Explanation for each parameter + Tips
\n", 852 | "
prompt frog:0.5 dog:0.7) The weights control how much the prompt is supposed to be paid attention to. Try to stick around 0.7-1.3 when all weights are added up for \"normal\" results. [[red]] (frog))\n",
858 | " Also included are prompt tags for easy prompt design, beginner prompt..alchemists may not have an intuition for how to change something simple (like \"red bird\") into a prompt that works well. Included are the tags: {artstation}: for that \"trending on artstation\" feel {overwatch}: a painterly overwatch-like fanart style {ghibli}: like a ghibli movie {intricate}: self explanatory. init_img \n",
869 | " classifier_guidance = True
\n",
870 | " lpips_guidance = True
\n",
871 | " lpips_scale = 8
\n",
872 | " loss_scale = 0
\n",
873 | " \n",
874 | " this will push the model to use that sort of pose but not necessarily keep every detail about it, this can give higher fidelity results but at some time cost. (Also it doesn't always .. keep the pose but it still increases fidelity anyway? I don't know how to explain it 100%) (example: link) size [height, width] steps start_step init_img section, start_step should be a number below whatever you set steps to. It just skips that far into the generation and should be used when you have an init_img set.\n",
890 | " perlin_multi scale seed batches lpips_scale init_img. Pushes the image to be structurally similar to the init_img. If 0 the lpips model will not be loaded, and the generation time will be faster. See the init_img documentation for tricks with this.\n",
916 | " clip_scale clip_text_prompt and clip_image_prompt. If 0 the CLIP model will not be loaded, and the generation time will be faster. I recommend a max of 0.2 and an average around 0.1. CLIP guidance is NOT DETERMINISTIC! Outputs are not 100% reproducible.\n",
920 | " tv_scale clamp_max \n",
929 | " magnitude = cond_grad.square().mean().sqrt()
\n",
930 | " cond_grad = cond_grad * magnitude.clamp(max=args.clamp_max) / magnitude\n",
931 | " \n",
932 | " quick_guidance clip_text_prompt prompt.\n",
940 | " clip_image_prompt clip_text_prompt and prompt, it should be denoted with a vertical bar. (url|weight)\n",
944 | " clip_model_name and clip_model_pretrained clip_scale isn't equal to 0. [('ViT-B-32', 'openai'),
\n",
950 | " ('ViT-B-32', 'laion400m_e31'),
\n",
951 | " ('ViT-B-32', 'laion400m_e32'),
\n",
952 | " ('ViT-B-32', 'laion2b_e16'),
\n",
953 | " ('ViT-B-32', 'laion2b_s34b_b79k'),
\n",
954 | " ('ViT-B-32-quickgelu', 'openai'),
\n",
955 | " ('ViT-B-32-quickgelu', 'laion400m_e31'),
\n",
956 | " ('ViT-B-32-quickgelu', 'laion400m_e32'),
\n",
957 | " ('ViT-B-16', 'openai'),
\n",
958 | " ('ViT-B-16', 'laion400m_e31'),
\n",
959 | " ('ViT-B-16', 'laion400m_e32'),
\n",
960 | " ('ViT-B-16-plus-240', 'laion400m_e31'),
\n",
961 | " ('ViT-B-16-plus-240', 'laion400m_e32'),
\n",
962 | " ('ViT-L-14', 'openai'),
\n",
963 | " ('ViT-L-14', 'laion400m_e31'),
\n",
964 | " ('ViT-L-14', 'laion400m_e32'),
\n",
965 | " ('ViT-L-14', 'laion2b_s32b_b82k'),
\n",
966 | " ('ViT-L-14-336', 'openai'),
\n",
967 | " ('ViT-H-14', 'laion2b_s32b_b79k'),
\n",
968 | " ('ViT-g-14', 'laion2b_s12b_b42k')]
\n",
969 | " I would suggest 'ViT-B-32', 'laion2b_s34b_b79k'
cutn