The Red Lotus model operates without specific required keywords. However, like other models based on the Pony architecture, it utilizes the Pony scoring system, which significantly influences the overall style of the generated images. For optimal results, I personally use "score_9, score_8_up, score_7_up" in the positive prompts and "score_6_up, score_5_up, score_4_up" in the negative prompts; however, experimenting with different combinations can yield even better results.
While the base model excels at producing 2.5D images, it is also capable of generating 2D images by employing the "source anime" tag or tags like "nidy." Additionally, you can achieve 2D outputs by including negative prompts with tags such as "3D, game cg, realistic, photorealistic." This responsiveness allows the model to adapt to various artistic needs and preferences.
In certain scenarios, the inclusion of specific tags is crucial; without them, slight biases may occur in the outputs. For the best performance, I recommend setting the CFG scale between 5 and 7, with 6.9 being optimal. The steps should ideally range from 25 to 40, with 30 suggested for balanced quality.
While extensive testing with other samplers is still pending, I have found that the Euler sampler works exceptionally well with this model.
The Red Lotus model operates without specific required keywords. However, like other models based on the Pony architecture, it utilizes the Pony scoring system, which significantly influences the overall style of the generated images. For optimal results, I personally use "score_9, score_8_up, score_7_up" in the positive prompts and "score_6_up, score_5_up, score_4_up" in the negative prompts; however, experimenting with different combinations can yield even better results.
While the base model excels at producing 2.5D images, it is also capable of generating 2D images by employing the "source anime" tag or tags like "nidy." Additionally, you can achieve 2D outputs by including negative prompts with tags such as "3D, game cg, realistic, photorealistic." This responsiveness allows the model to adapt to various artistic needs and preferences.
In certain scenarios, the inclusion of specific tags is crucial; without them, slight biases may occur in the outputs. For the best performance, I recommend setting the CFG scale between 5 and 7, with 6.9 being optimal. The steps should ideally range from 25 to 40, with 30 suggested for balanced quality.
While extensive testing with other samplers is still pending, I have found that the Euler sampler works exceptionally well with this model.