Guide to the Next “Beyond Experimental” Update

Hello everyone,

I’m excited to share what’s coming in the next update for Beyond Experimental Update (Alpha) 10.1

What I learned from the last version

In a previous build, I didn’t have enough balanced data. I had way more anime images than realistic ones, and that imbalance caused the anime style to overwrite the realistic style during training. I did get one model to work, but it wasn’t trained properly, so results could be blurry or disfigured. Looking back, I could’ve fine-tuned with higher-quality images; I just wasn’t thinking in that direction at the time.

What’s new: precise style control

This update introduces clear trigger pairs so you can reliably switch styles without guesswork.

Realistic images

• Positive trigger: source_realistic

• Negative trigger: BeyondNegativeR (cleans digital noise, motion blur, etc.)

Example

• Positive: source_realistic, a cinematic photo of a mountain landscape, masterpiece

• Negative: BeyondNegativeR

Anime images

• Positive trigger: source_anime

• Negative trigger: BeyondNegativeA (cleans messy lines, inconsistent details, etc.)

Example

• Positive: source_anime, a key visual of a magical girl, masterpiece

• Negative: BeyondNegativeA

A note on training & feasibility

I want to be transparent: I’m not 100% sure I can run the full training pass in one go. Training ~1,000,000 images across both sources is extremely resource-heavy. My tests push over ~100 GB of VRAM and ~86 GB of system RAM just to get through a single pass.

Most likely plan:

• Split the dataset so each half stays under 100 GB VRAM during training.

• Train the halves separately.

• Fine-tune afterward to keep both styles strong and stable.

This approach should keep the process manageable while preserving quality across realism and anime