Safety Filter
The model is great so far, but it's dependence of json prompting sucks. This feature should be only an advanced technic of prompting, not a mandatory stuff. This makes a ridiculous prompt like "a cute cat" being blocked by the "Safety Filter". Please, fix this, because this model have an insane potential to be best, even with the non-commercial license.
Yes the filter can be annoying, especially as it does all the compute and then blocks. Now json it not necessary and sometimes json seems worse.
Json
Text prompt only
Here a cute cat ;) Granted it did mess with json and and input image fir that ;)
Yes the filter can be annoying, especially as it does all the compute and then blocks. Now json it not necessary and sometimes json seems worse.
Json
Text prompt only
That's what I'm looking for in this model. You can see it, It works nice with natural prompting, but for some reason they didn't optimized that :']
Here a cute cut ;) Granted it did mess with json and and input image fir that ;)
It's a nice car by the way ‘‚:)
The model is totally useless because of safety filter. I can generate absolutely nothing. It is useless. I hope there will be some heretic versions soon.
The model is totally useless because of safety filter. I can generate absolutely nothing. It is useless. I hope there will be some heretic versions soon.
Yeah, I hope too. But note that filter is built in the model, so if we have a "heretic" version it will be probably worse than the original one. The situation gets even worse when you know they said that any sensitive output generated by the model is prohibited and should be reported for them refine this "safety filter". Again, it's so sad they censored the open weights of this model :[
The model is totally useless because of safety filter. I can generate absolutely nothing. It is useless. I hope there will be some heretic versions soon.
In my opinion, they are caring too much for what kind of content we generate with the open weights model. Obviously, any ilegal content generated and published by this model should be reported and repudiated, but not the standard nsfw thing, like nudity and sex, you know? It's also weird the fact that this model is completely nsfw capable but not able to generate it (even with sfw themes and natural prompting, lol). So why they trained the model with such images? The answer is that this is what makes this model so amazing, like GPT Image. Even for sfw outputs the nsfw stuff is necessary. Blocking this ability for an open weights model is like take off the wings of an angel...
Maybe, although the filter triggers so easily with completely innocent stuff like fantasy things. Seems much less of an issue with JSON prompts.
Now I just do it like: description -> LLM generate JSON based on it -> Ideogram4
Although still they look like paintings somehow. Not bad, but I ask for photorealistic and get that. Here's the example:
Anyone knows why it is that?
It makes this useless, it even bombs out on simple things. The false positives is insane. And furthermore, this should be should be banned from the earth and those who created it also. NO ONE should dictate what i want to generate. Screw you. Idiots.
It makes this useless, it even bombs out on simple things. The false positives is insane. And furthermore, this should be should be banned from the earth and those who created it also. NO ONE should dictate what i want to generate. Screw you. Idiots.
Releasing unusable software is not a crime.
JSON prompts work significantly better.
It makes this useless, it even bombs out on simple things. The false positives is insane. And furthermore, this should be should be banned from the earth and those who created it also. NO ONE should dictate what i want to generate. Screw you. Idiots.
Releasing unusable software is not a crime.
JSON prompts work significantly better.
Yeah, actually the model only works with this prompting technic, which is pretty bad. I hope they fix this.