# Model Training Experiments and Full Guide and Tutorial: Fine-Tuning vs LoRA Comparison This repository contains experimental results comparing Fine-Tuning/DreamBooth and LoRA training approaches. I am sharing how I trained this model with full details and even the dataset: please read entire post very carefully. This model is purely trained for educational and research purposes only for SFW and ethical image generation. The workflow and the config used in this tutorial can be used to train clothing, items, animals, pets, objects, styles, simply anything. The uploaded images have SwarmUI metadata and can be re-generated exactly. For generations FP16 model used but FP8 should yield almost same quality. Don't forget to have used yolo face masking model in prompts. ## How To Use Download model into diffusion_models of the SwarmUI. Then you need to use Clip-L and T5-XXL models as well. I recommend T5-XXL FP16 or Scaled FP8 version. A newest fully public tutorial here for how to use : https://youtu.be/-zOKhoO9a5s ## Additional Resources - Installers and Config Files : https://www.patreon.com/posts/112099700 - FLUX Fine-Tuning / DreamBooth Zero-to-Hero Tutorial : https://youtu.be/FvpWy1x5etM - FLUX LoRA Training Zero-to-Hero Tutorial : https://youtu.be/nySGu12Y05k) - Complete Dataset, Training Config Json Files and Testing Prompts : https://www.patreon.com/posts/114972274 - Click below link to download all trained LoRA and Fine-Tuning / DreamBooth checkpoints for free - https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline/tree/main ## Environment Setup - Kohya GUI Version: `021c6f5ae3055320a56967284e759620c349aa56` - Torch: 2.5.1 - xFormers: 0.0.28.post3 ## Dataset Information - Resolution: 1024x1024 - Dataset Size: 28 images - Captions: "ohwx man" (nothing else) - Activation Token/Trigger Word: "ohwx man" ## Fine-Tuning / DreamBooth Experiment ### Configuration - Config File: `48GB_GPU_28200MB_6.4_second_it_Tier_1.json` - Training: Up to 200 epochs with consistent config - Optimal Result: Epoch 170 (subjective assessment) ### Results - [Realism Test Part 1](https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline/blob/main/Dwayne_Fine_Tune_Realism_Test_Part1.jpg) - [Realism Test Part 2](https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline/blob/main/Dwayne_Fine_Tune_Realism_Test_Part2.jpg) ## LoRA Experiment ### Configuration - Config File: `Rank_1_29500MB_8_85_Second_IT.json` - Training: Up to 200 epochs - Optimal Result: Epoch 160 (subjective assessment) ### Results - [LoRA Realism Test Part 1](https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline/blob/main/Dwayne_LoRA_Realism_Test_Part1.jpg) - [LoRA Realism Test Part 2](https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline/blob/main/Dwayne_LoRA_Realism_Test_Part2.jpg) ## Comparison Results ### LoRA Epochs Comparison - [LoRA 90 vs 160 vs Fine-Tuning 170 Comparison](https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline/blob/main/LoRA_90_Epoch_vs_LoRA_160_Epoch_vs_Fine_Tuning_170_Epoch.jpg) ### Precision Testing Compared different precision formats in LoRA training: - FP8 vs FP16 vs FP32 LoRA configurations - [View Precision Comparison Grid](https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline/blob/main/LoRA_Precision_FP8_vs_FP16_vs_FP32_Grid.jpg) ### Model Variant Analysis Tested various model variants with LoRA (FP32 Version): - FP8 FLUX DEV Base - FP8 Scaled - GGUF 8 - FLUX DEV - [View Model Variants Comparison Grid](https://huggingface.co/MonsterMMORPG/Model_Training_Experiments_As_A_Baseline/blob/main/Model_Variants_Tests_Grid.jpg) - Works best with FP16 DEV base model, then GGUF 8 base model and then FP8 raw base model and FP8 scaled model sometimes works better sometimes worse ### Key Observations - LoRA demonstrates excellent realism but shows more obvious overfitting when generating stylized images. - Fine-Tuning / DreamBooth is better than LoRA as expected. - FP8 almost yields perfect quality as FP32 with LoRA - I have used Kohya GUI to convert FP32 saved LoRAs into FP16 and FP8 - Here full public article : https://www.patreon.com/posts/115376830 ## Model Naming Convention ### Fine-Tuning Models - `Dwayne_Johnson_FLUX_Fine_Tuning-000010.safetensors` - 10 epochs - 280 steps (28 images × 10 epochs) - Batch size: 1 - Resolution: 1024x1024 - `Dwayne_Johnson_FLUX_Fine_Tuning-000020.safetensors` - 20 epochs - 560 steps (28 images × 20 epochs) - Batch size: 1 - Resolution: 1024x1024 ### LoRA Models - `Dwayne_Johnson_FLUX_LoRA-000010.safetensors` - 10 epochs - 280 steps (28 images × 10 epochs) - Batch size: 1 - Resolution: 1024x1024 - `Dwayne_Johnson_FLUX_LoRA-000020.safetensors` - 20 epochs - 560 steps (28 images × 20 epochs) - Batch size: 1 - Resolution: 1024x1024 ## Some Example Images - They have MetaData on CivitAI CivitAI : https://civitai.com/models/911087