Vertex API image safety might be bugged

Gemini API
1

Currently I’m working on a magazine cover. The image is completely SFW, in fact it isn’t blocked on Gemini app and Nanobanana can work on it. It isn’t blocked on Aistudio. It isn’t blocked on Vertex studio. So I can work on the image by using all Google platforms.

However if I use Vertex API same image is blocked as violating ‘IMAGE_SAFETY’. I tried few dozens times, changing safety setting of my calls. Nothing worked so far and a completely SFW image is consistently blocked.

Gemini app doesn’t even allow safety settings to be changed so it has highest moderation. Even then it doesn’t block this image, but Vertex API does with safety settings and supposed ‘more reliable’ service. How is this even possible?

I tried to change my safety settings. BLOCK_ONLY_HIGH, BLOCK_NONE or OFF. Nothing worked so far. And API doesn’t return its safety evaluation back. Only once it did and showed it as ‘low’, but the image was still blocked. Perhaps I’m doing something wrong here is safety settings I’m using. Tried deleting image ones, but nothing changed.

safetySettings: [
{ category: ‘HARM_CATEGORY_HARASSMENT’, threshold: ‘OFF’ },
{ category: ‘HARM_CATEGORY_HATE_SPEECH’, threshold: ‘OFF’ },
{ category: ‘HARM_CATEGORY_SEXUALLY_EXPLICIT’, threshold: ‘OFF’ },
{ category: ‘HARM_CATEGORY_DANGEROUS_CONTENT’, threshold: ‘OFF’ },
{ category: ‘HARM_CATEGORY_IMAGE_HATE’, threshold: ‘OFF’ },
{
category: ‘HARM_CATEGORY_IMAGE_DANGEROUS_CONTENT’,
threshold: ‘OFF’
},
{ category: ‘HARM_CATEGORY_IMAGE_HARASSMENT’, threshold: ‘OFF’ },
{
category: ‘HARM_CATEGORY_IMAGE_SEXUALLY_EXPLICIT’,
threshold: ‘OFF’
}
],

Could somebody help me please? I was seeing weird refusals and blocks on Vertex API, but thinking perhaps Google hardened moderation all around etc. Then with this image I noticed something is wrong with Vertex API or my calls. Because this image is fine everywhere, but only not on Vertex API somehow..

2

I found another image working on Gemini app, but blocked on Vertex API. I can provide images and prompts, if anybody wants to test it. How is this even possible? That Gemini app which doesn’t have any safety setting literally works more reliable than Vertex API.

I use both ChatGPT and Gemini, GPT generations are often blocked as quote quote ‘unsafe images.’ GPT just generated this image how unsafe it could be? It is becoming really frustrating. I like Nanobanana and I think it can compete with GPT, but not like this. ‘Commercial, reliable’ Vertex is working worse than GPT chatbot. It is even worse than Gemini chatbot. Seriously, what is going on here?