Type | |
Stats | 572 249 |
Reviews | (118) |
Published | Sep 24, 2023 |
Base Model | |
Trigger Words | qingS, bodysuit, asymmetrical legwear, single mechanical arm, long_sleeves, cowl, gloves, headdress, shoulder_armor, boots, 1girl, brown hair, long hair, yellow eyes, hair_between_eyes |
Hash | AutoV2 19318C4BE9 |
请阅读使用说明，请使用google翻译。Please read the instructions, please use google translate。
説明書を読んで、Google翻訳を使用してください 。사용설명서를 읽고 Google 번역을 사용하세요.
权重Weights：0.6-0.85，推荐recommend: 0.65-0.7
NSFW:0.55-0.7
————————————————————————————————————————————————————————
推荐使用：高分辨率修复 (Hires. fix)、IMAGE-TO-IMAGE / controlnet-tile 放大重绘
【推荐使用的Checkpoint】：参考XYZ图
【负面-embaddings】:EasyNegative、negative_hand-neg
【CFG】：7
【VAE】: kl-f8-anime2
——————————————————————————————————————————————————————
【Trigger Words】
1.正面提示词：qingS, 1girl, bodysuit, asymmetrical legwear, single mechanical arm, long_sleeves, cowl, gloves, headdress, shoulder_armor, boots, brown hair, long hair, yellow eyes,
2.负面提示词：3D_rendering, white background, simple background, weapon, holding sword, holding handgun, gear
3.负面embaddings: EasyNegative, negative_hand-neg
【参考预设，您可以在此之上进行修改】
1.正面提示词：
qingS, 1girl, solo, (bodysuit:1.2), asymmetrical legwear, single mechanical arm, long_sleeves, cowl, gloves, headdress, shoulder_armor, boots, brown hair, long hair, yellow eyes,looking at viewer, (naughty_face:1.2), standing, (cowboy_shot1.2), park, tree, flowers, blue_sky, day, dappled sunlight, (masterpiece,best quality,beautiful and aesthetic:1.2), anime coloring, <lora:qingS-v1:0.65>
2.负面提示词：EasyNegative (negative_hand-neg:1.2) 3D_rendering, white background, simple background, weapon, holding sword, holding handgun ,(worst quality,low quality,normal quality:1.4) , gear
3.其它参数：DPM++2M SDE Karras 40 steps , 512X768 / 512x1024， 启用Hires.fix, R-ESRGAN 4X+Anime6B, 放大倍数：1.5-2，重绘幅度0.3-0.4
——————————————————————————————————————————————————————
【目前存在的问题与建议】
1.（1）在绘制远景镜头的图片时（full body），人物的眼睛与面部精度有概率出现瑕疵，但可以用以下方式【方法3】修复。
（2）人物服饰有概率出现错误，但不会影响整体人物观感。 提示词：“single mechanical arm”的还原效果不理想。如果不想露肩，请把提示词“shoulder_armor”改为“armor”。根据实际情况决定是否将提示词“gear”加入到负面提示词中
（3）实例使用的质量提示词仅供参考，您可以使用其它的质量提示词达成不同的效果
(4) 根据您使用的模型与实际出图情况来调整权重
(5) 如果您生成的图片与示例差距过大，请检查权重值、文件名是否正确，是否使用了负面embaddings
2.在绘制远景或全身画面时，请适当增大出图尺寸与步数，这有助于提升人物面部精度。
3.如果以上方法【2】的效果不理想，请直接将生成的图片发送至image-to-image进行放大重绘【效果很好】，方法如下：
（1）尽量使用与原图一致的提示词、角色lora、seed
（2）启用controlnet--tile, 启用 放大脚本 Ultimate SD upscale
（3）一个参考设置参数：Steps:30-40, 重绘幅度0.2-0.4，放大算法：4x-UltraSharp 或 R-ESGAN 4x+anime6B, 放大倍数1.25--
4.在创作NSFW内容时，只需使用提示词【3】，请根据实际情况调整权重，同时使用更多更精确的NSFW相关提示词，增加NSFW相关提示词的权重，根据出图情况，将提示词中的衣服类Trigger Words 加入负面提示词。
5.当您感觉图片细节不足、模糊时，请适当增加迭代步数，或调整采样器及vae,也可以尝试【3】的方法进行I2I重绘。
6.Checkpoint兼容性问题，您可能需要调整LORA权重以及各个提示词的权重、增减提示词来达成效果
————————————————————————————————————————————————————————————————————————
Weights：0.6-0.85，recommend: 0.65-0.7
NSFW:0.55-0.7
——————————————————————————————————————————————————————————————————————
Recommended use: High-resolution fix (Hires. fix), IMAGE-TO-IMAGE / controlnet-tile enlargement and redrawing
[Recommended Checkpoint]: Refer to the XYZ diagram
【Negative-embaddings】:EasyNegative、negative_hand-neg
【CFG】：7
【VAE】: kl-f8-anime2
——————————————————————————————————————————————————————————————————
【Trigger Words】
1. Positive prompt words: qingS, 1girl, bodysuit, asymmetrical legwear, single mechanical arm, long_sleeves, cowl, gloves, headdress, shoulder_armor, boots, brown hair, long hair, yellow eyes,
2. Negative prompt words: 3D_rendering, white background, simple background, weapon, holding sword, holding handgun, gear
3. Negative embaddings: EasyNegative, negative_hand-neg
[Reference default, you can modify it]
1. Positive prompt words:
qingS, 1girl, solo, (bodysuit:1.2), asymmetrical legwear, single mechanical arm, long_sleeves, cowl, gloves, headdress, shoulder_armor, boots, brown hair, long hair, yellow eyes,looking at viewer, (naughty_face:1.2), standing, (cowboy_shot1.2), park, tree, flowers, blue_sky, day, dappled sunlight, (masterpiece, best quality, beautiful and aesthetic:1.2), anime coloring, <lora:qingS-v1:0.65>
2. Negative prompt words: EasyNegative (negative_hand-neg:1.2) 3D_rendering, white background, simple background, weapon, holding sword, holding handgun, (worst quality, low quality, normal quality: 1.4), gear
3. Other parameters: DPM++2M SDE Karras 40 steps, 512X768 / 512x1024, enable Hires.fix, R-ESRGAN 4X+Anime6B, magnification: 1.5-2, redraw amplitude 0.3-0.4
——————————————————————————————————————————————————————————————————————————————
[Current problems and suggestions]
1. (1) When drawing a long shot picture (full body), there is a probability that the character's eyes and facial accuracy will be flawed, but it can be repaired in the following way [Method 3].
(2) There may be errors in character clothing, but it will not affect the overall character look and feel.
Tip word: The restoration effect of "single mechanical arm" is not ideal.
If you don't want to show your shoulders, please change the prompt word "shoulder_armor" to "armor".
Decide whether to add the prompt word "gear" to the negative prompt words according to the actual situation
(3) The quality prompt words used in the example are for reference only. You can use other quality prompt words to achieve different effects.
(4) Adjust the weight according to the model you use and the actual drawing situation
(5) If the difference between the image you generated and the example is too large, please check whether the weight value and file name are correct, and whether negative emaddings are used
2. When drawing distant views or full-body scenes, please increase the drawing size and number of steps appropriately, which will help improve the facial accuracy of the characters.
3. If the effect of the above method [2] is not satisfactory, please directly send the generated image to image-to-image for enlargement and redrawing [the effect is very good], the method is as follows:
(1) Try to use prompt words, characters lora, and seed that are consistent with the original picture.
(2) Enable controlnet--tile, enable the amplification script Ultimate SD upscale
(3) A reference setting parameter: Steps: 30-40, redraw amplitude 0.2-0.4, amplification algorithm: 4x-UltraSharp or R-ESGAN 4x+anime6B, amplification factor 1.25--
4. When creating NSFW content, just use the prompt word [3]. Please adjust the weight according to the actual situation. At the same time, use more and more precise NSFW-related prompt words to increase the weight of NSFW-related prompt words. According to the picture production situation, Add negative prompt words to the clothing Trigger Words in the prompt words.
5. When you feel that the picture details are insufficient or blurry, please increase the number of iteration steps appropriately, or adjust the sampler and VAE. You can also try the method [3] for I2I redrawing.
6. Checkpoint compatibility issues, you may need to adjust the LORA weight and the weight of each prompt word, and increase or decrease prompt words to achieve the effect.