I had the same experience on their site, even following the given example! "Show me a horse" -> a brown horse appears. "Make it black" -> now it has black fur. "Make it white" -> I'm sorry, racist content is not permitted, you are a bad person
Pixelwave generates much better horse-people, in the event I wanted that. Admittedly it doesn't have an edit function; unfortunately the illustrations I want to edit all have people in them.
Ever since OpenAI showed (but did not release) this type of multimodal output with 4o, I have been waiting for this to be available to the general public.
It seems like really combining visuals at the level of generation capability means language understanding is fully grounded in a richer world model.
I am hoping for a step up in real world common sense intelligence areas like those covered by SimpleBench. Although they are static images, so there might still be room for improvement ad far as physics understanding.
Also, if they can get it to the point of really accurate (probably larger models), this unlocks whole industries in terms of being able to do useful work.
I was really hoping that there would be more character consistency, given the fact they mention it in the blog. It also doesn't seem to reliably follow styles like "watercolor illustration" or "line and wash".
It just tells me 'content not permitted'.
For context, I was attempting to put a cup of hot chocolate into the hands of an anime character.
I had the same experience on their site, even following the given example! "Show me a horse" -> a brown horse appears. "Make it black" -> now it has black fur. "Make it white" -> I'm sorry, racist content is not permitted, you are a bad person
I had exactly the same issue, seems like it's practically useless for humanoid illustrations/characters.
I think it's definition on what's a "person" is way too broad. I've had similar problems, it's pretty shitty.
It allowed me to do a horse, but-
https://usercontent.irccloud-cdn.com/file/DkQ5SSdT/image.png
Pixelwave generates much better horse-people, in the event I wanted that. Admittedly it doesn't have an edit function; unfortunately the illustrations I want to edit all have people in them.
Realistic photos work better, though it still doesn't beat Flux.1-dev: https://usercontent.irccloud-cdn.com/file/ZsouXNpn/image.png
Ever since OpenAI showed (but did not release) this type of multimodal output with 4o, I have been waiting for this to be available to the general public.
It seems like really combining visuals at the level of generation capability means language understanding is fully grounded in a richer world model.
I am hoping for a step up in real world common sense intelligence areas like those covered by SimpleBench. Although they are static images, so there might still be room for improvement ad far as physics understanding.
Also, if they can get it to the point of really accurate (probably larger models), this unlocks whole industries in terms of being able to do useful work.
If it can do diagrams, charts, etc, with any kind of accuracy, it would have far more impact.
Eg. "I suggest moving the boiler from point A to B on the below map of the factory to reduce piping costs and heat loss"
It's on the right track, but lacks precision.
https://aistudio.google.com/app/prompts?state=%7B%22ids%22:%...
Previous try with some interesting introspection:
https://drive.google.com/file/d/1SCBbpDo1dAJBAz7bFABk4yBZBuz..., https://aistudio.google.com/app/prompts?state=%7B%22ids%22:%...
I was really hoping that there would be more character consistency, given the fact they mention it in the blog. It also doesn't seem to reliably follow styles like "watercolor illustration" or "line and wash".