<h1 id="description">Description</h1>Long time no see.Because of work, I have not had much time recently to deeply explore and study Stable Diffusion models. But now I am back with a new concept.This page introduces the BASE version of my model series.If this is your first time seeing this model, you can check the preview images and decide whether it matches the workflow you are looking for.<h1 id="about-the-model">About the Model</h1>First of all, what does BASE mean?Does it mean this model is mainly designed as a finished creative model for direct image generation?Not exactly.Is it suitable for being used together with your LoRAs?Yes. Very suitable.That is the main purpose of this model.The positioning of this BASE model is to work as a stable and compatible foundation model, especially for users who want to combine it with LoRAs and reproduce the style, character, or concept of their own datasets more accurately.Instead of pushing a very strong fixed style, this model is designed to give LoRAs more room to perform. It aims to help your LoRA show its intended effect more clearly, while still keeping good image quality, color, lighting, and overall visual stability.<h1 id="latest-update">Latest Update</h1><h2 id="base-version">BASE Version</h2>This model has a strong relationship with Illustrious-v2 / ILL V2.Because of that, LoRAs trained on ILL V2-based models should usually be able to achieve good results on this BASE model. In particular, character LoRAs, style LoRAs, and dataset-specific LoRAs based on the ILL V2 ecosystem may benefit from better compatibility and more faithful reproduction.I have also adjusted and repaired the key parameters of this model, so it can be used directly as a base model for LoRA training.If you want to train your own LoRA and need a clean, stable, and compatible foundation, this model may be a suitable choice. It can help your dataset features appear more naturally and give you more control over the final output.<h1 id="model-positioning">Model Positioning</h1>This is not a model that tries to do everything by itself.It is not meant to completely replace your LoRA, your prompt structure, or your creative direction.Instead, it is designed to support them.The main goal of this BASE model is:<ul><li>better LoRA compatibility</li><li>better reproduction of dataset features</li><li>stable performance with ILL V2-related LoRAs</li><li>suitable use as a LoRA training base model</li><li>balanced image quality without overly suppressing LoRA effects</li><li>more flexible results for both anime-style and semi-realistic workflows</li></ul><h1 id="about-compatibility-with-loras">About Compatibility with LoRAs</h1>This model is especially recommended for LoRAs trained on ILL V2-based models.In most cases, it should be able to reproduce the LoRA’s intended style or character features more accurately than models with a very strong built-in style.However, the final result still depends on the LoRA itself, the training quality, the dataset, the prompt, and the sampling settings.If you are stacking multiple LoRAs together, I still recommend keeping the total weight under control. Do not force every LoRA to run at full strength. Lower and more balanced weights usually produce more stable and natural results.<h1 id="about-lora-training">About LoRA Training</h1>This BASE model can be used directly as a LoRA training base model.The relevant parameters have already been adjusted and repaired, so you do not need to treat it only as a merged model for image generation.If your goal is to train a LoRA and make the final result closer to your dataset, this model may help you achieve a more faithful and controllable output.Of course, training results still depend heavily on your dataset quality, caption structure, training parameters, and workflow.<h1 id="recommended-settings">Recommended Settings</h1>There is no single perfect setting for every LoRA or prompt.For general use, I recommend starting with moderate CFG values and adjusting based on the type of image you want.If the image looks too weak or lacks prompt control, you can slightly increase CFG.If the image looks too rigid, overcooked, or the LoRA effect becomes too strong, you can lower CFG or reduce LoRA weight.For LoRA testing, I recommend changing only one variable at a time, such as CFG, LoRA weight, sampler, or steps. This makes it easier to understand how the model and LoRA are interacting.<h1 id="notes-on-use">Notes on Use</h1>This model is more like a foundation than a final answer.It is suitable for creators who want to build on top of it, train LoRAs, test LoRA compatibility, or explore different dataset-driven effects.If you only want a model with a very strong fixed visual style, this may not be the best choice.But if you want a flexible base that allows your LoRA to express itself more clearly, this model is exactly made for that purpose.<h1 id="personal-note">Personal Note</h1>It has been a while since my last update.During this time, I was busy with work and did not have enough energy to continue deep model research. But I still care about AI image generation, model sharing, and the creative possibilities behind them.This BASE model is a new attempt for me.I hope it can be useful not only for image generation, but also for people who want to train their own LoRAs and present their own ideas more accurately.If you create something interesting with this model, feel free to share it.Your feedback, tests, and ideas are always valuable to me.

q7_AIO_base

Fixed the issue with the base model's unet regarding dark environment generation.

Files with SHA256: ca4c041fdeca3add5c3cddd654b19e8d4009ea73c480b680059a6d8065701c92