI is supported, but we will add more providers.
FGitHub link is 404
PThank you! Public now
FLooks great at a quick glance. I'd also announce this in -> #what-i-built
RThe limit is "fuzzy". Around 2.5-3MB depending on the content. Seems to apply to all models. (
JSON.stringify(inputs).length
KYour error mentions
8b8bSounds like you typo'd the model name, as you mentioned it's
@cf/meta/llama-3.2-3b-instruct
MThe image is around 3,5mb so you recommend compressing them?
RYeah, you could try that. The limit for compressed/random data is roughly 2.6MB. In my experience, sending smaller images (file size) has resulted in a significant improvement in latency.
FThese are still gigantic image sizes for LLMs, it should be possible to downscale them to ~100-300kb size and get similar results.
FMost likely the AI API will scale those images down anyway before processing.
MWhat format do you recommend to compress them to jpg? With what setting
IJPEG would work well. Keep in mind that LLMs actually use very small sizes internally anyway - a size of 640 pixels is good enough, no need for 4k images.
MWith the object classification from fb i feel like if you resize the box dimensions dont work anymore
REither way will work. Compressing or downscaling. The main thing is to get the file size down. When I tested I could shave of half a second of latency simply by sending the (1024x1024) image as a ~200kb jpg instead of a 2MB png.
IThe box dimensions should be based on the pixels you give, so it should be consistent? 
MYeah but if I want to then visualize the boxes on the original image it wouldnt match anymore right
II'd recommend resizing over compressing. A compressed 4k image is still going to be resized internally by the model anyway, and by aggressively compressing big images you're going to lose some of the details.
IJust make sure to calculate the resize ratio.
MI guess yeah
IIf you have an image that is 3840 and resize to 640 then just multiple the returned coordinates by 6 (3840 divided by 640)
IJust an FYI that I'm not picking 640 pixels arbitrarily - that's the actual size used by lots of models (ex. YOLO object detection)
MKind of difficult since the images I want to classify are user input and could differ but I’m sure ill figure it out
IHmmm yeah. So depending on what your input pipeline looks like, I'd take the longest dimension (height or width) and then divide that by 640, and then resize to 640, send the image, and multiple all results by that ratio
IMight be tricky if your pipeline is browser based but there's still options even then
MUsers upload through an app, it gets stored in R2 and then i’ll likely run some automation on it for it to be classified
IIf you're using a Worker for that automation you can try
sharpProbably be a job on our api thatll send it to the worker and update the db with the response
RJpeg blocks are 8x8 pixels, so you're unlikely to lose details if the image is resized by a large factor. I believe resnet expects 224x224 as input, which would amount to ~150kb of uncompressed image data. If you're going to handle the resizing yourself you still have to decide on which kernel to use and in which color space to do the operation - e.g. lancoz, linear gamma. Worst case scenario is if you happen to resize to image dimensions that doesn't quite match the model, since the low res image will be resized once more. So idk, I'd probably go with lossy compression. Although if the image dimensions are huge, it could make sense to bring them down a bit.
HHey, when I send
I'll get: Answer HTTP 400:
I'll get: Answer HTTP 400:
Hso it says content must be a string, but Openai-API says string or Array https://platform.openai.com/docs/api-reference/messages/createMessage#messages-createmessage-content
Aany idea how I can get more useful error messages? No matter what, I only get
AiError: AiError: Unknown internal errorim trying to use a lora adapter
Aplease ping me if you reply
TAny one manage to get flux to work? When I run the exactly example in the documentation it errors with NSFW:
F
FAm I the only one that can't get
@cf/black-forest-labs/flux-1-schnell
DHey, is anyone else still seeing this with flux?
IWe're going to try escalating it
DIsaac has escalated this, we'll let you know if we have any updates
CLooks like the example is categorized as nsfw but most other inputs aren't
RThe example code doesn't work. Flux returns a json object with the image encoded as base64. So you need to decode it.
FThank you, got it working now. 
RA lot of prompts triggers the NSFW error. "cyber", "bar", "toe", "arm", "nose", "drill", "boot", "spy", "cow", "chicken" just from trying a few random one word prompts. "a cyberpunk cat" is ok, "cyberpunk cat" isn't.
DThe team is looking into the filter, for the time being they recommend using longer prompts as they are less likely to be caught - they recommend at least 4-5 words
Aany ideas?
RThis. Although I wish https://developers.cloudflare.com/workers-ai/privacy/ were more clear. It feels well-intended and as if it is written to provide peace of mind, but "Cloudflare does not use .. to" can be read as "does use, except for these two things" in the same way that "may be stored by Cloudflare if you specifically use a storage service" can be read as "does not store, except if".
RRight now, however, I'm more concerned about Cloudflare throwing NSFW errors and wonder if this filter is Cloudflare's own doing, because if so then IMO Cloudflare is on a slippery slope of engaging in content moderation, and I can't find anything about NFSW in terms and conditions. So I would very much like to have some clarity regarding this. Might be raising an "unpopular option" thing here, but well.. I feel it is important.
