Diffusers current/future

[vladmandic]

Starting a brainstorming thread on ideas for longer-term priorities for diffusers...

• Better LoRA support
  Use of LoRA is very near top of priority for any user
  Right now, diffusers are pretty fast supporting LoRAs trained by diffusers for any new models
  But 95%+ of loras are NOT trained by diffusers and tools like onetrainer, kohya, ai-toolkit, simpletrainer (in no particular order)
  are much more widely adopted
  Also, like-it-or-not, CivitAI is de-facto standard for LoRA distribution so asking for LoRA to be uploaded to
  HF hub just so it can be tested is a no-no
  Right now, out-of-the-box, diffusers support very small part of LoRAs and its highly impacting its userbase
  It should become a priority, not an afterthought
• Support single-file everywhere
  There was a big initiative a year ago to make it happen and it resulted in a lot of refactoring
  But guess what? We're back in a position where a lot of new models are supported only using from_pretrained with HF URL
  Again, like-it-or-not, large majority of users prefer single-file safetensors
  And I can't say I truly blame them as HF folder-style is non-portable
  E.g. if I have a model, moving it somewhere is a nightmare (and frequently breaks due to hf libs usage of symlinks)
  Also, many users already have single-file downloaded for use in other apps such as ComfyUI
  But if they want to use it in diffusers, guess what? download again
• Focus on consumer GPUs
  Almost all new models are large and diffusers is fast to bring support
  But it takes very deep know-how to make it work on normal consumer GPUs as that is never part of original dev/test cycle
  E.g. even advanced users are left wondering how quantize each component separately
  and which quants are even allowed for each component
  And make recommendations so users are not left starting at dry docs which don't help them
• Create differentiators
  Adding new support for new model is me-too, but what is the value-add of diffusers other than its all in the same codebase?
  Initiatives as unified guiders (currently in draft) are very welcome as they bring such value add
  Remote VAE is another very positive initiative
  I'd even suggest going deeper - recently I've implemeted Nunchaku engine
  with its SVDQuant which has native 4bit execution via custom CUDA kernels
  Result? FLUX.1 works 3-5x faster. Yes, that's actual 300-500%!
  But although the approach in general is applicable on any DiT-based model, that team simply cannot do that
  as they do not have resources - but this is where diffusers team could step in!
  Nitpick the best-of-the-best solutions and actually port them to most popular pipelines

Anyhow, this is not intended to be exhaustive list - this is pretty much just a conversation starter
Any feedback and/or ideas are more than welcome

cc: @a-r-r-o-w @yiyixuxu @sayakpaul

[sayakpaul]

I will chime in for the LoRA bits.

I think we're also fast in supporting non-diffusers LoRAs and that is always a priority for us. If we check the timeline of the PRs we have opened after we got requests for supporting them, it would suggest our dedication towards making them a priority.

Regarding the Hub upload ask, that was my preference. None of my teammates do that, I think. I still think it's a fair ask and I will stick to my point. Even when a contributor doesn't provide the Hub URL to the LoRA, we will just do it ourselves (and we have done it multiple time; will continue in doing so).

Not providing a minimal reproducible snippet is a no-no for me and I will not indulge myself into further explaining it or any kind of other arguments whatsoever. If I were reluctant, I would not have tried figuring out a reproducible snippet myself (reference).

---

For consumer GPUs, we're on it. In the current release, we have shipped a ton of memory optimization-related stuff, specifically targeting consumer GPUs. We agree that the docs are currently not great and #11385 is a step forward towards that. We are also thinking of "auto" offloading strategies based on a given accelerator specification. More on that in the coming days.

[vladmandic]

my message was not an attack, its simply my view of the current situation.

re: lora - if you say we should open separate feature request for every lora that doesn't work, ok, we'll dedicate time for that - but my ask is opposite of that - to be more proactive as we know what user expectation is ahead of time. this is not about code snippets or feature requests or uploads to hub - its about being proactive when we know in advance where the biggest demand will be. and how do we know that? by seeing how many loras exist for any specific model on civitai - that is a very simple metrics.

re: consumer gpu - there is tons shipped (and i really do like quant support in diffusers).
but its difficult to use without any clear guidance. right now, every module needs to be loaded separately and then combined into pipeline and there is no note which modules and/or gpus support which quant types. i can deal with that on my side, but i wouldn't expect that from normal user.

[ParagEkbote]

Thank you for starting this discussion @vladmandic.

I had a couple of queries about the diffusers community pipelines. Firstly, does it make sense to deprecate the older/broken pipelines which have not been updated since the changes introduced in #6984?

Secondly, the newer community pipelines have not been added with respect to the API design of

pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", custom_pipeline="filename_in_the_community_folder")

Could they be refactored as well?

[vladmandic]

i would say at least diffusers need slow-test for community pipelines to see which one is working and which one is not and split the list into working/nonworking sections. then maybe community members can pick-and-choose which ones to fix as right now there is no way to know what's good and what's not.