Tired of your LoRAs looking like a goblin coughed them up? Do your datasets resemble a dragon's hoard – impressive to the untrained eye, but full of worthless trinkets? Fear not, intrepid creator! Dragon Diffusion UK Ltd. has forged a tool of arcane power: the Dragon Diffusion Dataset Validator!
https://github.com/DragonDiffusionbyBoyo/Dragon-Diffusion-Dataset-Validator
The Human Eye: A Deceiving Sorcerer
We humans, with our squishy brains and limited processing power, are easily fooled. We see a picture of a majestic dragon breathing fire and think, "Yes, this is magnificent! Throw it in the training set!" But what does the machine see? Probably a blurry mess of pixels, noise artifacts, and a caption that's about as relevant as a lute solo at a death metal concert.
Our eyes, bless their fleshy little hearts, fill in the gaps. They imagine detail where there is none. They are, in short, unreliable narrators.
The AI Eye: A Cold, Calculating Dragon
The Dragon Diffusion Dataset Validator, however, is a different beast entirely. It's an "AI eye," a merciless judge that sees the world in pure, unadulterated math. It doesn't care if a picture looks cool. It cares about cosine similarity, noise levels, blur metrics, and a whole host of other arcane measurements that would make a wizard's head spin.
This tool surgically dissects your dataset, analyzing each image with the precision of a dragon hoarding gold. It compares images to their captions using a vast semantic VLM, ensuring that "dragon" actually looks like a dragon, and not a slightly singed potato. If you're after a specific face, it'll make sure it's that face, and not some random peasant.
The GUI: A Color-Coded Prophecy
The Dragon Diffusion Dataset Validator's GUI is a mystical oracle, speaking in the language of RED, AMBER, and GREEN.
GREEN: All is well! This image is worthy of the dragon's hoard.
AMBER: Proceed with caution. This image might have a few flaws (maybe it's a dragon's backside, missing crucial facial features), or the caption might be a bit off (perhaps it contains a trigger word that mess with the semantic test).
RED: Abandon all hope, ye who enter here! This image is pure, unadulterated garbage. It will corrupt your LoRA like a cursed artifact.
The Verdict: Science, Not the Force
The Dragon Diffusion Dataset Validator isn't just a fancy tool; it's a paradigm shift. It's about replacing the subjective with the objective, the guesswork with science. It's about trusting the math, even when your eyes are screaming, "But it looks so cool!"
The tool also provides a detailed CSV report, a scroll of forbidden knowledge, explaining the exact test scores for each image. You can then use the "move to hold" function to isolate all the RED-flagged images.
What happens next is up to you, brave creator. Do you attempt to resurrect the fallen image with upscaling and repair? Or do you accept the machine's cold, hard verdict and cast it into the digital abyss? The choice is yours.
But remember: we are wielding science, not the Force. Trust the Dragon's Eye, and your LoRAs will soar to new heights of glory.
This was tested on 24 Gig of Vram. Please let me know if you have OOM or performance issues with smaller GPUs, if I do get constructive accurate feedback I will look into performance enhancement methods. Currently it is gobbling up all images no matter the size but a low Vram version is possible if this does not run. I am not going to make a low Vram version though as I do not know if it actually does or does not work on Low Vram as of yet.
Other constructive feedback and requests will be considered on Merit. There is a technical gate of building a Venv and installing at the moment so that if there are issues I have not encountered yet I can resolve before I make it one click installer.
Enjoy