VirtualMagicGirl

This model features the high level of creativity typical of IllustriousXL-based models while remaining highly realistic; it offers greater creative freedom without sacrificing a remarkable degree of realism. I consider it an excellent balance between creativity and realism for this type of architecture—or at least, the best balance I have managed to achieve so far. Although it is still a work in progress, it allows for the generation of striking images that can be used to create impressive videos. It is compatible with Illustrious and XL LoRA models, though I recommend not overusing LoRAs and sticking to only those strictly necessary to add concepts not present in the base model. Notably, the model is already capable of independently generating a wide range of scenarios that would typically require a LoRA. It is also quite flexible and handles high levels of prompt emphasis well, though moderation is still advised. It tends to lean toward NSFW content if detailed descriptions of intimate body parts are included; this tendency can be partially mitigated by avoiding superfluous descriptions and adding "NSFW" to the negative prompt with a high weight (1.7–2.0) when aiming to create SFW content and avoid depictions of intimate areas.

Recommended parameters:

Sampling method: Euler A, DPM++ 2M, Heun

Schedule: Automatic (Karras)

Steps: 25–35

Skip: 2

CFG Scale: 4–7 (max 7.5)

Upscaler: RealESRGAN (or original)

Hires steps: 20–35

Denoising strength: 0.15–0.4

The following VAE is recommended:

fixFP16ErrorsSDXLLowerMemoryUse_v10.safetensors