Home Creators Posts Import Register Favorites Logout

Content

NVFP4 models run over 100% faster in ComfyUI and SwarmUI with CUDA 13; this full tutorial compares NVFP4 quality on FLUX 2, Z Image Turbo and FLUX 1 in grids.

Shows how to upgrade ComfyUI and SwarmUI to CUDA 13 and Torch 2.9.1 with precompiled Flash Attention, Sage Attention and xFormers for Windows and Linux, from GTX 1650 to RTX 5000 series GPUs. Also introduces SimplePod, a new cheap cloud GPU service that works like RunPod.

Full tutorial: https://www.youtube.com/watch?v=yOj9PYq3XYM

NVFP4 With CUDA 13 Full Tutorial, 100%+ Speed Gain + Quality Comparison & New Cheap Cloud SimplePod

Finally NVFP4 models has arrived to ComfyUI thus SwarmUI with CUDA 13. NVFP4 models are literally 100%+ faster with minimal impact on quality. I have done grid quality comparison to show you the difference on FLUX 2, Z Image Turbo and FLUX 1 of NVFP4 versions. To make CUDA 13 work, I have compiled Flash Attention, Sage Attention & xFormers for both Windows and Linux with all of the CUDA archs to support literally all GPUs starting from GTX 1650 series, RTX 2000, 3000, 4000, 5000 series and more.

In this full tutorial, I will show you how to upgrade your ComfyUI and thus SwarmUI to use latest CUDA 13 with latest libraries and Torch 2.9.1. Moreover, our compiled libraries such as Sage Attention works with all models on all GPUs without generating black images or videos such as Qwen Image or Wan 2.2 models. Hopefully LTX 2 presets and tutorial coming soon too. Finally, I introduce a new private cloud GPU platform called as SimplePod like RunPod. This platform has all the features of RunPod same way but much faster and cheaper.

📂 Resources & Links:

⏱️ Video Chapters:

  • 00:00:00 New ComfyUI installer (CUDA 13, Torch 2.9.1, Triton + attention libs)

  • 00:00:19 NVFP4 speedup claims vs real tests; why CUDA 13 enables new models

  • 00:00:34 Prebuilt FlashAttention/SageAttention/xFormers for many GPUs (Windows + Linux)

  • 00:01:00 Quality roadmap: FLUX2 Dev, Z Image Turbo, FLUX Dev (BF16/FP8/GGUF/NVFP4)

  • 00:01:23 Downloader adds NVFP4: FLUX2 Dev, FLUX Dev (Context/Dev), Z Image Turbo

  • 00:01:51 SimplePod AI intro: RunPod-style pods, cheaper rates, permanent storage

  • 00:02:36 Musubi Tuner FP8 Scaled: quality myths vs GGUF + why scaled matters

  • 00:03:10 Quantization & precision (FP32/BF16/FP8/GGUF) + Qwen3 low-VRAM encoders

  • 00:03:34 ComfyUI v73 zip: CUDA 13 included; update NVIDIA drivers only (v72 deprecated)

  • 00:04:13 Update steps: overwrite zip, delete venv, run install/update .bat

  • 00:05:02 Python: 3.10 recommended (supports 3.10-3.13); fresh vs update

  • 00:06:02 New installer flow: uv speed, standalone use, backend libs detected

  • 00:07:12 Stability flags: --cache-none vs --disable-smart-memory (OOM/stuck fixes)

  • 00:07:54 SwarmUI presets: 32 presets supported; drag/drop + auto model downloader

  • 00:08:25 Update SwarmUI model-downloader zip (extract + overwrite)

  • 00:08:49 Download bundles/models (Z Image Turbo Core + NVFP4 options)

  • 00:09:25 Update/launch SwarmUI; point to updated ComfyUI backend + set args

  • 00:10:32 Live gen test: Z Image Turbo BF16 @1536x1536

  • 00:11:29 Switch to NVFP4: VRAM cache behavior; 1024x1024

  • 00:12:36 FLUX2 Dev quality: FP8 Scaled vs NVFP4 side-by-side comparisons

  • 00:13:33 Speed chart: FLUX2 NVFP4 about 193% faster than FP8 Scaled

  • 00:14:10 Z Image Turbo quality: BF16 vs NVFP4 vs FP8 Scaled (quant method)

  • 00:15:25 FLUX Dev: FP8 Scaled approx GGUF Q8; NVFP4 currently shows degradation

  • 00:16:45 What precision means + model size examples (FP32/BF16/FP8 Scaled/NVFP4)

  • 00:18:07 Practical recommendations: BF16 best; avoid FP16; raw FP8 vs FP8 Scaled

  • 00:19:43 GGUF explained: block quant, slower runtime; use only when RAM is too low

  • 00:21:36 Precision hierarchy recap + when to pick FP8 mixed/scaled over GGUF

  • 00:21:58 SimplePod setup: register, add credits, open template link

  • 00:22:31 Template config + RunPod price comparison (disk, ports, GPU selection)

  • 00:24:02 Persistent volume: create + mount to /workspace

  • 00:25:11 Launch RTX Pro 6000 pod; SimplePod vs RunPod pricing differences

  • 00:26:29 Temp vs persistent disk: deleting instance wipes temp data - backup!

  • 00:26:55 JupyterLab: upload zips, apt install zip, unzip ComfyUI in workspace

  • 00:27:48 Run install script; unzip SwarmUI; start the model downloader

  • 00:29:02 Downloader path for ComfyUI + folder structure; download Z Image Turbo bundle

  • 00:30:08 Start ComfyUI; confirm CUDA 13 + Torch 2.9.1; connect via port 3000 Direct

  • 00:31:08 Preset demo: Z Image Turbo Quality 1; fix VAE path; monitor VRAM

  • 00:33:18 File Browser Direct: download outputs/models fast; upload files back

  • 00:34:41 Restart server; install/start SwarmUI; open Cloudflared URL

  • 00:36:26 SwarmUI backend: /workspace/ComfyUI/main.py + args; import presets

  • 00:37:27 Download FLUX2 Core + NVFP4; share model paths between SwarmUI & ComfyUI

  • 00:39:27 FLUX2 NVFP4 generation @2048x2048; VRAM usage + step speed

  • 00:40:43 Cloud GPU pitfall: diagnosing a power-capped GPU

  • 00:41:28 Resume: re-run template w/ volume; reconnect fast

  • 00:45:02 Wrap-up: SimplePod pros (direct/secure, cheaper storage)

Screenshots and Info

New_ComfyUI

New_ComfyUI_2

Model_Speeds

ComfyUI_Features

our_supported_gpu_list

1410001-slighty smiling photo of ohwx man wearin-Qwen_Base_2512_Model_FineTuning-2

Related posts

All posts in this topic: ComfyUI: 1-Click Installer, Presets and Tutorials · New to SECourses? Start Here: Best and Most Popular SECourses Posts

Files

NVFP4 With CUDA 13 Full Tutorial, 100%+ Speed Gain + Quality Comparison & New Cheap Cloud SimplePod

Finally NVFP4 models has arrived to ComfyUI thus SwarmUI with CUDA 13. NVFP4 models are literally 100%+ faster with minimal impact on quality. I have done grid quality comparison to show you the difference on FLUX 2, Z Image Turbo and FLUX 1 of NVFP4 versions. To make CUDA 13 work, I have compiled Flash Attention, Sage Attention & xFormers for both Windows and Linux with all of the CUDA archs to support literally all GPUs starting from GTX 1650 series, RTX 2000, 3000, 4000, 5000 series and more. In this full tutorial, I will show you how to upgrade your ComfyUI and thus SwarmUI to use latest CUDA 13 with latest libraries and Torch 2.9.1. Moreover, our compiled libraries such as Sage Attention works with all models on all GPUs without generating black images or videos such as Qwen Image or Wan 2.2 models. Hopefully LTX 2 presets and tutorial coming soon too. Finally, I introduce a new private cloud GPU platform called as SimplePod like RunPod. This platform has all the features of RunPod same way but much faster and cheaper. 📂 Resources & Links: ComfyUI Installers: [ https://www.patreon.com/posts/ComfyUI-Installers-105023709 ] SimplePod: [ https://simplepod.ai/ref?user=secourses ] SwarmUI Installer, Model Auto Downloader and Presets: [ https://www.patreon.com/posts/SwarmUI-Install-Download-Models-Presets-114517862 ] How to Use SwarmUI Presets & Workflows in ComfyUI + Custom Model Paths Setup for ComfyUI & SwarmUI Tutorial: [ https://youtu.be/EqFilBM3i7s ] SECourses Discord Channel for 7/24 Support: [ https://discord.com/invite/software-engineering-courses-secourses-772774097734074388 ] NVIDIA NVFP4 Blog Post More: [ https://developer.nvidia.com/blog/introducing-nvfp4-for-efficient-and-accurate-low-precision-inference/ ] ⏱️ Video Chapters: 0:00 New ComfyUI installer (CUDA 13, Torch 2.9.1, Triton + attention libs) 0:19 NVFP4 speedup claims vs real tests; why CUDA 13 enables new models 0:34 Prebuilt FlashAttention/SageAttention/xFormers for many GPUs (Windows + Linux) 1:00 Quality roadmap: FLUX2 Dev, Z Image Turbo, FLUX Dev (BF16/FP8/GGUF/NVFP4) 1:23 Downloader adds NVFP4: FLUX2 Dev, FLUX Dev (Context/Dev), Z Image Turbo 1:51 SimplePod AI intro: RunPod-style pods, cheaper rates, permanent storage 2:36 Musubi Tuner FP8 Scaled: quality myths vs GGUF + why scaled matters 3:10 Quantization & precision (FP32/BF16/FP8/GGUF) + Qwen3 low-VRAM encoders 3:34 ComfyUI v73 zip: CUDA 13 included; update NVIDIA drivers only (v72 deprecated) 4:13 Update steps: overwrite zip, delete venv, run install/update .bat 5:02 Python: 3.10 recommended (supports 3.10-3.13); fresh vs update 6:02 New installer flow: uv speed, standalone use, backend libs detected 7:12 Stability flags: --cache-none vs --disable-smart-memory (OOM/stuck fixes) 7:54 SwarmUI presets: 32 presets supported; drag/drop + auto model downloader 8:25 Update SwarmUI model-downloader zip (extract + overwrite) 8:49 Download bundles/models (Z Image Turbo Core + NVFP4 options) 9:25 Update/launch SwarmUI; point to updated ComfyUI backend + set args 10:32 Live gen test: Z Image Turbo BF16 @1536x1536 11:29 Switch to NVFP4: VRAM cache behavior; 1024x1024 12:36 FLUX2 Dev quality: FP8 Scaled vs NVFP4 side-by-side comparisons 13:33 Speed chart: FLUX2 NVFP4 about 193% faster than FP8 Scaled 14:10 Z Image Turbo quality: BF16 vs NVFP4 vs FP8 Scaled (quant method) 15:25 FLUX Dev: FP8 Scaled approx GGUF Q8; NVFP4 currently shows degradation 16:45 What precision means + model size examples (FP32/BF16/FP8 Scaled/NVFP4) 18:07 Practical recommendations: BF16 best; avoid FP16; raw FP8 vs FP8 Scaled 19:43 GGUF explained: block quant, slower runtime; use only when RAM is too low 21:36 Precision hierarchy recap + when to pick FP8 mixed/scaled over GGUF 21:58 SimplePod setup: register, add credits, open template link 22:31 Template config + RunPod price comparison (disk, ports, GPU selection) 24:02 Persistent volume: create + mount to /workspace 25:11 Launch RTX Pro 6000 pod; SimplePod vs RunPod pricing differences 26:29 Temp vs persistent disk: deleting instance wipes temp data - backup! 26:55 JupyterLab: upload zips, apt install zip, unzip ComfyUI in workspace 27:48 Run install script; unzip SwarmUI; start the model downloader 29:02 Downloader path for ComfyUI + folder structure; download Z Image Turbo bundle 30:08 Start ComfyUI; confirm CUDA 13 + Torch 2.9.1; connect via port 3000 Direct 31:08 Preset demo: Z Image Turbo Quality 1; fix VAE path; monitor VRAM 33:18 File Browser Direct: download outputs/models fast; upload files back 34:41 Restart server; install/start SwarmUI; open Cloudflared URL 36:26 SwarmUI backend: /workspace/ComfyUI/main.py + args; import presets 37:27 Download FLUX2 Core + NVFP4; share model paths between SwarmUI & ComfyUI 39:27 FLUX2 NVFP4 generation @2048x2048; VRAM usage + step speed 40:43 Cloud GPU pitfall: diagnosing a power-capped GPU 41:28 Resume: re-run template w/ volume; reconnect fast 45:02 Wrap-up: SimplePod pros (direct/secure, cheaper storage)

Comments

Tyler

Is there an NVFP4 for SRPO?

David

He said not for now.

SECourses

i didnt find it sadly. i am gonna look for how to convert ourselves if possible

SECourses

true not yet sadly but i am expecting soon hopefully

SECourses

true not yet sadly but i am expecting soon hopefully

SECourses

started searching how to generate NVFP4. so we can add into our app hopefully

SECourses

yep working. i tested on few loras :D

David

Okkk good, we wait your updates 👍🏽

Tyler

Wait, did you already add the SRPO!? I see it in your notes on the email that went out today. Can the NFP4 SRPO model be used as plug and playu for previous comfy workflows that used the SRPO model?

SECourses

Yes I added. Hopefully will add v2 even better quality working on it

Tyler

Amazing! Is the model available to download solo or only through your auto downloader?

SECourses

it is available in downloader you can download there solo. also i am compiling NVFP4 Mixed precision version atm. it will have hopefully better quality same speed. will update downloader app once ready hopefully

Roni Bliss

NVFP4 only works on Blackwell chips brother

SECourses

true it brings speed only on Blackwell RTX 5000 series at the moment.

Roni Bliss

Also, I don't want to sound like I am digging at you, I love your work you are the best in the busines and so valuble to the community, but just need for others to understand. It seems that since I updated my ComfyUI to PT 129 & Cude 130 many many nodes are broken (RELSFY etc) this is due to flash-attention-cuda not aligning with the new cuda/PT. https://github.com/Dao-AILab/flash-attention - Needs update before many of these nodes will be working again Just a heads up people need to know! -- This is very cutting edge haha

SECourses

i compiled flash attention with cuda 13 and i tested it works. perhaps your custom node broken. i tested it with RELSFY . for example i generated images with flash attention with qwen on 5090 and 3090.

JS

What 5090 driver are you using? Are there any other updates required? The last update to Cuda and python I have done was over a year ago, I don't know if other updates are required between now and this most recent update or if this update described in this video will cover everything needed?

SECourses

currently nvidia driver should be sufficient. i am using Driver Version: 591.44 but newer can be used too