Roach 1.5 | Richard Meier Style

V1.1 and V1.2 models have relatively weakened changes compared to V1.0, and are more sensitive to cues for aluminum plate curtain walls; the situation of a bluish tone seems to have improved; V1.2 has weakened generalization ability, slightly reducing the yield rate. If you want to diverge more for early-stage use in various scenarios, it is recommended to use the Lora version of V1.0.

Originally, I wanted to improve the weakening of V1.1's Mayer style, so I refined the Dan of V1.2. But the current situation seems imperfect.

V1.2 uses the SD1.5 official base model for refining, mainly for recommending large models in a realistic style.

Training resolution is 768*768.

————————————————

V1.2 recommendation: Lao Wang 0.5 base model: weight 0.6-0.7 (occasionally able to reach 0.8 withou* ***ng controlnet), iteration steps 50.

       Doll 1.7 base model: not recommended.

       SD1.5 official base model: weight 0.6-0.9, iteration steps 20-30.

       chilloutmix: weight 0.6-0.9, iteration steps 20-50.

V1.1 recommendation: Lao Wang 0.5 base model, weight 0.6-1.0, iteration steps 20-50.

       Doll 1.7 base model: weight 0.6-0.8, iteration steps 20-50.

       SD1.5 official base model: weight 0.6-0.8, iteration steps 20, not good effect.

       chilloutmix: weight 0.6-0.8, iteration steps 20-50.

V1.0 recommendation: Lao Wang 0.5 base model, weight 0.6-1.0, iteration steps 20-50.

       Doll 1.7 base model: weight 0.6-1.0, iteration steps 20-50.

       SD1.5 official base model: weight 0.6-1.0, iteration steps 20-50.

       chilloutmix: weight 0.6-1.0, iteration steps 20-50.

The following are the performances of the V1.0/V1.1/V1.2 models under various base models (Lora weight 0.8, iteration steps 50 steps+CN):

Before finding a better alchemy method, the Lora of 1.5 has stopped updating, detailed model recommendations are at the bottom of the page.

Wish everyone a pleasant playtime.

————————————————