This upload serves as a demonstration of advanced image generation techniques using the publicly available svdq-int4-flux.1-dev model (original source: https://huggingface.co/mit-han-lab/svdq-int4-flux.1-dev/tree/main) in conjunction with my Finesse2FP16Lora and NSFW2FP16Lora. It generates 5X FASTER images than Flux1.Dev while using less than 8GB VRAM on budget GPUs.
Please note I am not the original creator of svdq-int4-flux.1-dev. The key innovation here is the application of SWDQuant (https://github.com/mit-han-lab/nunchaku), which optimizes inference for this model far beyond simple quantization, ensuring the preservation of the original flux1-dev quality. Nunchaku isn't just another Flux model accelerator—it's a paradigm shift. The presence of a separate BF16 preserved layers file is evidence of this careful optimization.
The showcased images were generated using this optimized svdq-int4-flux.1-dev with my Finesse2FP16 Lora in SwarmUI, 20 steps, CFG 1 To replicate these results, please create a diffusion_models/svdq-int4-flux.1-dev subdirectory and place the two .json and two .safetensors files from the provided Hugging Face link inside it.
On my Nvidia 3060, generating an image with 20 steps takes only about 50 seconds (flux1.dev-fp8 takes about 300 seg, 6x more). This represents a significant leap in efficiency (speed and size) without any noticeable degradation in quality or the characteristic flux1-dev aesthetic.
This upload serves as a demonstration of advanced image generation techniques using the publicly available
svdq-int4-flux.1-devmodel (original source: https://huggingface.co/mit-han-lab/svdq-int4-flux.1-dev/tree/main) in conjunction with my Finesse2FP16Lora and NSFW2FP16Lora. It generates 5X FASTER images than Flux1.Dev while using less than 8GB VRAM on budget GPUs.Please note I am not the original creator of
svdq-int4-flux.1-dev.The key innovation here is the application of SWDQuant (https://github.com/mit-han-lab/nunchaku), which optimizes inference for this model far beyond simple quantization, ensuring the preservation of the original
flux1-devquality. Nunchaku isn't just another Flux model accelerator—it's a paradigm shift.The presence of a separate BF16 preserved layers file is evidence of this careful optimization.
The showcased images were generated using this optimized
svdq-int4-flux.1-devwith my Finesse2FP16 Lora in SwarmUI, 20 steps, CFG 1To replicate these results, please create a
diffusion_models/svdq-int4-flux.1-devsubdirectory and place the two.jsonand two.safetensorsfiles from the provided Hugging Face link inside it.On my Nvidia 3060, generating an image with 20 steps takes only about 50 seconds (flux1.dev-fp8 takes about 300 seg, 6x more). This represents a significant leap in efficiency (speed and size) without any noticeable degradation in quality or the characteristic
flux1-devaesthetic.If you want to install SwarmUI: In X:\ (not in C:\Program files) put https://github.com/mcmonkeyprojects/SwarmUI/releases/download/0.9.5-Beta/install-windows.bat into it, then execute it, will create a X:\SwarmUI. Wait some minutes.
A guide for ComfyUI users https://www.reddit.com/r/StableDiffusion/comments/1j7dzhe/nunchaku_v014_svdquant_comfyui_portable/