Home Creators Posts Import Register Favorites Logout

Content

Qwen Image 2512 is the new king for training: the video compares trained Base, 2512, Edit 2509 and Edit 2511 models and shows a big leap in realism.

Also covers the new FP8 quantization in SECourses Musubi Tuner, Qwen Image 2512 UHD realism presets, and all SwarmUI presets converted into drag-and-drop ComfyUI workflows with models from the 1-click downloader.

Full tutorial: https://www.youtube.com/watch?v=RcoXd9v1t_c

Qwen Image 2512 & 2511 Training Results Are Next Level + 33 ComfyUI Presets for Image & Video Gen

Qwen Image 2512 Text to Image model is a massive upgrade in quality just as Qwen Image Edit 2511 was for image editing tasks based on commands. I have trained both Qwen Image older Base, Qwen Image 2512 new base, Qwen Image 2509 older edit model and Qwen Image 2511 newer edit model and compared them in this video. The results are astonishing. Moreover, we have converted all of our premium SwarmUI image and video generation presets into ComfyUI workflows. Just drag and drop the workflow and start using immediately. All models are downloaded with our premium 1-click to use model downloader app. ComfyUI and SwarmUI are also installed with 1-click installers and our ComfyUI fully supports Sage Attention, Flash Attention, xFormers, Triton and more with RTX 5000 series and all GPUs like 3000, 2000, 4000 series, etc for both Linux and Windows.

📂 Resources & Links:

⏱️ Video Chapters:

  • 00:00:00 Introduction to Qwen Image Base & 2512 Model Training Findings plus Qwen Image Edit 2509 & 2511 News

  • 00:00:16 Announcement of Training Support for Qwen Image Edit Models & Upcoming Training Guide Overview

  • 00:00:32 New FP8 Quantization Feature in SECourses Musubi Tuner & Full ComfyUI Preset Support Announcement

  • 00:00:48 Converting All SwarmUI Presets into Drag-and-Drop ComfyUI Workflows & HTML Helper File Introduction

  • 00:01:10 How to Use the New HTML File to Identify and Download Necessary Model Bundles via SwarmUI Downloader

  • 00:01:24 Qwen Image Base Model Training Analysis: Observing Overfitting and Quality Degradation at 240 Epochs

  • 00:02:04 Qwen Image 2512 Model Training Results: Massive Quality Boost & Reduced Noise Compared to Base Model

  • 00:02:38 Showcasing New Qwen Image 2512 UHD Realism Presets Available in Both ComfyUI and SwarmUI

  • 00:03:01 Visual Comparison: The Massive Leap in Realism Quality Between Previous Models and New 2512 Model

  • 00:03:38 Instructions on Updating the Training Application via Bat File to Unlock New Qwen Model Support

  • 00:04:03 Selecting Specific Model Versions in the Trainer: Text-to-Image vs Image Edit Plus 2509/2511 Variants

  • 00:04:24 Using the Updated Model Downloader Application to Fetch New Qwen 2512 & 2511 Checkpoints Automatically

  • 00:04:40 Wan 2.2 Training Tutorial Reference & Overview of Recently Updated High-Quality Pr- sets for All Models

  • 00:05:05 Training Dataset Configuration: Using Medium Quality Settings & OHWX Caption St- ategy for Best Results-

  • [00:05:22](https://youtu- be/RcoXd9v1t_c?t=322) Testing Qwen Image Edit 2509 vs 2511 Models with GTA 5 Style Transfer Training on 120 and 180 Epochs

  • 00:05:46 Analyzing GTA 5 Style Reproduction Accuracy on Both Edit Models Without Using Any Control Images

  • 00:06:11 Style Transfer Experiment: Converting Input Image to Learned Style Using Edit Models Without Image-to-Image

  • 00:06:40 Critical Findings on Edit Models: Lower Epoch Requirements & How Overtraining Alters Input Image Fidelity

  • 00:07:09 Resolution Limitations for Style Transfer: Why 1328x1328 is the Optimal Sweet Spot for Editing Tasks

  • 00:07:30 Reviewing Specific Presets Used for Qwen Image Edit Text-to-Image Generation vs Image Editing Tasks

  • 00:08:02 How to Install and Update ComfyUI with the Latest Standalone Installer Script & Overwriting Old Files

  • 00:08:15 Downloading Model Bundles for ComfyUI Presets Using the SwarmUI Downloader Tool & HTML Guide

  • 00:08:39 Navigating the HTML Guide to Select Correct Model Bundles like Flux 2, Z Image Turbo or Qwen Core

  • 00:09:12 Understanding Preset Limitations: Why Inpainting & Outpainting Workflows Require SwarmUI Editing Tab

  • 00:09:57 Live Demonstration: Dragging and Dropping Flux SRPO Preset into ComfyUI Interface for Instant Setup

  • 00:10:11 Configuring the Model Downloader to Target Your Specific ComfyUI Installation Path for Direct Downloads

  • 00:10:47 Handling Potential VAE Subfolder Issues & Verifying High-Speed Downloads with 100% Hash Verification

  • 00:11:06 Generating Images in ComfyUI: Python 3.10-3.13 Support, RTX 5000/4000/3000 Compatibility & Speed

  • 00:11:54 Recommendation to Use SwarmUI with ComfyUI Backend for Best Experience & Access to All Features

  • 00:12:20 Upcoming Tutorials Teaser: Trellis 2 3D Models & Video/Image Upscaler Guides Coming Soon

Some Generated Images

01

Related posts

All posts in this topic: LoRA and Fine-Tuning Training: Kohya, Musubi Tuner, OneTrainer, AI Toolkit · New to SECourses? Start Here: Best and Most Popular SECourses Posts

Files

Qwen Image 2512 & 2511 Training Results Are Next Level + 33 ComfyUI Presets for Image & Video Gen

Qwen Image 2512 Text to Image model is a massive upgrade in quality just as Qwen Image Edit 2511 was for image editing tasks based on commands. I have trained both Qwen Image older Base, Qwen Image 2512 new base, Qwen Image 2509 older edit model and Qwen Image 2511 newer edit model and compared them in this video. The results are astonishing. Moreover, we have converted all of our premium SwarmUI image and video generation presets into ComfyUI workflows. Just drag and drop the workflow and start using immediately. All models are downloaded with our premium 1-click to use model downloader app. ComfyUI and SwarmUI are also installed with 1-click installers and our ComfyUI fully supports Sage Attention, Flash Attention, xFormers, Triton and more with RTX 5000 series and all GPUs like 3000, 2000, 4000 series, etc for both Linux and Windows. 📂 Resources & Links: ComfyUI Installer: [ https://www.patreon.com/posts/ComfyUI-Installers-105023709 ] SwarmUI Installer, Model Auto Downloader and Presets: [ https://www.patreon.com/posts/SwarmUI-Install-Download-Models-Presets-114517862 ] SwarmUI & ComfyUI Setup Guide for Windows: [ https://youtu.be/c3gEoAyL2IE ] SwarmUI & ComfyUI Setup Guide for RunPod & Massed Compute: [ https://youtu.be/bBxgtVD3ek4 ] Qwen Image Models Training Tutorial: [ https://youtu.be/DPX3eBTuO_Y ] Wan 2.2 Model Training Tutorial: [ https://youtu.be/ocEkhAsPOs4 ] Z Image Turbo Model Training Tutorial: [ https://youtu.be/ezD6QO14kRc ] FLUX Models Fine Tuning / DreamBooth Training Tutorial: [ https://youtu.be/FvpWy1x5etM ] FLUX Models LoRA Training Tutorial: [ https://youtu.be/nySGu12Y05k ] Upload / Download Big Files Guide for RunPod & Massed Compute: [ https://youtu.be/X5WVZ0NMaTg ] ⏱️ Video Chapters: 0:00 Introduction to Qwen Image Base & 2512 Model Training Findings plus Qwen Image Edit 2509 & 2511 News 0:16 Announcement of Training Support for Qwen Image Edit Models & Upcoming Training Guide Overview 0:32 New FP8 Quantization Feature in SECourses Musubi Tuner & Full ComfyUI Preset Support Announcement 0:48 Converting All SwarmUI Presets into Drag-and-Drop ComfyUI Workflows & HTML Helper File Introduction 1:10 How to Use the New HTML File to Identify and Download Necessary Model Bundles via SwarmUI Downloader 1:24 Qwen Image Base Model Training Analysis: Observing Overfitting and Quality Degradation at 240 Epochs 2:04 Qwen Image 2512 Model Training Results: Massive Quality Boost & Reduced Noise Compared to Base Model 2:38 Showcasing New Qwen Image 2512 UHD Realism Presets Available in Both ComfyUI and SwarmUI 3:01 Visual Comparison: The Massive Leap in Realism Quality Between Previous Models and New 2512 Model 3:38 Instructions on Updating the Training Application via Bat File to Unlock New Qwen Model Support 4:03 Selecting Specific Model Versions in the Trainer: Text-to-Image vs Image Edit Plus 2509/2511 Variants 4:24 Using the Updated Model Downloader Application to Fetch New Qwen 2512 & 2511 Checkpoints Automatically 4:40 Wan 2.2 Training Tutorial Reference & Overview of Recently Updated High-Quality Presets for All Models 5:05 Training Dataset Configuration: Using Medium Quality Settings & OHWX Caption Strategy for Best Results 5:22 Testing Qwen Image Edit 2509 vs 2511 Models with GTA 5 Style Transfer Training on 120 and 180 Epochs 5:46 Analyzing GTA 5 Style Reproduction Accuracy on Both Edit Models Without Using Any Control Images 6:11 Style Transfer Experiment: Converting Input Image to Learned Style Using Edit Models Without Image-to-Image 6:40 Critical Findings on Edit Models: Lower Epoch Requirements & How Overtraining Alters Input Image Fidelity 7:09 Resolution Limitations for Style Transfer: Why 1328x1328 is the Optimal Sweet Spot for Editing Tasks 7:30 Reviewing Specific Presets Used for Qwen Image Edit Text-to-Image Generation vs Image Editing Tasks 8:02 How to Install and Update ComfyUI with the Latest Standalone Installer Script & Overwriting Old Files 8:15 Downloading Model Bundles for ComfyUI Presets Using the SwarmUI Downloader Tool & HTML Guide 8:39 Navigating the HTML Guide to Select Correct Model Bundles like Flux 2, Z Image Turbo or Qwen Core 9:12 Understanding Preset Limitations: Why Inpainting & Outpainting Workflows Require SwarmUI Editing Tab 9:57 Live Demonstration: Dragging and Dropping Flux SRPO Preset into ComfyUI Interface for Instant Setup 10:11 Configuring the Model Downloader to Target Your Specific ComfyUI Installation Path for Direct Downloads 10:47 Handling Potential VAE Subfolder Issues & Verifying High-Speed Downloads with 100% Hash Verification 11:06 Generating Images in ComfyUI: Python 3.10-3.13 Support, RTX 5000/4000/3000 Compatibility & Speed 11:54 Recommendation to Use SwarmUI with ComfyUI Backend for Best Experience & Access to All Features 12:20 Upcoming Tutorials Teaser: Trellis 2 3D Models & Video/Image Upscaler Guides Coming Soon

Comments

David

I was doing a LoRA training with Qwen 2512 on an RTX 5090 and kept hitting a CUDA illegal memory access crash around step ~70. The issue was caused by torch.compile / TorchDynamo with dynamic shapes becoming unstable in the attention path. Disabling torch.compile (and dynamic shapes) fixed the problem completely. After that, the training runs stable well past the original crash point, can be useful for someone 😬 9 hours and it's done. I haven't applied further optimizations yet, but I'll experiment later to see if the training time be reduced 👍🏽

Klaus The.

This is absolutely amazing. The advantage in image quality is a big step forward. You think LORA training for Qwen 2512 needs completly new parameters/settings or can they be trained with same settings as previous 2509 model?

SECourses

thanks for info. i hope this issue gets fixed

David

Another issue: during the same training I ran into an apparent freeze at epoch transitions (GPU going idle, no new logs). On Windows this seems related to DataLoader workers, especially with persistent workers enabled. Disabling persistent DataLoader workers and reducing data loading workers to 1 resolved the hang so far. Still monitoring to confirm it’s fully stable through the entire training. Of course ChatGPT is helping me with these settings, I update you, thank you.

SECourses

i didnt have both of the issues before but it is great you are solving them

David

Yes I believe you, it's Windows... every system is different! Even with same windows version 🤣

David

Update: after changing several settings last night to improve stability (disabling torch.compile, adjusting DataLoader workers, using SDPA, and enabling block swapping), the training ran much further but still hit a CUDA illegal memory access later on during the night (around epoch ~87 / step ~2170). This suggests the issue is not only an early-stage configuration problem, but likely related to long-run kernel stability (possibly FP8 on RTX 5090). I’m continuing to test resumes and configurations to see how late the failure can be pushed or fully eliminated.

David

Testing the LoRA checkpoint at epoch 75 already produced very good results, so the training was clearly progressing well before the crash.

Julius

Are the image examples shown here from LoRA or full finetuning? What tier of your configs?

SECourses

These are bf16 fine tuning. but fp8 and lora are almost same. i used 200 epoch : Tier1_29400_MB.toml

SECourses

i wonder if that is the result weights getting exploded

SECourses

so weird to fail after that long

David

I’m not sure if it’s exploding weights. What I can say is that the training got stuck/crashed multiple times at different points. Each time I just resumed from the latest available epoch/checkpoint and continued. In the end I managed to complete the training by doing it in multiple passes, always restarting from the last saved epoch when it stopped.

SECourses

interesting i wonder what causes

geddon

I'm running Qwen_Image_2512_Quant_FP8_Scaled on my RTX 5090 and getting nothing but static with the Presets. Any suggestions?

SECourses

you are not using our comfyui installer. make a fresh install with python 3.10 and use it as a backend. use latest zip file

geddon

Deleting my old Comfy_UI folder and installing it fresh did the trick. Thanks!

David

Since I'm still hitting frequent freezes/crashes on Windows, I'm planning to try the training on Linux to see if stability improves. Do you have a Linux distro you'd recommend for this setup, and can your Windows installation scripts be easily adapted for Linux?

SECourses

yes for training linux is definitely faster. you can use ubuntu 22 or ubuntu 24. my massed compute scripts are linux and you can use them. just make sure to change python code to python 3.10 venv

David

Thanks a lot for the clarification! I'll try moving the training to Linux (Ubuntu 22). I'll follow your Linux scripts and set everything up with Python 3.10 in a venv. Let's see if this improves stability 😅 I will let you know.

SECourses

sure let me know. i am expecting really good gains

David

Update!!! I finally completed the training successfully by moving everything to Ubuntu. Dataset: 30 images, 6000 steps. Total training time: ~6 hours 20 minutes, average speed ~3.8 s/step. On Windows the same setup was slower (~5.7–7 s/step) and I kept hitting random CUDA crashes. On Ubuntu the training was fully stable from start to finish. I used your Linux setup approach, Python 3.10 venv, and after fixing a couple of missing parts in the install script, everything worked perfectly. Thanks for the guidance! Linux definitely made the difference 🤩🤩🤩 If you want furthermore info just contact me, in pvt too. Thank you boss 😬

David

One minor issue on Linux it’s hardware-related, in my specific case. My GPU AIO uses proprietary fans/controllers (Lian Li), so on Linux I can’t control those fans yet. Because of that I capped the GPU to 400W to keep temperatures stable during training. With this setup the GPU stayed around ~70°C, while on Windows (with full fan control) it usually stays around ~55°C even under heavy load. The training still completed successfully, and once I can fully control the fans via USB, I may be able to remove the power cap and squeeze out a bit more training speed as well 😬

David

Oh, that was the exact issues I had! Nvidia...😅

SECourses

yep. i wish we could get more community pressure for them to fix

David

Yes I agree. Anyway Ubuntu is not that bad, only bad thing is that it's not fully supported from a lot of company.

SECourses

true. but i NVIDIA fixes the issue Windows will be really good too :D

David

That's sure😅🤣

David

I have 2 questions: 1) Why is it that when I use a LoRA, even with a low weight like 0.7, it still makes other people in the image resemble that LoRA’s face? 2) When I apply two LoRAs, why does it blend them together, like one’s hair and the other’s face, instead of keeping them distinct? I trained with Qwen 2512.

SECourses

due to this models training internal structure. they bleed into all tokens of trained images even if you dont caption. that is why we just use ohwx it will both bleed and blend and sometimes not. sadly no way to fix this issue atm except during training have 2 person in same photo for it to learn seperately

David

Perfect, thanks for the explanation 👍🏽

Michael

Would you say it's better training Qwen 2512 with about 60-80 high quality dataset, or WAN 2.2 - what you think which models performs over all better? when model size and generation speed are irrelevant, because renting GPUs?

SECourses

Qwen 2512 is better for image training.

Taiga

GEE, THANKS A BUNCH. Finally worked out the python11 switch and your installer was screwed anyway! What, now I have to subscribe again just to get the fix. 26 May 2026 - Version 12.0 With new zip file below errors fixed Token: not provided Note: pyannote gated files require --token or HF_TOKEN on first download. [ERROR] Dependency source does not expose pytorch_model.bin: MonsterMMORPG/Wan_GGUF/pyannote_segmentation3 2 May 2026 - Version 11.1 Some installer bugs fixed Enable Background Music Remover Filter - fixed Enable Silero VAD Filter - fixed

SECourses

welcome back. if any issues let me know i shall fix ASAP. sorry for errors