Chimera. Merged my CatEar and SDXL, then merged that with hakurei/waifu-diffusion-xl · Hugging Face, then trained for forever for 1.0. Then merged in Blue-Pencil @ .25, then trained a lot more. Every so often I'll fold base SDXL or wdxl back in and start training again. Dataset is a curated and growing selection of images from various artists and 30 or so pics of top 100 rated pictures for top 100 characters from Danbooru. I just recently had to restart from v.2 since I tried to up the learning rate, and ended up with completely fried images (identical tags would equal near identical images. :/)

@Korewaai Thanks! - I'll have to check out the merge so I can heart them~

@D00derino584  Not that many - I think first run was ~16000? Now its up to 32694. Thought process is finetuning should be done with well captioned, curated, and high quality images, and not just throwing tons and tons of images from danbooru at it (Thankfully, WD and SDXL did that part). I do ~30 - 50 images of the best images from each artist (with more from ones I like/I find interesting) well captioned with Art by Artist (SD upscaled to at least 2160*2160 -- one trainer downscales them, of course). Then a bunch of the best quality anime, retro anime or cartoon images with generic captions. Then run and hope the RNG gods smile upon me. I train the text encoder2 at 1/4 the learning rate.

@Korewaai Mainly to be dummy-proof -- didn't want to do the math for what size I had to do for what aspect ratio, and I wanted some leeway in case I had to crop face, etc. I just made a quick script, below,

# Add the necessary assembly for System.Drawing Add-Type -AssemblyName System.Drawing

# Define the directory path, new folder for small images, and report file path

$directoryPath = "C:\path\to\your\images\directory"

$newFolderPath = Join-Path $directoryPath "SmallImages"

$reportFilePath = "C:\path\to\your\sortedImagesReport.txt"

# Create the new folder if it doesn't exist if (-not (Test-Path -Path $newFolderPath)) { New-Item -Path $newFolderPath -ItemType Directory } # Get all image files in the directory $imageFiles = Get-ChildItem -Path $directoryPath -File | Where-Object { $_.Extension -match 'jpg|jpeg|png|gif' }

# Process each image file foreach ($file in $imageFiles) { $image = [System.Drawing.Image]::FromFile($file.FullName) $pixelCount = $image.Width $image.Height $image.Dispose() if ($pixelCount -lt 4665600) { # Move images with less than 4,665,600 pixels to the new folder Move-Item -Path $file.FullName -Destination $newFolderPath } }

# Generate a report of remaining images sorted by pixel count $imageDetails = Get-ChildItem -Path $directoryPath -File | Where-Object { $_.Extension -match 'jpg|jpeg|png|gif' } | ForEach-Object { $image = [System.Drawing.Image]::FromFile($_.FullName) $pixelCount = $image.Width $image.Height $image.Dispose() [PSCustomObject]@{ Name = $_.Name PixelCount = $pixelCount } } $sortedImages = $imageDetails | Sort-Object -Property PixelCount $sortedImages | Out-File -FilePath $reportFilePath # Display the report file content (optional) Get-Content -Path $reportFilePath

to filter all images less than 2160*2160 into a separate folder so I could SD upscale them at .01 if they were too small. Didn't want to run into the issue where it tried to bucket a like 256 * 1024 image or something crazy.

@fikxzer Originally, I used WD vit with a .75 confidence level. The idea was this gave the non-artist tagged images a loose concept and gave SD a broad sense of understanding of those concepts. In my current tests, I duplicate the image, then have one version of the image tagged with WD-vit-3 at .7, and the duplicate tagged at .55. Then I caption the image tagged with .55 using COGVLM with the following prompt, "Describe the image using the following tags as guidance: TAGS TAGS TAGS. Ensure the background is well described." Followed by human correction. I use no quality tags -- I never understood using quality tags as I would assume that would cause concept bleed between prompts, and what is the difference between a "good" image and a "masterpiece".

I used python and a vision LLM to do this at first -- but tag gui meets all of those needs now, and is less hassle. GitHub - jhc13/taggui: Tag manager and captioner for image datasets