Dall-E 2 false positives for content violations

also given enough fine tuning and effort, you could probably make it work with babbage or curie, further reducing the cost. I’m sure they have their reasons for using a wordlist but having somthing that can understand context would help a ton. Maybe use a wordlist as a backup gpt fails? The added latency shouldnt be a problem since it already takes a bit of time to generate the image.

1 Like

I highly doubt that. It’s simply easier to implement (no train - just yeet wordlist in and done) and costs nothing, but that’s not a valid reason imo.

Meanwhile tens of people have already been banned for no reason at all, hurting OpenAI’s revenue. Most don’t come back after that, considering that there are alternatives which are cheaper, work almost as good and the user doesn’t need to wait a team’s manual decision to get unbanned.

It’s kinda sad. I love OpenAI, but this decision in particular I just don’t get it.

2 Likes

I had my account suspended, filed an appeal a week ago. Was just trying to edit the rib of an angel with inpainting. I can’t figure out anything wrong I’ve done.

1 Like