Hyphoria Real [Illu] - v0.9

The focus of Hyphoria Real was first and foremost prompt adherence, just like Hyphoria. That is, what you put in the prompt box, the model will do its damndest to make it happen. I built this on top of Hyphoria, merging in various Illustrious based realistic models.

v0.9

I'm excited to share the accomplishments of the new Hyphoria Real model, which achieves an exceptional balance of prompt adherence and flexibility while enhancing realism.

The default output delivers impressive image quality - detailed and realistic visuals that seamlessly integrate with the Hyphoria family aesthetic. The model maintains high flexibility while producing remarkably lifelike results.

One of the most significant achievements is reducing the number of attempts needed to get your desired image. Hyphoria Real demonstrates excellent prompt adherence and handles traditionally challenging areas like fingers and faces with remarkable competence. When these elements are prominently featured, the model performs exceptionally well. Many base renders are ready to use without modifications, though upscaling and detailers can further enhance the final images.

While I continue to refine lighting and camera effects flexibility (with specific training data being collected to address this), the current version represents a sweet spot of quality, realism, and creative control.

If you discover scenarios where Hyphoria Real could be improved - especially areas where your favorite realistic models excel - please let me know so I can incorporate specific training data to enhance those capabilities.

There are three weaknesses with this version, which I will be addressing in v1.0, first is sampler performance, some samplers can cause unwanted noise like DPM++ 2M non-SDE (as on civit) in some cases, secondly is the teeth leave a lot to be desired and detailing or hiresfix is a must when they are not prominently displayed. The other is the somewhat analogue natural blurry look, even when you hires with it. It enhances the realism but at the same time can feel weird when you expect it to not be blurry after hires.

v0.5

This was my first forray into realistic models. I learned a ton making it. I learned even more seeing all the interesting things people make with it. It was always a stepping stone towards better versions but it was a great step towards my ultimate vision.

Recommended Settings

• Steps: 20-35 (I have always liked 35, but lower is just as good)

• CFG: 2.5-6

  • I use 3, it is my sweet spot, but CFG can be used to affect the look more so than prompt adherence.

• Sampler: DPM++ 2M SDE (or Euler A base with potentially DPM++ 2M for hiresfix on Civit)

  • If you want a more realistic image, please use DPM++ 2M SDE and not just Euler A

• Scheduler: Karras

• Resolution: I recommend 1024x1024 - 1280x1280 for normal generation. Supports up to 1536x1536 without breaking, but can suffer some body extension. Any aspect ratio that equals the same number of pixels should work.

• Prompting: This model listens to your prompts, so avoid the fluff and keep your prompts focused.

• Recommended Positives:

  • photorealistic, realistic - For a more realistic photo look and proportions (especially with anime sourced characters)

  • best quality - Less is more, this model was tuned to not need as many quality tags.

• Recommended Negative:

  • I typically don't use negatives with this model outside of shota, loli so keep it minimal or focused on the image you're trying to make, otherwise:

    • worst quality

• You can often omit pos or neg quality tags depending on what you're trying to make. Some of the other quality tags do work if you want those types of images.

Next Version & Future Plans

I am still collecting and setting up a dataset for training more expressive lighting, camera controls, and various realism focused features. I would like for anyone using Hyphoria Real to not need commonly used loras to get the most realistic and good looking images. Until then, most loras should work, if you do find any that don't please let me know.

As a long term goal I want to do a more comprehensive fine tune to cover the gaps that almost all realistic models have with prompt understanding, but beyond that I want my models to keep up with complex prompts but stay easy to use and fill in the gaps when you don't want to prompt for things specifically.

License & Use

SDXL - CreativeML Open RAIL++-M
Illustrious - https://freedevproject.org/faipl-1.0-sd/
NoobAI-XL - https://civitai.com/models/license/1140829