Furkan Gözükara
patreon
DreamBooth Training on SDXL Anime Model & Comparing Ground Truth Regularization Images Effect
🕑 Added 2023-12-25 15:03:58 +0000 UTCJoin discord and tell me your discord username to get a special rank : SECourses Discord
It was requested of me to test ground truth Regularization / Classification images effect during Stable Diffusion XL (SDXL) DreamBooth training.
So for this experimentation I have used an anime SDXL model name as : Starlight XL 星光 Animated V3.
I used the configuration I shared here currently. It has been discovered after 120 different full trainings experimentation : https://www.patreon.com/posts/very-best-for-of-89213064
Hopefully I will make a full big tutorial of DreamBooth training and show all the settings.
For training Kohya SS GUI is used. If you don’t know how to install and use Kohya SS GUI here the tutorial you need > https://youtu.be/sBFGitIvD2A
And here is the tutorial that shows how to use shared configuration both on Windows and RunPod > https://www.youtube.com/watch?v=EEV8RPohsbw
So I used the below images as training images dataset.
This training dataset is at best medium quality because it has repeating backgrounds, repeating clothing and missing fully body pose.
Then for training configuration I did the following.
When training with regularization images, I did 1 epoch and 150 repeating. So more regularization images are used. And saved checkpoints at every 30 epochs. To make it that way calculate the number of steps you need to save. It is easy. 15 * 30 * 2 + 1
When training without regularization images, i set the repeating 1, trained up to 150 epochs and saved checkpoints at every 30 epoch. That is also why model names has different format. One has epoch number 1 has steps number.
I have used ultra high quality manually picked Unsplash real images having man dataset as regularization images dataset. This means they were ground truth regularization images like how Stable Diffusion was initially trained.
You can download dataset from > https://www.patreon.com/posts/massive-4k-woman-87700469
5200 images for both man and woman. Perfect quality cropped and resized to many common resolutions.
All training images and reg images were 1024x1024 pixels.
I used the 275 Fooocus amazing styles file to generate images with models so that I could pick later on.
You can find styles file (styles . csv) and styles file generator script here : https://www.patreon.com/posts/all-amazing-of-95143823
I saved both After Detailer (ADetailer) before and after images. So some images are ADetailer used and some not. This is in settings panel to set both. For after detailer I used 0.7 inpainting denoise and 80 steps manually set. Improves quality.
And here the results.
First I will compare anime
The png info as below. I used batch PNG info extractor to get them easier : https://www.patreon.com/posts/batch-image-for-95176238
parameters: picture of (ohwx man:1.1) with shirt in anime style, anime movie, anime cartoon, anime battle, muscular, anime screenshot
Negative prompt: shirtless,topless,nsfw
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 9, Seed: 1382869474, Size: 1024x1024, Model hash: f31347de70, Model: me_15_StarlightXLAnimatedV3_noREG-000030, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer mask only top k largest: 1, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.5, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use separate steps: True, ADetailer steps: 80, ADetailer version: 23.11.1, Script: X/Y/Z plot, X Type: Checkpoint name, X Values: "me_15_StarlightXLAnimatedV3_noREG-000030.safetensors [f31347de70],me15img\\model\\me_15_StarlightXLAnimatedV3-step00000901.safetensors [4610bf5599]", Version: v1.7.0-133-gde03882d
As you can see above, the regularization images affect is making it more like us. However, it reduced the stylization capability. It adds realism and more likeliness.
Now here second prompt compared. This one is realistic image instead of anime.
parameters: luxury product style portrait of ohwx man . elegant, sophisticated, high-end, luxurious, professional, highly detailed
Negative prompt: cheap, noisy, blurry, unattractive, amateurish
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 2811753849, Size: 1024x1024, Model hash: f31347de70, Model: me_15_StarlightXLAnimatedV3_noREG-000030, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer mask only top k largest: 1, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.7, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use separate steps: True, ADetailer steps: 80, ADetailer version: 23.11.1, Script: X/Y/Z plot, X Type: Checkpoint name, X Values: "me_15_StarlightXLAnimatedV3_noREG-000030.safetensors [f31347de70],me15img\\model\\me_15_StarlightXLAnimatedV3-step00000901.safetensors [4610bf5599]", Version: v1.7.0-133-gde03882d
As can be seen above again, the reg images improved likeliness and realism quality.
And here some other good images I generated
parameters: retro game art portrait of ohwx man . 16-bit, vibrant colors, pixelated, nostalgic, charming, fun
Negative prompt: realistic, photorealistic, 35mm film, deformed, glitch, low contrast, noisy
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 2542262511, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer mask only top k largest: 1, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.7, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use separate steps: True, ADetailer steps: 80, ADetailer version: 23.11.1, Version: v1.7.0-133-gde03882d
parameters: breathtaking illustration from adult comic book presenting portrait of ohwx man. fabulous artwork. best quality, high resolution
Negative prompt: deformed, ugly, low quality, low resolution
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 427502473, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, Version: v1.7.0-133-gde03882d
parameters: closeshot picture of (ohwx man:1.1) wearing shirt, 3D render, 3D art, 3d character, muscular game, cgi, artstation
Negative prompt: ugly, deformed, noisy, blurry, low contrast, grunge, sloppy, unkempt, photograph, photo, realistic, nsfw, shirtless, topless
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1171263453, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, Version: v1.7.0-133-gde03882d
parameters: picture of (ohwx man:1.1) in Photorealistic chibi art style, facing the viewers , artistic, artist, simple volumetric light background, swirling clouds, ColorART, many cute ghosts with tongues out, chibi,
Negative prompt: gaussian noise worst quality lowres oversaturated undersaturated overexposred underexposed grayscale bw bad photo bad photography bad art blur blurry grainy morbid ugly asymmetrical mutated malformed mutilated poorly lit bad shadow draft cropped out of frame cut off censored jpeg artifacts out of focus glitch duplicate pixelated soft focus color fringing overprocessed oversharpened
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 3576768970, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, Version: v1.7.0-133-gde03882d
parameters: portrait of (ohwx man:1.1), painting by sargent and rhads and leyendecker and greg hildebrandt evening sky
Negative prompt: gaussian noise worst quality lowres oversaturated undersaturated overexposed underexposed grayscale bw bad photo bad photography bad art blur blurry grainy morbid ugly asymmetrical mutated malformed mutilated poorly lit bad shadow draft cropped out of frame cut off censored jpeg artifacts out of focus glitch duplicate pixelated soft focus color fringing overprocessed oversharpened
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 1252309971, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, Version: v1.7.0-133-gde03882d
parameters: Watercolor style painting, portrait of ohwx man , visible paper texture, colorwash, watercolor
Negative prompt: ugly, deformed, noisy, blurry, low contrast, photo, realistic
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 3204333295, Size: 1024x1024, Model hash: 29bbb5ce43, Model: me_15_StarlightXLAnimatedV3-step00001802, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer mask only top k largest: 1, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.7, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use separate steps: True, ADetailer steps: 80, ADetailer version: 23.11.1, Version: v1.7.0-133-gde03882d
parameters: Flat 2D Art, portrait of ohwx man , simple flat color, 2-dimensional, Flat 2D Art Style
Negative prompt: ugly, deformed, noisy, blurry, low contrast, 3D, photo, realistic
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 998589366, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, Version: v1.7.0-133-gde03882d
parameters: pixel-art portrait of ohwx man . low-res, blocky, pixel art style, 8-bit graphics
Negative prompt: sloppy, messy, blurry, noisy, highly detailed, ultra textured, photo, realistic
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 3068779067, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, Version: v1.7.0-133-gde03882d
parameters: comic portrait of ohwx man . graphic illustration, comic art, graphic novel art, vibrant, highly detailed
Negative prompt: photograph, deformed, glitch, noisy, realistic, stock photo
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 2542262511, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer mask only top k largest: 1, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.7, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use separate steps: True, ADetailer steps: 80, ADetailer version: 23.11.1, Version: v1.7.0-133-gde03882d
parameters: breathtaking illustration from adult comic book presenting portrait of ohwx man. fabulous artwork. best quality, high resolution
Negative prompt: deformed, ugly, low quality, low resolution
Steps: 20, Sampler: DPM++ 2M SDE Karras, CFG scale: 7, Seed: 427502468, Size: 1024x1024, Model hash: f8f06f0d73, Model: me_15_StarlightXLAnimatedV3-step00002703, Version: v1.7.0-133-gde03882d
Comments
Furkan Gözükara
amazing. by the way 9 epoch 40 repeat = 2.4 epoch 150 repeat roughly
Ec Jep
thx - yesterday I compared 1 epoch at 150 repeats and 9 epochs at 40 repeats and so far the 9 is better than 8 (which I have always stopped at). I still can't believe how good training can be. 1 epoch gives a good likeness but 8-9 epochs is quite a bit better on my 4090 24gb.
Furkan Gözükara
i recommend 1 epoch and 150 or 200 repeat and save checkpoints and find best checkpoint for realism.
Ec Jep
what is your conclusion with this testing? For photo realism do you still recommend 8 epochs and 40 repeats or 1 epoch and 150 repeats?
Furkan Gözükara
Well these are downloaded from civitAI. i only trained myself extra :)
Hash Bookies
Looks great! Quick question, I am new to this, how can I get all your awesome models to try one by one ? 😅
