【CHECKPOINT】Henmix_Real真实感女生融合henmixReal_v5c[B2EB92023A]

资源详情：
2024.01.20 ： V5C版本
该模型是在短暂中断后创建的。
它是以重新开始的心态开发的。它可能感觉与现有模型有很大不同，直到 v4。
请理解，我只执行了一个简单的合并。
由于合并模型的版权问题，我遵守非常严格的版权政策。
感谢您的理解。     
我使用了扩展 BMAB
https://github.com/portu-sim/sd-webui-bmab?tab=readme-ov-file
采样器：DPM++ SDE Karras、DPM++ 2M Karras、DPM++ 2S a
步数 ： 20~25
CFG ： 7.5 ~ 12.5

Hash parameter SHA256 B2EB92023AE246729DFABB655E84F85323C37829A1ABC224A55C1BC2BE320DAA CRC32 EFD87315 BLAKE3 C420DBFC04B5FC11F99D31F13C6F4601F46CBDD6CCEDDDDC4CBB2D07B3B55EBF AUTOV1 884F5E3B AUTOV2 B2EB92023A

作品展示：

Generation Data：

1. 1girl, (korean cute girl:1.2), (22 years old:1.3), (film grain:1.2), (bottomless:1.2), snowing, heavy snow, looking at viewer, (ugg boots:1.2)
   
2. Negative prompt: 3d, Drawings, abstract art, cartoons, surrealist painting, conceptual drawing, graphics, (low resolution:1.3), (worst quality:1.3), (low quality:1.3), collage, (watermark:1.3), letter, nude, dog, puppy
   
3. Steps: 50, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Size: 768x512, Seed: 129564112, Model: henmix_real_v5c, "Module: none, Sampler: DPM++ 2M SDE Karras, VAE hash: c6a580b13a, CFG scale: 18.5, Clip skip: 2, Model hash: b2eb92023a, BMAB noise end: 0.4, BMAB noise begin: 0, BMAB noise strength: 0.3, BMAB controlnet mode: lineart, BMAB_controlnet_option: "enabled=True,with_refiner=False,noise=True,noise_strength=0.3,noise_begin=0,noise_end=0.4"

复制代码

Generation Data：

1. 1girl,(RAW photo, best quality), (realistic, photo-realistic:1.2), 1girl, high quality, (high detailed skin:1.4), puffy eyes, gorgeous hair, <lora:epi_noiseoffset2:1.8>, (dark room:1.3), (rim lighting:1.3), (dimly lit:1.3), (dark night:1.3), indoors, portrait, black hair, dark background, short hair, <lora:koreanDollLikeness_v15:0.4>, <lora:fashionGirl_v47:0.4>
   
2. Negative prompt: (worst quality:1.4), (low quality:1.4), (monochrome:1.1), EasyNegative, bad-artist bad-image-v2-39000, bad_prompt_version2, ng_deepnegative_v1_75t, verybadimagenegative_v1.2-6400, vile_prompt3, bad-hands-5, painting, illustration,two face,2 girl
   
3. Steps: 20, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Size: 1024x1536, Seed: 2657663265, Model: henmixReal_v5c, Version: v1.7.0, Sampler: DPM++ 2M SDE Karras, VAE hash: c6a580b13a, CFG scale: 7, Clip skip: 2, Model hash: b2eb92023a, fashionGirl_v47: f3db8bc6be56", "epi_noiseoffset2: d1131f7207d6, koreanDollLikeness_v15: 3c8b9d1e9b98

复制代码

Generation Data：

1. (realistic:1.3) , finely detailed, quality, (masterpiece:1.2) , (photorealistic:1.2) , (best quality) , (detailed skin:1.3) , (intricate details) , ray tracing, dramatic, cute girl, As the intrepid traveler navigated the winding mountain trails, she marveled at the breathtaking views of the valleys below.
   
2. Negative prompt: 3d, Drawings, abstract art, cartoons, surrealist painting, conceptual drawing, graphics, (low resolution:1.3), (worst quality:1.3), (low quality:1.3), collage, (watermark:1.3), letter, nude
   
3. Steps: 20, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Size: 512x768, Seed: 3357147463, Model: henmix_real_v5m, "Module: none, Sampler: DPM++ 2S a, VAE hash: c6a580b13a, CFG scale: 12.5, Model hash: b2eb92023a, BMAB noise end: 0.4, BMAB noise begin: 0, BMAB noise strength: 0.4, BMAB controlnet mode: lineart, BMAB_controlnet_option: "enabled=True,with_refiner=False,noise=False,noise_strength=0.4,noise_begin=0,noise_end=0.4"

复制代码

Generation Data：

1. (by Callie Fink:1.1) , 1girl, korean cute Ethereal Female, Foggy conditions, Trendercore, side light, film grain, Instax, rich color
   
2. Negative prompt: (low resolution:1.3), (worst quality:1.3), (low quality:1.3), collage, (watermark:1.3), letter, nude,
   
3. Steps: 50, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Size: 512x768, Seed: 4212470241, Model: henmix_real_v5c, "Module: inpaint_only+lama, Sampler: DPM++ 2M SDE Karras, CFG mode: Half Cosine Down, VAE hash: c6a580b13a, CFG scale: 17.5, Clip skip: 2, Mimic mode: Cosine Up, Model hash: b2eb92023a, Mimic scale: 9, Interpolate Phi: 1, CFG scale minimum: 12.5, Scaling Startpoint: MEAN, Mimic scale minimum: 0, Variability Measure: AD, Threshold percentile: 99.9, Separate Feature Channels: True, Dynamic thresholding enabled: True, BMAB process_resize_by_person: 0.8, BMAB process_resize_by_person_ratio: 1.1

复制代码

Generation Data：

1. (RAW photo, best quality), (realistic, photo-realistic:1.2), 1girl, high quality, (high detailed skin:1.4), puffy eyes, gorgeous hair, <lora:epiNoiseoffset_v2:1.8>, (dark room:1.3), (rim lighting:1.3), (dimly lit:1.3), (dark night:1.3), indoors, portrait, black hair, dark background, short hair, <lora:koreanDollLikeness_v15:0.4>, <lora:breastinclassBetter_v13:0.4>, <lora:fashionGirl_v47:0.4>,
   
2. Negative prompt: (worst quality:1.4), (low quality:1.4), (monochrome:1.1), EasyNegative, bad-artist bad-image-v2-39000, bad_prompt_version2, ng_deepnegative_v1_75t, verybadimagenegative_v1.2-6400, vile_prompt3, bad-hands-5, painting, illustration,
   
3. Steps: 40, VAE: vae-ft-mse-840000-ema-pruned.ckpt, Size: 720x1024, Seed: 3589893419, Model: henmixReal_v5c, Version: v1.7.0, Sampler: DPM++ 2M SDE Karras, VAE hash: c6a580b13a, CFG scale: 7, Clip skip: 2, Model hash: b2eb92023a, bad-hands-5: aa7651be154c", vile_prompt3: 2d26d46daa6d, "EasyNegative: 66a7279a88dd, fashionGirl_v47: f3db8bc6be56", bad-image-v2-39000: 5b9281d7c6ad, bad_prompt_version2: 6f35e7dd816a, ng_deepnegative_v1_75t: 54e7e4826d53, "koreanDollLikeness_v15: 3c8b9d1e9b98

复制代码

下载地址：

游客，如果您要查看本帖隐藏内容请回复

哇，漂亮的模型。谢谢。V5c 创建有吸引力的图像。

I have been using these setting for Dynamic thresholding, but i see they are from V2.2, Are they still good for V4?

Dynamic thresholding enabled: True, Mimic scale: 10, Threshold percentile: 99.95, Mimic mode: Cosine Up, Mimic scale minimum: 0, CFG mode: Half Cosine Down, CFG scale minimum: 15

等待就是结果 发表于 2024-2-2 11:41
I have been using these setting for Dynamic thresholding, but i see they are from V2.2, Are they sti ...

我在 v4 中使用了它

启用动态阈值：True，模拟比例：10，阈值百分位数：99.9，模拟模式：余弦上，模拟比例最小值：0，CFG 模式：半余弦下，CFG 刻度最小值：12.5

这个模型非常偏向亚洲女性，但质量非常好。

我已经尝试了所有支持 NSFW 的模型（Reliberate、ICBINP、ConsistentFactor、FantasticMix 等）——这是最好的。默认情况下生成亚洲人，但能够生成带有提示（和“亚洲”否定提示）的所有种族。关闭剪辑跳过有助于更好地遵循提示，但点击次数会更少。还能够以令人印象深刻的准确性生成各种各样的公众人物。即使在CFG20上也取得了出色的成绩！

我认为预览图片的名称和缺乏多样性使人们认为这种模式只能是亚洲人和轻微的动漫风格，但事实并非如此。

随着版本的增加，学习进展顺利，但图像类似于 chiloutmix。就我个人而言，我认为 v1.0 更现实。

这个模型很真实，但是必须要结合Adetailer或者DDetailer使用，不然画面非常渣。另外还有两个问题，一个是出图速度极其的慢，二是用了这两个插件以后，跑3张图只有第一张能被controlnet控制。

我的照片是涂鸦，我用什么VAE？谢谢