Anitrait FX

"Please read this before using, thank you."

Generate simple anime style portraits for Flux Dev.

(Currently Works for Simple Prompts Only!)

This is v1.1 release. This version is intended to gain more flexibility in terms of different prompt usage compared to v1. I do hope this works better when being used on Shakker's Zeno model since I can't do a test on it. I do have some changes in the dataset, but it is still based on Animagine 3.1. This version is more or less the same as v1, so you can still use v1 if this doesn't suit you. v1.1 might be the last of version 1 since I'm thinking of using Animagine 4 as dataset for the next version.

Use settings:

> Put trigger words at the beginning or the end of prompt (preferably at the beginning).

> Weight from 0.6 to 1. Using 0.6 is okay if you put "anime" in prompt but otherwise start from 0.7.

> Use Euler. Basically any Euler works, but somehow for this version, I think Euler normal is better. I never tested others, but if you're comfortable with certain samplers, you can definitely try with that.

> 20 steps is fine. 25 is also fine. 30 is fine. I never tested more than 30 since it's expensive.

> You don't really need upscaler but you can use like R-ESRGAN 4x + Anime 6B if you want higher fidelity.

> Clip skip... I don't really know. I usually use 1, but 2 also works.

> Recommended for portrait images. I'm not talking about image orientation, but like, that portraits. I hope you understand this. But if you want to create something else then feel free to do it since I'm not really a good prompter and can't test it with good prompt.

Changelog (v1.1):

> Even smaller size! I managed to cut it to around 150 MB but still performs quite well.

> Slightly better dark images (I added some dark images in the data).

> Trigger word changed into "ry012_style". No reason for this other than me trying things out XD .

> Slightly improved flexibility compared to v1.

> Changes to the training parameter resulting in slight difference in behavior (compared to v1), but other than that, it's the same.

Known issue:

> Bare feet still somewhat sucks.

> I don't know if it's an issue, but it might produce inconsistent results in different tools (A1111, regular AI Generator, Comfy). Flux likes it at 0.6; A1111 likes it at 0.7-0.8 even though 0.6 is alright; so you need to play with the weight more. But the result between different weights is quite consistent compared to previous versions.

> You tell me

Thank you for using this model for your image generation.

====================================================