The accelerator of FLUX-dev, can reduce steps from 20 to 8 and have same quality.
https://huggingface.co/ByteDance/Hyper-SD
For V2
My optimized version
1. The weight can be 1.0, and no overload noise will be generated.
2. LORA has no weakening effect on the image.
3. The size of LORA is adjusted to about 250 MB.
For V1
Note that its weight should not exceed over 0.3
Description
dev
FAQ
Comments (10)
why this version is 600mb but the two on huggingface are 1,4gb?
I made them less and add some self made thing inside.
@Aseer 🤔 anyway. i tryed in both forge and comfy and is not working for me, both DEV and Q8.
i even tryed download the original from ByteDance ass tiktok company, same stuff. no luck.
@LatentDream You need to use an FP16 checkpoint. NF4, FP8 does not work
@LatentDream you need to use this lora with 0.125-0.35 weight. if it goes more than 0.4 (As I wrote on introduction), it will ruin your images.
@TiwazM FP8 is fine, NF4 not suitable to any lora at all. and for dev-20 steps, you can reduce from 20 steps to 8 steps, for schnell, you can have more details.
It works great. Just set the lora weights.，0.125-0.35 weight.
will gguf q4 models work with this?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.