Patreon
Home / Videos / Train Stable Diffusion on 15 Photos of Your Face: Full SDXL

Train Stable Diffusion on 15 Photos of Your Face: Full SDXL Fine-Tuning With OneTrainer (2024)

Timestep sampling shift explained for diffusion models: Qwen Image uses an exponential shift and FLUX a linear one, and 2.205 matches Qwen at 1664x928.

Published Apr 09, 2024 · 50,041 views · Watch on YouTube

Get the app and files

Decoding the Timestep Shift in Diffusion Training: Qwen Image & FLUX

Decoding the Timestep Shift in Diffusion Training: Qwen Image & FLUX

Timestep sampling shift explained for diffusion models: Qwen Image uses an exponential shift and FLUX a linear one, and 2.205 matches Qwen at 1664x928.

InstantID Gradio App: 1-Click Face Transfer With Xinsir ControlNet

InstantID Gradio App: 1-Click Face Transfer With Xinsir ControlNet

InstantID Gradio app transfers a face from one photo into new SDXL images with Xinsir ControlNet; block swap lowers 1280x1280 runs to 8.25 GB VRAM.

NVIDIA SANA 1-Click Installers & Gradio App: 4K Images on 8 GB VRAM

NVIDIA SANA 1-Click Installers & Gradio App: 4K Images on 8 GB VRAM

NVIDIA Labs SANA text-to-image model in an improved SECourses Gradio app; with all 4 VRAM optimizations, 1K images need 4 GB, 2K 6 GB and 4K 8 GB GPUs.

V-Express Talking Avatar App: 1-Click Image to Animation Installers

V-Express Talking Avatar App: 1-Click Image to Animation Installers

Tencent AI Lab V-Express animates a still portrait into a talking video from audio; after VRAM optimizations 8 GB GPUs can make 20-second videos or longer.

EasyAnimate 1-Click Installer: Image-to-Video and Text-to-Video App

EasyAnimate 1-Click Installer: Image-to-Video and Text-to-Video App

EasyAnimate 5.1 app for text-to-video, image-to-video with start and end frames and video-to-video, at 512, 768 and 960 resolutions; 512 needs a 24 GB GPU.

OmniGen Unified Image Generation: 1-Click Windows & Cloud Installers

OmniGen Unified Image Generation: 1-Click Windows & Cloud Installers

OmniGen is a unified image generation model that makes a wide range of images from multi-modal prompts; the Gradio app runs in 5.4 GB VRAM with model offload.

Chapters

  1. 0:00 Introduction
  2. 3:54 Intro to instructions GitHub readme
  3. 4:32 How to register Massed Compute (MC) and start virtual machine (VM)
  4. 5:48 Which template to choose on MC
  5. 6:36 How to apply MC coupon
  6. 8:41 How to install OT on your computer to train
  7. 9:15 How to verify your Python, Git, FFmpeg and Git installation
  8. 12:00 How to install ThinLinc and start using your MC VM
  9. 12:26 How to setup folder synchronization and file sharing computer and MC VM
  10. 13:56 End existing session in ThinClient
  11. 14:06 How to turn off MC VM
  12. 14:24 How to connect and start using VM
  13. 14:41 When use end existing session
  14. 16:38 How to download very best OT preset training configuration for SD 1.5 & SDXL models
  15. 18:00 How to load configuration preset
  16. 18:38 Full explanation of OT configuration and best hyper params for SDXL
  17. 24:10 How to setup training concepts accurately in OT
  18. 24:52 How to caption images for SD training
  19. 30:17 Why my training images dataset is not great and what is a better dataset
  20. 31:41 How to make DreamBooth effect in OT with reg images concept
  21. 32:44 Effect of using ground truth regularization images dataset
  22. 34:41 How to set regularization images repeating
  23. 35:55 Explanation of training tab configuration and parameters
  24. 41:58 What does masked training do and how to do masked training and generate masks
  25. 44:53 Generate samples during training setup
  26. 46:05 How to save checkpoints during training to compare and find best one later
  27. 47:11 How to save your configuration in OT
  28. 47:22 How to install and utilize nvitop to see VRAM usage
  29. 48:06 Why super slow training happens due to shared VRAM and how to fix it
  30. 48:40 How to reduce VRAM usage before starting training
  31. 49:01 Start training on Windows
  32. 49:11 Starting to setup everything on MC same as on Windows
  33. 49:37 Upload data to MC
  34. 51:11 Update OT on MC
  35. 52:33 How to download regularization images
  36. 53:42 How to minimize all windows on MC
  37. 54:00 Start OT on MC
  38. 54:20 Setting everything on MC same as Windows
  39. 55:22 How to set folders on MC VM
  40. 56:31 How to properly crop and resize your training images
  41. 57:47 Accurate Auto1111 Models folder on MC
  42. 58:05 Copy file & folder path on MC
  43. 58:54 All of the rest of the config on MC
  44. 63:29 How to utilize second GPU if you have
  45. 65:45 Checking back again our Windows training
  46. 66:06 How to use Automatic1111 (A1111) SD Web UI on MC and Windows
  47. 71:35 How to use default Python on MC
  48. 71:55 Checking training speed and explaining what it means
  49. 72:13 How many steps we are going to train explanation
  50. 73:40 First checkpoint and howe checkpoints named
  51. 74:15 How to fix A1111 errors
  52. 75:44 How to start A1111 Web UI and use it with Gradio Live share and locally
  53. 77:45 What to do if model loading takes forever on Gradio and how to fix it
  54. 79:01 Where to see status of the training of OT
  55. 79:43 How to upload checkpoints / anything into Hugging Face for permanently saving
  56. 86:21 How to auto upgrade A1111 and install ADetailer and ControlNet extensions
  57. 89:10 How to use trained model checkpoints on Massed Compute
  58. 90:08 How to test checkpoints to find best one
  59. 92:15 Why you should use After Detailer (adetailer) and how to use it properly
  60. 94:48 How to do proper highres fix upscale
  61. 96:19 Why anatomy inaccuracy happens
  62. 97:07 How to generate images forever in A1111
  63. 98:02 Where the generated images are saved and download them
  64. 100:30 Super Important
  65. 105:16 Analyzing x/y/z checkpoint comparison results to find best checkpoint
  66. 108:20 How to understand if model is overtrained
  67. 112:27 How to generate different expressions having photos
  68. 114:53 How to do inpainting in Stable Diffusion A1111
  69. 116:34 How to generate LoRA from your trained checkpoint
  70. 118:03 Windows OneTrainer training completed so how to use them on your computer
  71. 120:24 Best SD 1.5 models Fine-Tuning / DreamBooth training configuration / hyper-parameters
  72. 123:50 How can you know you have sufficient VRAM?
  73. 125:36 What to do before terminating MC VM
  74. 126:55 How to terminate your VM to not spend anymore money
  75. 128:35 How to do style, object, etc training
  76. 129:47 What to do if your thin client don't synch

About this video

In this tutorial, I am going to show you how to install OneTrainer from scratch on your computer and do a Stable Diffusion SDXL (Full Fine-Tuning 10.3 GB VRAM) and SD 1.5 (Full Fine-Tuning 7GB VRAM) based models training on your computer and also do the same training on a very cheap cloud machine from MassedCompute if you don't have such computer.

SDXL Configs Updated to Better Version After 20 New Experiment ⤵️

https://www.patreon.com/posts/96028218

Tutorial Readme File ⤵️

https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/Tutorials/OneTrainer-Master-SD-1_5-SDXL-Windows-Cloud-Tutorial.md

Register Massed Compute From Below Link (could be necessary to use our Special Coupon for A6000 GPU for 31 cents per hour) ⤵️

* https://bit.ly/SECoursesMassedCompute

* https://vm.massedcompute.com/signup?linkId=lp_034338&sourceId=secourses&tenantId=massed-compute

Coupon Code for A6000 GPU is : SECourses

0:00 Introduction

3:54 Intro to instructions GitHub readme

4:32 How to register Massed Compute (MC) and start virtual machine (VM)

5:48 Which template to choose on MC

6:36 How to apply MC coupon

8:41 How to install OT on your computer to train

9:15 How to verify your Python, Git, FFmpeg and Git installation

12:00 How to install ThinLinc and start using your MC VM

12:26 How to setup folder synchronization and file sharing computer and MC VM

13:56 End existing session in ThinClient

14:06 How to turn off MC VM

14:24 How to connect and start using VM

14:41 When use end existing session

16:38 How to download very best OT preset training configuration for SD 1.5 & SDXL models

18:00 How to load configuration preset

18:38 Full explanation of OT configuration and best hyper params for SDXL

24:10 How to setup training concepts accurately in OT

24:52 How to caption images for SD training

30:17 Why my training images dataset is not great and what is a better dataset

31:41 How to make DreamBooth effect in OT with reg images concept

32:44 Effect of using ground truth regularization images dataset

34:41 How to set regularization images repeating

35:55 Explanation of training tab configuration and parameters

41:58 What does masked training do and how to do masked training and generate masks

44:53 Generate samples during training setup

46:05 How to save checkpoints during training to compare and find best one later

47:11 How to save your configuration in OT

47:22 How to install and utilize nvitop to see VRAM usage

48:06 Why super slow training happens due to shared VRAM and how to fix it

48:40 How to reduce VRAM usage before starting training

49:01 Start training on Windows