Description:
This model can generate more impactful and epic images, with much better image details than the basic XL, more impactful image composition, richer elements, better hand understanding than the basic XL, and more accurate understanding of prompt words. The effect is more obvious for words generated by GPT4O labeling. If you need to use this model as a base model to train Lora, please use GPT4 labeling, which is easier to fit.
Model self-evaluation:
3D large scenes, CG pictures have excellent output effects
IP images, PVC characters, blind box design effects have excellent output effects
Automobiles, mechas, and machinery have excellent output effects
Horror, disasters, and monsters have excellent output effects
Product design and industrial design have excellent output effects
Interior design, architectural design, and natural scenery have good output effects
Logo design and graphic design have good output effects
Real portraits, average output effects
Graphic illustrations and 2D comics have average output effects
For good or average output effects, lora can be used to improve quality and effects.
Parameter recommendation:
CFG: 7
Iteration steps: 20-30 for the display
Sampling method: Euler a (balanced, no bright spots or weaknesses)
DPM++ 2M Karras (suitable for flat images)
DPM++ SDE Karras (recommended, suitable for 3D/realism)
HD upscaling algorithm:
R-ESRGAN_4x+, Anime6B