Text-to-Image

this model looks like garbage (old title: diffusers format should be standard)

#2
by bghira - opened

huggingface isn't just a place to dump your models, it's a community of developers.

please stop releasing models without diffusers equivalent weights.

Have to agree here, this is Huggingface, it's for diffusers and open model research. Seems like we're kind of just dumping models here to use the site bandwidth for downloads. Where is SD3, also? cosXL? Is this like the third or fourth fancied up version of the same model with the same parameters?

to be fair i think the only folks remaining at SAI are either business-oriented or just normal finetuners like Lykon or Nitrosocke and old sysadmins like TwoDukes so we're just seeing the best that they can do, and it's this

Yes, maybe I'm being a little hard on them, it's just annoying. They closed the waitlist for SD3, so not sure what that means now... Kind of been waiting for them to drop it, but with the departures it may have got pushed to the side for the time being. Hopefully they can get some more brains on board and salvage the project, I know it's hard work and requires a lot of time and patience.

from what we've observed on the dedicated discord server that's just for SD3, the outputs really aren't good. it looks like SD 2.2 with better prompt adherence, the problem is it always looks clipart composed together. like a collage of magazine photos.

it still can't make hands correctly, it can't identify every day objects correctly. like you ask for a hammer and a nail, and get a complete trainwreck.

Jesus christ can you be more stuck up

@asadas i'll wait til you publish something worth talking about before you become a part of the convo

Yes Huggingface is supposed to be a community of developers yet you are using it as a dump for your uninformed ramblings ptx0.

I prefer single file format which typical users also prefer. Diffusers also supports it very well for SDXL as far as I know so I'm not even sure why you are complaining.

comfy, i would expect someone who works at a dying company to be a bit more welcoming to community members.

"uninformed ramblings" is what you told a LAION developer recently - that he "doesn't know how text encoders work".

are you capable of having a positive interaction with the community? or are you just toxic?

@asadas i'll wait til you publish something worth talking about before you become a part of the convo

As opposed to who, you?.

comfy, i would expect someone who works at a dying company to be a bit more welcoming to community members.

"uninformed ramblings" is what you told a LAION developer recently - that he "doesn't know how text encoders work".

are you capable of having a positive interaction with the community? or are you just toxic?

That's rich coming from you hahaha

comfy, i would expect someone who works at a dying company to be a bit more welcoming to community members.

"uninformed ramblings" is what you told a LAION developer recently - that he "doesn't know how text encoders work".

are you capable of having a positive interaction with the community? or are you just toxic?

That's rich coming from you hahaha

assuming this is devilismyfriend, another SAI worker that is upset about the state of their company and lashing-out at others.

need I remind you how your StableTuner development efforts were perceived? your own tone that you take when interacting with others? that the project is now abandoned?

do I need to remind SAI that SD3 isn't even released yet? why are you all on here arguing with some "nobody"?

the diffusers weights for CosXL aren't even out yet.

there's a lot you guys can be doing to save your dying company, and yet, you're here arguing with an idiot.

comfy, i would expect someone who works at a dying company to be a bit more welcoming to community members.

"uninformed ramblings" is what you told a LAION developer recently - that he "doesn't know how text encoders work".

are you capable of having a positive interaction with the community? or are you just toxic?

That's rich coming from you hahaha

assuming this is devilismyfriend, another SAI worker that is upset about the state of their company and lashing-out at others.

need I remind you how your StableTuner development efforts were perceived? your own tone that you take when interacting with others? that the project is now abandoned?

do I need to remind SAI that SD3 isn't even released yet? why are you all on here arguing with some "nobody"?

the diffusers weights for CosXL aren't even out yet.

there's a lot you guys can be doing to save your dying company, and yet, you're here arguing with an idiot.

... What?, is everything a conspiracy to you?

just gonna wait for the Diffusers layout / model configs and ignore the BS from the SAI workers who are doing everything but their jobs

Stability AI org

If anyone wants to know how to use the model with diffusers, feel free to check it here:
https://huggingface.co/spaces/multimodalart/cosxl/blob/main/app.py#L10-L37

I've built a demo for it and works like a charm as @comfyanonymous mentioned, works great with a single file 🤗

Beautiful, thank you @multimodalart !

re @Leomn This is a side model we were putting together before SD3 took over development focus, and figured it would be best to quietly put out there in the short period of time where XL is still the leading open model arch, before SD3 takes the throne. SD3 will likely be out in somewhere in the range of a few weeks (no promises yet, bug chrLaf on twitter if you want more info).

@ptx0 The random hostility is unnecessary. Especially literally guessing that a random user might be somebody else (ftr I'm pretty sure that's not anybody that works here?) and then going on a rant against the person you've guessed them to be. That's wildly beyond inappropriate in any situation, much less a technical thread on a model release page??

@AlexGoodwin please reorient the disdain toward those who derailed the thread and started out swinging with insults. please pretend to be morally superior elsewhere. maybe try pitching StableSwarm to more people that wont use it.

image.png

prompting solid black background shows some kind of signal leaking concentrates in the lower right corner.

so it still has problems that Bytedance solved last year, and that I've solved through a v-prediction retrain of SDXL last year as well.

image.png

prompt: pure white background

better luck next time i guess guys

ouchies, high frequency noise galore. was this trained on a laptop?
image.png

top open model arch, lol
image.png

image.png

CosXL looks worse than ptx0/terminus-xl-velocity-v1 which is also v-prediction, trained from scratch (not the VAE though) on >7.9 million Midjourney v5.2 outputs, but uses zero-terminal SNR noise.

maybe something in Diffusers is messing up v-prediction results, maybe it's the training data, maybe it's a lot of things.

but it's not good, or state of the art, or even living up to the claims made on the model card

bghira changed discussion title from diffusers format should be standard to this model looks like garbage (old title: diffusers format should be standard)

This is an extremely toxic thread with no constructive comments and will be locked. If you don't like the model then move on quietly, this is for researchers to experiment with, not for people to create flame wars and be toxic.

Pseudo, if you truly wanted to uphold and represent the community of huggingface as you originally tried to posture with this thread you've ended up doing quite the opposite with your continued trolling and toxicity. Please refrain from continuing conversations in this manner.

TwoPerCent locked this discussion

Sign up or log in to comment