ℹ Notice - Do not reprint this model. This model can be merged and shared, however derivatives cannot be used for commercial purposes (generative services, commissioned models, etc.)

💡 Alpha 1 - A starting point for the Amber Light model series focusing on creative anime styling and capability.

Info

This is a model series I had worked on at one point last month as a means of getting back into some SDXL testing. A lot of testing, many variations, as well as some custom test LoRA implementation in sync with the timely release of Civitai's LoRA training features.

I've decided to use this version in the meantime as a starting point as I re-familiarize myself with where I left off and each version again. This still may take some time on and off though. I do have a couple versions in mind to keep playing around with as improved variants currently. I may go back a bit into the recipe and cut some things out or tone down weights as well.

Feel free to give some detailed feedback for this current iteration through comments and reviews, as they are generally pretty helpful! 👍

Recommended Use

Use this model as you see best. You can try a wide range of prompt styles, samplers/steps/scale, and other parameters to fully see what capabilities this model currently has. The below information is just a general starting point which I happened to primarily use while testing variations.

Samplers and cfg scale variation can change up the styling a bit. I've primarily used Euler a with ~7 scale, but feel free to try out others. You may find more illustrative type styles using lower cfg scale and/or some of the karras samplers.

As always, you can reference the preview/gallery images, or use the values below as a solid starting point.

~20-25 steps, ~6.5-8 scale, hires fix denoising, Adetailer, clip skip 1, Euler a or DPM++ 2M Karras samplers

Hires Fix - Yes

I suggest using hires fix or alternatives if you are able to. This gives some more clarity to detail and shapes things up with a 2nd pass etc., even at a higher starting resolution. The values below are what I am able to run in a1111 using an 8gb vram gpu w/ ~768x1152 resolutions and the compatibility settings listed below.

upscaled by 1.3, denoising strength 0.22, hires steps 20, 4x-UltraSharp upscaler

Negative Embeddings - ~Yes

Negative TI do alter styling a bit, this may or may not be beneficial to you, all factors depending. I have primarily used UnaestheticXL_Sky3.1. Feel free to try others or none at all.

Adetailer - Optional

Adetailer can be a good tool for faces on characters at a mid/long distance, emphasizing detail/expressions, or other creative uses. There are also options for hands and other options you can add in.

I suggest using this tool sparingly however, try with and without if you would like. I've noticed solid faces can get messed up a bit due to the positives/negatives I had input with adetailer at times, or they just generally got altered too much from the base prompt.

You may notice this in general with some of the similar facial expressions in the preview images. Some of this may also just be part of the model.

With that said, this tool can also be widely beneficial at times if used properly and strategically alongside your prompt. One example being to add more of an anime style (or any style) to faces if your prompt wildly opposes such styling; and any current negative TI and/or model in general aren't providing enough base styling.

I have not noticed this as an issue with this model series thus far, but something to keep in mind. I have had limited testing with both negative TI and adetailer turned off however.

Below is a configuration example:

model face_yolov8n
inpaint denoising strength ~0.25-0.36
positive anime, cute
negative worst quality, blurry, you can also try using negative TI here.

Prompt Terms - Optional

Natural language + tag oriented prompts should work fine with this model, let me know if you have any additional thoughts. Keep in mind SDXL responds to various things in a lot of different ways. I generally just used any sort of prompt while testing though, as seen in the preview images.

Make note of the styling direction you are giving the model with your prompt, negative TI and other parameters, etc. and any potential conflicts that don't align with the styling your going for. This model/version has a pretty anime focused neutral styling so you shouldn't need anime or illustration style tokens to achieve this. However they can always help or emphasize the styling more so if you choose to use them.

positive - none or terms such as anime, anime opening, 1girl, solo, screencap
negative - (bad quality), (worst quality), watermark, (blurry), (unaestheticXL_Sky3.1:0.8)

Refiners - Try

I have not tried using any refiner models with this model/version, feel free to try any!

Workflows - Try

I also haven't used this model/version using ComfyUI or within any SDXL specialized workflows. I imagine those who do and have some solid ones might be able to make wonderful use of this model with them.

Compatibility Settings - Optional

If you're having trouble running SDXL models, you can try adding these startup flags while running in a1111 (via editing the webiui-user batch file).

set COMMANDLINE_ARGS= --xformers --medvram-sdxl

This is what I use personally and have little to no issues as long as I don't overdo resolution scaling with the allotted vram for my gpu. There are some other flags available you can use for further optimization if you do a bit of searching. You can also try using ComfyUI for running SDXL models which I had no issues with in the past with other models and testing.

I use the compatibility setting use old karras scheduler signmas (0.1 to 10)within a1111 for the preview images.

Note: this version/model is not finetuned to maximum detail accuracy.

Disclaimer

This model(s) may output NSFW content unintentionally depending on parameters used. Make sure you tailor your prompts accordingly. For example "nsfw" in the negative prompt.

The purpose of sharing this model is not to showcase obscene material in a public forum. The use of this learning model is entirely at the discretion of the user, and they have the freedom to choose whether or not to create SFW or NSFW content. The decision of whether to engage with SFW or NSFW content lies with the user and their own personal preferences. The AI model(s) do not contain explicit visual content that can be accessed easily.

Note: this is a general precautionary disclaimer; as SDXL models change and other merges are added, this may or may not be more relevant.

Resources

This model used the following resources - all of these are great themselves, props to each creator.

Ambience_a1
Custom LoRA (AnimeStyles_SDXL)
Custom LoRA (StyleBuff_SDXL)
BreakDomainXL_v05g (via br_d)
SDXLAnime_v3.1 (via Nullcaster)
Kohaku-XL_beta5 (via kblueleaf)

📝 Detailed recipes may go up on a Hugging Face page in the future.

License & Use

CreativeML Open RAIL++-M

Use Restrictions

You agree not to use the Model or Derivatives of the Model:
- In any way that violates any applicable national, federal, state, local or international law or regulation
- For the purpose of exploiting, harming or attempting to exploit or harm minors in any way
- To generate or disseminate verifiably false information and/or content with the purpose of harming others
- To generate or disseminate personal identifiable information that can be used to harm an individual
- To defame, disparage or otherwise harass others
- For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation
- For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics
- To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm
- For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories
- To provide medical advice and medical results interpretation
- To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use).

Terms of use

- You are solely responsible for any legal liability resulting from unethical use of this model(s)
- If you use any of these models for merging, please state what steps you took to do so and clearly indicate where modifications have been made.

Note

If you see any conflicts or corrections to be made, please let me know.