Hyper Flux 8 step Lora - CivArchive (CivitAI Archive)

Hyper Flux 8 step Lora - v1.0

NSFW

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)

yLoraNov 17, 2024

CivitAI

Great... Now we need to find a way to quantize LoRAs

BilboTagginsJan 23, 2025

I think this may be what you are looking for: https://civitai.com/models/693544/hypersd-flux1-dev-bytedance-optimized-size?modelVersionId=776158

civitai3463Oct 1, 2025

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.

in_the_zone_againMar 28, 2026

@civitai3463 can your give us an optimized version of this for chroma please?

flo11ok874Jun 27, 2025· 4 reactions

CivitAI

This Lora works with Flux kontext which is great for speed. /same settings: 0.125/

Creepybit

Author

Jun 27, 2025· 1 reaction

Great info, thanks!

SassyAIAug 27, 2025· 1 reaction

CivitAI

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?

Daddy_WaifuDec 18, 2025

CivitAI

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

ferrrett33Apr 13, 2026

CivitAI

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

LORA

Flux.1 D

by Creepybit

Download (Beta) View on CivitAI

assets

Details

Downloads

4,074

Platform

CivitAI

Platform Status

Available

Created

10/12/2024

Updated

8/18/2026

Deleted

Files

Hyper-FLUX.1-dev-8steps-lora.safetensors

Size:

1.29 GB

SHA256:

e0ab0fdf569cd01a382f19bd87681f628879dea7ad51fe5a3799b6c18c7b2d03