An impressively fast model
Prints nine 640x832 images in less than 3 seconds, decodes nine 640x832 images in less than half a second; Meine Respektierung to the architecture chef(s)
note: some underclocked 5070 Ti slapping 640x832 x9 batch at 4 NFEs as a quick little test
I look forward to whatever funstuffs gets done with this model, it smells good.
Now the question is, how uncensored is it, how easy is it to make loras...
how easy is it to make loras...
I guess nobody has yet, need to wait a bit. sad to say, but I found it dumber than Klein 4B, will it get fine-tuned, yes, but I don't think it will replace Klein
how uncensored is it
Can generate toplessness, I have not tested it much yet ...
how easy is it to make loras...
I guess nobody has yet, need to wait a bit. sad to say, but I found it dumber than Klein 4B, will it get fine-tuned, yes, but I don't think it will replace Klein
how uncensored is it
Can generate toplessness, I have not tested it much yet ...
Yeah, this model is very bad with hands and another concepts. I think that Krea 2 is definitely the best open weight model for now.
Yeah, this model is very bad with hands and another concepts. I think that Krea 2 is definitely the best open weight model for now.
Well, if you make it print promptless outputs, it's little wonder that it is quite rough;
it prints mostly webpage screenshots, gibberish text, and powerpoint slides :-)
Here's some promptless outputs from Anima and Krea2
the former is mostly danbooru pictures, with occasionally something else thrown in there - pretty much as described by its author
Krea2 is a mixed bag of "eye-catching picture" sorts
and then, we have mageflow;
don't forget; Garbage in, Garbage out :- )
The promising part here is not so much the performance of the base model, but the slick speedy architecture with the important bits in place.
The world's laziest 'fine-tune' would be copying end-user outputs of Krea2; as they happen to share the same text encoder, 'it's a prettier base model' (less 'garbage in'), and it's very popular
Yeah, this model is very bad with hands and another concepts. I think that Krea 2 is definitely the best open weight model for now.
Well, if you make it print promptless outputs, it's little wonder that it is quite rough;
it prints mostly webpage screenshots, gibberish text, and powerpoint slides :-)
Here's some promptless outputs from Anima and Krea2
the former is mostly danbooru pictures, with occasionally something else thrown in there - pretty much as described by its authorKrea2 is a mixed bag of "eye-catching picture" sorts
and then, we have mageflow;
don't forget; Garbage in, Garbage out :- )
That's an interesting way to test a model ‘‚:)
Running a model without guidance nor prompt for roughly as many steps as it takes to converge on an image (usually 30 is enough with Gradient Estimation as the sampler) is a very telling method of making the model spit out what secrets it holds, and what it's likely to be good at.
If it spits out varied characters in various risque poses, then it's likely good at that. If it spits out cool scenes, it's likely good at that. If it spits out furry pr0n and questionable content, then you know who's who.
If it spits out powerpoint slides, then it was probably overtrained on text examples to make a "hey look at me, I can be a big boy who does text too!!" public presentation.
the best parts of the model is the crazy inference speed, surprisingly low cost of resolution scaling (i.e. crazy inference speed, spelled differently), an excellent text encoder neighbour to steal homework from, supposedly a very good flux2 VAE space, and the fact that you already get functioning editing and reference image inputs.
Running a model without guidance nor prompt for roughly as many steps as it takes to converge on an image (usually 30 is enough with Gradient Estimation as the sampler) is a very telling method of making the model spit out what secrets it holds, and what it's likely to be good at.
If it spits out varied characters in various risque poses, then it's likely good at that. If it spits out cool scenes, it's likely good at that. If it spits out furry pr0n and questionable content, then you know who's who.
If it spits out powerpoint slides, then it was probably overtrained on text examples to make a "hey look at me, I can be a big boy who does text too!!" public presentation.
the best parts of the model is the crazy inference speed, surprisingly low cost of resolution scaling (i.e. crazy inference speed, spelled differently), an excellent text encoder neighbour to steal homework from, supposedly a very good flux2 VAE space, and the fact that you already get functioning editing and reference image inputs.
Yeah. As I had tested, Mage flow is a nice model, but it's lack of knowledge in another general concepts makes it worse (for me). I wonder, what if this model had 6B/8B instead 4B? It could be probably slower and heavyweight but I think they could overcome Z-Image.
Note: I had tested the model on hugging face before its space censorship. I didn't get nice results on ComfyUI due some issues.