Type | |
Stats | 3,873 |
Reviews | (414) |
Published | Aug 1, 2023 |
Base Model | |
Training | Steps: 175,000 |
Hash | AutoV2 34E356F16D |
The new version of MBBXL has been trained on >18000 training images in over 18000 steps. It's probably the most significant fine-tune of SDXL so far and the one that will give you noticeably different results from SDXL for every prompt.
The model tends towards a "magical realism" look, not quite photo-realistic but very clean and well defined.
Very flexible model that looks good with most prompts (including simple prompts).
Use of the SDXL Refiner is not recommended with this model. Shouldn't be needed either.
Recommended negative:
blurry, blurred detail, poorly hands, poorly face, enhanced hands, missing fingers, mutated hands, fused fingers, deformed, malformed limbs, disfigured, watermarked, text, extremely grainy, very chromatic aberration, oversaturated
Recommended settings: 1024x1024 or 1336x768.
DPM++ SDE 30 steps or more.
Works very well with hires.fix, try 768x768 upscaled to 2x!
Limitations:
Skin textures are not strong, to come later probably as an add-on lora.