Greatly reduces the steps required for Flux dev.
Recommended LoRA scales around 0.125 that is adaptive with training and guidance scale could be kept on 3.5.
Note: I did not create this LoRa, all credits goes to ByteDance
Description
FAQ
Comments (9)
Great... Now we need to find a way to quantize LoRAs
I think this may be what you are looking for: https://civitai.com/models/693544/hypersd-flux1-dev-bytedance-optimized-size?modelVersionId=776158
I did in fact make a small change to the LoraConvert-node of ComfyUI-KJNodes and found that you can convert almost all LoRAs to FP8 with little to no loss in quality, i.e. halving or quartering the space requirement. These LoRAs function without any problems in ComfyUI.
@civitai3463 can your give us an optimized version of this for chroma please?
This Lora works with Flux kontext which is great for speed. /same settings: 0.125/
Great info, thanks!
It's very hard to control LoRAs when the Hyper LoRA is active, even at 0.125, other LoRAs go completely wild and you lose control. Anyone found a good way to use LoRAs consistently alongside the Hyper one on Forge?
Your Loras works perfectly in Flux dev quantized (gguf - Q3-K_S), truly an incredible result, thank you for the amazing work!
setting: Comfyui
strengh0.125
So this is pretty awesome, it's the Flux analog of DMD2 for SDXL, exactly the same concept. You really need this if you have an older video card like 6GB and still want to run Flux.Dev.
Testing now comparing to DPM++ 2M on well known standard Flux.Dev NSFW models (Q4 GGUF versions to fit my VRAM).
Results are great, but the scheduler matters a lot.
I find that Euler KL-OPTIMAL is far better than Euler Beta (in most cases). These images are on par with the regular model at 20 steps and DPM++ 2M.
Will update this as I do more testing...
Winner: Euler Kl-Optimal (super nice)
Second: Euler Beta
Some models: Euler SGM-Uniform
Details
Files
Hyper-FLUX.1-dev-8steps-lora.safetensors
Mirrors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX-8steps.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
bytedance-hyper-flux-acceleration-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Hyper-FLUX.1-dev-8steps-lora.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.















