BF16 weights?
Hi there, thanks for publishing this great model!
I’ve experienced significant quality loss using FP8 vs BF16 weights. Is there any chance you could also provide the BF16 weights?
Thanks!
yea, why no bf16 weights if real open model
it was likely trained in fp8 only
Essential for mac compatibility
Would appreciate bf16, thanks.
it was likely trained in fp8 only
nowhere in the documents stated this, any proofs? if it is real native fp8 training, they should have bragged about that already
Hi there, thanks for publishing this great model!
I’ve experienced significant quality loss using FP8 vs BF16 weights. Is there any chance you could also provide the BF16 weights?
Thanks!
Just out of curiosity where and how exactly did you test the BF16 weights ? I've not seen them being published anywhere, any of the ones that are available are simply dequantized from FP8.
BF16 would be great because then i could convert to int8 for my 3090
Funny how not a single comment on this repo has received any sort of response from the official team. Either they're just watching and actually working on the stuff we're requesting or they clearly don't give a fk, I don't mean to sound rude but the entire community is excited about a model you release and have some requests or queries, the least you could do is reply.
Another vote for BF16 weights here.
it was likely trained in fp8 only
nowhere in the documents stated this, any proofs? if it is real native fp8 training, they should have bragged about that already
well training it with fp8 matmul works great, but the trainable params are bf16 - i have no real proof, just suggesting there in fact may be no bf16 weights to speak of.
I assume this fp8 version was made from the official Ideogram-4 model repo using a conversion script. Perhaps it is resource-intensive, but surely it is not a secret or proprietary method, correct?
I have the cycles to try and make a bf16 conversion and share if there are instructions.
FP8 is a lossy format. So, making a BF16 from FP8 means data will have been lost, permanently, versus an original BF16 file.
Hoping we get bf16 and a team member to answer some of these posts.
For anyone wondering, the AI community has been deemed not worthy in the slightest of the bf16 version of the model - https://www.youtube.com/live/gO-D5eO8VlA?t=2263
Sadly the community has been deemed unworthy and only useful for model testing, experimentation and optimization, so in a first in model release history, users are not worth the bf16 model being released. So no BF16 for you.
Looks like we have an answer to the question now. I just want to make a few points for anyone from the community or the Ideogram team (if they’re even reading it).
Some of the community responses I’ve seen to the missing BF16 weights are just insane. People feel entitled to demand free work from Ideogram while also hating on them for not meeting their demands. I get the frustration of having the best weights gatekept but there are better and more productive ways of expressing that frustration.
Never forget that these models take lots of money and resources to build, we should be grateful for having access to the FP8. FP8 weights on your local machine is still much better than BF16 behind an API. These negative and entitled comments only make things worse for everyone. Do you think Ideogram (or any other lab) is more or less likely to actually release the BF16 weights after seeing all the hate they’re getting?
The CEO said the reason they didn’t release BF16 is so it could be run on consumer hardware. It’s hard to buy that because the community can easily convert it to lower precision and quantize it. My guess is that they wanted to find a compromise between providing open weights while keeping a higher quality version privately so they can make money from it, and that’s fine, they’re a business not a charity. BFL does something similar with Flux but they do it through separate models and not reduced precision.
If anyone from Ideogram is reading this, thank you for the effort to provide open weights, even at FP8 the model is great. Please don’t be discouraged from the hate and negativity you’re getting about the precision and other things like the safety guardrails. There will always be haters and ungrateful people. For future releases, if any, I suggest you engage with the community more openly and be more clear about the limitations you impose and your reasoning behind it, that should clear a lot of the speculation and frustration from the community. Again, thanks for the model!
There will always be haters and ungrateful people.
Calling people names because they don't share your opinion isn't particularly pious. Models like Wan 2.2 set a standard in the community. When others don't live up to that expectation it shouldn't be a surprise when there is criticism.
Companies are under no obligation to provide BF16, FP32, nor anything else but people are also not under any obligation to remain silent if they have an opinion.
Nothing is free, including models that are released in any form. We live in an attention economy. Time and energy are resources. When people do things like create lora ecosystems for "free" models they are doing work to perpetuate that company's mindshare. If people don't think the model meets their expectations they have every right to make that known. What everyone should do, of course, is make their needs/wishes known in a professional manner.
you know who's unworthy? ... pricks that think they can exploit the good will of the ppl for great benefit, with paper thin PR bullshit... this is not an open source model! it's a lukewarm demo. Don't bother. fp8 is and always will be jank. you gotta do more to deserve playing a role in big productions.
our western model devs are severly misguided ... this goes VERY deep. you don't need no cfg ... 2b is all you need...
I AM the artist.. I TELL you what i need! not the other way around!
https://huggingface.co/Boogu/Boogu-Image-0.1-Edit
this respects you... it is not done yet and every feedback is welcome
BF16 base, apache 2 .. no post training... no silly filters ^^ ... the future
Judging from the silence from the gate keepers of the ideogram team, who have been watching all of this chaos incite, it's safe to assume Ideogram must've been eating dirt all these years while they were closed, so they released a few free samples and gained way more attention in the past few months than they ever did from the last 2 years.
"We've given away our free samples and yeah we're miserly so we only let the community sample fp8s, if you want the full weights start using our commercial, closed source shit, we will not reply to any of the madness that's been happening in our comments because were just as spineless as any other corporate dogshit, so don't even in your wildest dreams ever assume we'll be releasing the full open weights
Oh we're not just spineless but also let some NPCs roam amongst the community who will be sucking up to us on these comments, replying on behalf of us since we're incapable of coming up with an actual answer."