SAFIRE: Stylized Anime Faces In Real Environments X: v1+v2+v3 etc. These loras are intended for god-tier "anime-IRL" stylization and ultra-fidelity female characters. Suggested usage: - tested only with flux1-dev/fp8, 1024x1024 - use a low guidance of 1.8~2.2 for higher realism - chain together multiple loras with low weights, gradient increase based on version. This will be more flexible than using a single version because of the way the iterative training process works. Each version is trained by selecting the top ~10% of previous version's outputs along with a few gems from previous versions. These are the faces of mythological model collapse, so with higher weights or a single lora you will notice some prevalent biases. - There are no specific trigger words, but 'anime' is prevalent in training captions. Strengths: - "My waifus are too strong for you traveler. You can't handle my waifus they are too strong for you. You do not understand, traveler, you cannot handle my strongest waifus. You need to find a lora that generates weaker waifus, traveler. You cannot handle my strongest waifus, no one can." - These are a good visualization for robots to contrast against the 999 companies creating clunky faceless tincan droids, preluding the inevitable image-to-robot 3D printers powered by megamind catgirls. - 'anime IRL' is a difficult target, made 10x easier with these loras. - with proper usage, a prompt as simple as "anime sorceress" can yield extreme fidelity. Weakness: - how do hands work? - can overbake into pseudo-3D aesthetics - can sometimes invoke creepy lifeless dolls. - carries forward several quirks of various anime art styles into reality causing occasional artifacts. - biased towards certain clothing, jewelry, collars, and random hair ornaments. - biased towards certain framing, blurred backgrounds, architecture, etc. - general flux problems that are not entirely mitigated Side Effect: - tends to beautify everything a little (ex: a normal picture of the woods can become a magical paradise), which could be good or bad depending on what you are trying to make. --- feel free to use these however you want, train stuff with the outputs, summon an army of vtuber robots to run around in the street, etc. I consider these loras a proof of concept: simply training in a loop where directionally-optimal outputs are subjectively selected improves towards a target, with no outside references required. These were made with relatively weak/narrow prompts, low diversity, mediocre captions, and short training runs. To say that there is room for even more improvement beyond these would be an understatement. You cannot handle my strongest waifus, traveler!

SAFIRE-X

initial variety version trained on a larger pool of generated samples

Files with SHA256: f8d9a8ca2746badb3ba71ffb3aa0432ddd23bfd21d70d7d401ad646bb01caae6