Qwen Image Training - The Complete Detailed Guide

Part 1: Initial Setup & Installation

Introduction & Finding Resources
- Locating the main post with all necessary information.
- Finding the specific resources: the zip file, instructions, and the attachments section.
- Speaker's Note: A direct instruction to not start installation immediately and to read the entire post first.
Mandatory Prerequisites: The Requirements Tutorial
- The critical first step: Following the separate "requirements tutorial" linked in the main post.
- This sub-tutorial is presented as a separate video.
- Contents of the requirements tutorial: Installing Python, CUDA, FFmpeg, and other essential software.
- Confirmation that all links in the requirements tutorial are fully updated (example date given: 3 September 2025).
- The explicit instruction to return to the main post after completing the requirements tutorial.
Core Application Installation
- Moving the downloaded .zip file to the desired installation drive (e.g., Q drive).
- Extracting the zip file (using the standard Windows extractor is fine).
- Critical Step: Entering the newly extracted folder before running anything.
- Running the installer by double-clicking Windows install and update.bat (or selecting and hitting Enter).
- Crucial Warning: Do NOT run anything as an administrator, as it will break the installation.
- Explanation of the isolated virtual environment and how it protects your system.
- How to handle installation errors: Select all text in the CMD window (Ctrl+A, Ctrl+C), save to a text file, and send it for support via email, Patreon, or Discord.

Part 2: Acquiring the AI Models

Downloading Training Models
- Running the downloader script: Windows download training models.bat.
- Choosing which base model to download:
  - Option 1: Qwen base model.
  - Option 2: Qwen image edit plus model.
- Features of the custom downloader: High speed (16 simultaneous connections), fully resumable, automatic file merging, and hash verification to prevent corruption.
- Location of downloaded models: training/models/Qwen/ folder.
- Important Model Requirement: BF16 version models are mandatory for training. FP8 or GGUF versions are explicitly stated to not work.

Part 3: The Training Interface & Configuration

Starting and Navigating the UI
- How to update the application: Rerun Windows install and update.bat.
- How to start the application: Run Windows start up.bat.
- The importance of monitoring the CMD window for status and errors.
- Ensuring you are in the "Qwen image training" tab.
- User Interface Tip: Refresh the page before loading a new config to avoid issues.
Loading a Training Configuration (Preset)
- First step in the UI: Loading a configuration file (.toml).
- Navigating to the Qwen training configs folder.
- Choosing a training type: LoRA vs. Fine-tuning (Dreambooth).
- Choosing training duration/quality: 200 epochs (best quality, lower learning rate), 100 epochs, 50 epochs.
- Understanding tiered configs (Tier 1, 2, 3...) based on VRAM requirements.
- How to understand tier differences: Opening the Lora configs explanation.gpx file located inside the configs folder.
System & VRAM Preparation
- How to check total GPU VRAM: nvidia-smi command.
- How to check free VRAM: pip install nvitop and then running nvitop.
- The necessity of minimizing background VRAM usage (e.g., getting it under 2GB) by restarting the PC and disabling unnecessary startup apps in Task Manager.
Detailed Training Parameter Setup
- Accelerate Launch Settings: For multi-GPU training.
- Important Caveat: It's noted that multi-GPU on Windows is not working very well.
- Checkpoints & Output Settings:
  - Setting the Output folder for saved models.
  - Setting the Save every N epochs frequency.
- Definition of Epoch: One epoch is defined as the point where all training images have been trained one time.
- Saving Your Custom Configuration: How to save your modified settings into a new .toml file to use later, and the importance of saving to a new folder to avoid overwriting base configs.

Part 4: Dataset Preparation (The Most Critical Section)

Using the Ultimate Batch Image Processing Tool
- Downloading, installing, and running the separate image processing application.
- Stage 1: Smart Cropping: Using the tool to automatically zoom in on the subject (e.g., "person") via class selection (YOLO) or prompt (SAM 2) without cropping important parts.
- Stage 2: Exact Resizing: Taking the output from Stage 1 and resizing/cropping to the exact target resolution (e.g., 1328x1328).
Dataset Quality Guidelines (For Characters)
- Variety is Key: Include a mix of full body, half body, close-ups, clothing, backgrounds, emotions, poses, and angles.
- Quality is Paramount: Use images with good lighting and sharp focus. Avoid blurry or fuzzy backgrounds.
- Core Principle: The model will learn and repeat whatever is consistently present in the dataset. Only the subject should be consistent.
NEW - Using the Internal Image Pre-processing Tool
- A tab within the main trainer UI named "Image Preprocessing."
- Purpose: To visualize the exact final version of your images as the trainer will see them, especially useful for bucketing (multi-aspect ratio training).
- How to Use: Provide an input folder, enable bucketing, select the architecture (Qwen Image), and process.
- What it reveals: Inaccurate image orientation, padding added by bucketing, and the final resolution distribution.
- Debug Mode: A feature to have images pop up one-by-one during training to see exactly how they are being processed.
Dataset Structuring for the Trainer
- Creating a subfolder in your dataset directory named in the format: [repeats]_[triggerword] (e.g., 1_OHWX).
- Explanation of Repeats: Used to balance datasets with different numbers of images per concept. For a single subject, 1 is sufficient.
Captioning Your Dataset
- Recommended Method: Using only a single trigger word (e.g., OHWX) as the caption. The tutorial states this works best.
- Alternative (Detailed Captions): Using the "Image Captioning" tab (powered by Qwen VL) or the external Joy Caption tool.
- Generating the dataset_config.toml file within the UI is the final step of dataset setup.
Finalizing Model Paths and Settings
- Qwen Image Model Settings: Selecting the Base model, VAE, and Text encoder files.
- Block Swap: Adjusting this value if you get out-of-VRAM errors or slow speeds, as it controls swapping memory with system RAM.
- Training Settings: Setting the Maximum number of epochs (e.g., 200 is good for <50 images).

Part 5: The Training Process & Post-Training Workflow

Monitoring Training & Performance Optimization
- Advanced Speed Improvement Tricks:
  - Connect monitors to a secondary, weaker GPU to free up the primary GPU's idle VRAM.
  - Enable "Use pinned memory for block swapping" in the UI.
  - Disable "Hardware-accelerated GPU scheduling" in Windows Graphics Settings.
  - Overclocking the GPU with MSI Afterburner.
Testing and Finding the Best Checkpoint
- Prerequisite: Setting up Swarm UI with a Comfy UI backend via the linked tutorial.
- Using the Grid Generator tool in Swarm UI to compare multiple checkpoints against a set of test prompts.
- For LoRA: Vary the Lora parameter.
- For Fine-tune: Vary the Model parameter.
- Analyzing the generated grid to visually determine the best-performing checkpoint.
Resuming an Incomplete Training
- For LoRA: Set network_weights to the path of your last checkpoint. Adjust max_train_epochs to the additional number of epochs needed.
- For Fine-tuning: Set the base_model_path to your last fine-tune checkpoint. Adjust max_train_epochs similarly.

Part 6: Generating High-Quality Images

Image Generation Workflow in Swarm UI
- Using presets (e.g., Qwen image realism tier 2).
- Crucial Note for Fine-Tuned Models: After applying a preset, you might need to manually re-select your fine-tuned model, as the preset can override it.
- Pro-Tip: Generate smaller images first without upscaling to quickly find a good seed, then enable upscaling for the one you like.
Fixing Imperfections: Inpainting
- Automatic Face Fixing: Adding segment face to the prompt.
- Manual Inpainting: Using the "Edit Image" feature to manually mask and regenerate a specific area.

Part 7: Specialized Training Scenarios

Fine-Tuning (Dreambooth) Specifics
- Process is nearly identical to LoRA but uses fine tuning configs.
- Key Difference: Checkpoints are massive (~40 GB).
- Post-Training Step: Converting the BF16 models to FP8 scaled (tensor_wise), reducing size by 50%.
Training on the Qwen Image Edit Model
- Without Control Images: A simple process allowing your subject to work with command-based prompts (e.g., "replace his face with OHWX man").
- With Control Images (Teaching New Commands): A complex setup requiring a final image, a caption with the command, and one or more input images. This method uses significantly more VRAM and is much slower.
Style Training
- The key is the dataset: extremely consistent style with high variety in subjects (nothing should repeat except the style).
- Using presets like Qwen images stylized UHD tier 1 (better quality, more steps) or tier 2 (faster) for generation.
Product Training
- The key is the dataset: a mix of close-up shots (for details like text) and wider shots (for scale).
- Use detailed prompts during inference that describe logos and text on the product for higher accuracy.

Part 8: Conclusion & Community

Final Housekeeping & Recap
- LoRAs go in the SwarmUI/models/lora folder.
- Fine-tuned models go in the SwarmUI/models/diffusion folder after FP8 conversion.
Community and Support Channels
- Joining the Discord server.
- Following on GitHub (fork, star, watch, sponsor).
- Joining the Reddit community.
- Following on LinkedIn and subscribing on YouTube (and enabling notifications).
- Information about professional services: private lectures and company consultations.