... what is this? O_o
Curious what "DF11" means.
If you check my other model uploads, you should see links to what "DF11" means.
But essentially, DFloat11 is a technique for losslessly compressing BFloat16 models by about 30%, with absolutely no differences in output (unless there happens to be some weird bugs, like the case with Klein in ComfyUI). It works somewhat like a zip file, using entropy compression to compress the 8 exponent bits in BF16 down to about 3 bits, and the original weights can be 100% faithfully reconstructed from these compressed weights.
For your convenience, I have pasted the links in the screenshot above here: https://huggingface.co/DFloat11 and https://github.com/LeanModels/DFloat11.
Do check out my ComfyUI custom node as well, you can find it in the manager under the name "ComfyUI DFloat11 Extended". The GitHub repo is here: https://github.com/mingyi456/ComfyUI-DFloat11-Extended
please share the workflow for ace step in comfyui ... i installed your custome node in manager.. but wehre is the workflow?
@VivekLeon You can simply replace the "Load Diffusion Model" node in your usual Ace-Step workflow with the "Load DFloat11 Model" node, and that is it. The only thing is that LoRAs are currently not yet supported with Ace-Step. If you want LoRA support, do let me know and I will look into adding support for it.
Please add LoRA support for Ace Step. From now, I will only use DFloat11 models.
@PixelPlayer I just added LoRA support for Ace-Step1.5, but I have not tested it yet. Do give it a try and let me know if the output with LoRA is identical or not.
I can confirm that LoRA is now working. The outputs with LoRA enabled and disabled are definitely different (using the same seed), and with LoRA enabled, the style becomes 100% identical to the one LoRA was trained on. Everything is running smoothly; I haven't noticed any slowdown. Thank you very much!