[enhancement]: Optional acceleration for non-distilled models

[iwr-redmond]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

No response

What should this feature add?

It may be helpful to add support for easy access to distillation LoRAs on non-distilled architectures, e.g. for Strix Halo users and those with lower-end cards that have sufficient memory but slower rendering speeds. I reckon that this would be best done in stages:

1. Testing support for acceleration LoRAs (noting [bug]: Hyper and turbo flux loras fail to load and patch #8415)
2. Adding relevant LoRAs to the Starter Models list
3. Adding a checkbox to the linear UI that can be activated by users when desired
4. Guarding against incompatible CFG settings when acceleration is activated

Here is a list of likely LoRAs for inclusion:

Architecture	Steps	Huggingface Repository
Stable Diffusion 1.5	8 (CFG preserved)	ByteDance/Hyper-SD
Stable Diffusion XL	8 (CFG preserved)	ByteDance/Hyper-SD
Stable Diffusion 3.5	4 (CFG preserved)	ByteDance/Hyper-SD
Flux.1 Dev	4*	Lingyuzhou/Hyper_Flux.1_Dev_4_step_Lora
Anima	4 (requires CFG=1)	sorryhyun/anima-turbo-4step
Krea 2 Turbo	4*	lvladikov/Krea2-Turbo-Distill-4step-LoRA

* Base model already distilled

Alternatives

The CFG-preserved LoRAs listed usually have lower-step alternatives that require the CFG to be set to 1. I have opted here for the more conservative options.

Additional Content

• Fooocus previously bundled the 4-step LoRA with its SDXL implementation, making it available via a preset
• If this FR is implemented, it may be possible to support CPU-only generation on select architectures using 1-step LoRAs, as demonstrated in this existing Diffusers implementation
  • SD1.5
  • SDXL
  • Flux.2 Klein 4B