<h1 id="heading-3411">Version 7.</h1>Fixed the issue with body proportions. You still can obtain extreme body proportions with loras of course.Right now, I'm using:<ul><li>UI: ComfyUI with <a rel="ugc" href="https://github.com/dchatel/comfyui_facetools">comfyui_facetools</a> (a rotation aware face extractor/merger made by me) to fix the faces</li></ul><ul><li>positive quality prompt: masterpiece, best quality, drop_shadow, subsurface scattering</li><li>negative quality prompt: extra_hands, extra_arms, bad_anatomy, fat, overweight, abs, muscular, fat_rolls, pubic_hair badhandv4 realisticvision-negative-embedding epiCNegative</li><li>1st pass:<ul><li>sampler: dpm++ 3m sde</li><li>cfg: 3.8</li><li>steps: 20</li></ul></li><li>HiRes and subsequent (like face detailer) pass:<ul><li>sampler: dpm++ 2m sde</li><li>cfg: 4</li><li>steps: 8</li></ul></li><li>upscaler:<ul><li>model: 4x_NMKD-Siax_200k</li><li>controlnet: I know this can be a bit surprising, but it works really well in my use cases. The trick is to give some freedom to the inpaint controlnet with <code>Scaled Soft Weights </code> and inpaint the whole image, no mask. So, if there was no <code>Scaled Soft Weights</code>, or if <code>base_multiplier</code> was set to 1.0, it would have no effect, as it would repaint the same picture as the one provided in <code>image</code><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/730f051e-bcc6-40c5-9242-d2ccf83d6ac1/width=525/730f051e-bcc6-40c5-9242-d2ccf83d6ac1.jpeg" /></li></ul></li></ul><h1 id="heading-3412">Version 6.</h1>cfg scale can be quite different. I got nice results from 2.5 to 7 (I haven't really tried higher). Also v6 doesn't even need quality tokens in the positive prompt to give good results. At least as long as realisticvision-negative-embedding and epicrealism-neg are in the negative prompt.<h1 id="heading-3413">Version 5.</h1>The model should require less token in both positive and negative prompt to get realistic results, now.Low CFG scale ~2.5 gives a nice photo-realistic look to the picture and allows you to go a bit crazier with the weighting in the prompts.Heun sampler with as low as 8 steps also gives quite nice results.<hr /><h1 id="heading-3414">Version 4.</h1>Initially built to create lifelike portraits of characters from the novel I'm working on, whose title is Endless, thus the name of the model.This is a rather complex merge, and sadly I did not took notes.PromptingExample of prompt for a realistic image (replace {prompt} with your subject):<pre><code>professional photograph
{prompt}
((video grainyfilm)), natural cinematic lighting, ((professional photo highly defined))
realistic, best quality, photo-realistic
8k, best quality, masterpiece, realistic, photo-realistic</code></pre>You can add more details using this lora which gives pretty good results: <a target="_blank" rel="ugc" href="https://civarchive.com/models/82098/add-more-details-detail-enhancer-tweaker-lora">Detail Tweaker LoRA (细节调整LoRA)</a>Example of negative prompt for a realistic image:<pre><code>(text, signature, watermark, logo, ux, ui:1.3), monochrome, (2D), (3D), ((camera)), ((cartoony)), (CGI), ((drawing)), ((lowres)), ((painting)), ((sketchy)), ((The Sims)), (ugly), ((blur)) (worst quality, low quality, normal quality:1.3), lowres</code></pre>NSFW content is possible, nsfw, uncensored in the negative prompt can help get rid of them.More tips on promptingYou can actually use negative weights on lora/lycoris. Here is an obvious example with <a target="_blank" rel="ugc" href="https://civarchive.com/models/4602/microwaist">Microwaist Lora</a>:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/0eba4cf0-64dd-432d-bd9a-4e4b2d3d307b/width=525/0eba4cf0-64dd-432d-bd9a-4e4b2d3d307b.jpeg" />Using this, you can use <a target="_blank" rel="ugc" href="https://civarchive.com/models/69267/sharpness-tweaker-lora-lora">Sharpness Tweaker LoRA (锐度调整LoRA)</a> to lower the sharpness of the image, resulting in a somewhat more realistic picture as if it had been taken by an old camera. Here is an example showing the effect on girls of various ethnicities:<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c6b4431e-e864-4945-b858-a2243938c7e5/width=525/c6b4431e-e864-4945-b858-a2243938c7e5.jpeg" /><a target="_blank" rel="ugc" href="https://i.imgur.com/gfyXRfd.jpg">Full resolution of the above grid on imgur</a>

endlessReality

v2 is also known as stage9.This version has an added VAE.As previously, I tried to maintain high resolution quality and realism. I found it works pretty well with the 8x_NMKD-Superscale_150000_G upscaler which you can find here : <a target="_blank" rel="ugc" href="https://icedrive.net/s/43GNBihZyi">https://icedrive.net/s/43GNBihZyi</a>.Compared to the previous version, this one should probably have a bit sharper and less messy skin textures, while still maintaining the beautiful skin imperfections.You can play with the factor in (worst quality, low quality:1.7), reducing it to 1.3 for example will yield an image where skin is a less textured. This can be useful if you find that the skin is too shiny or has too many moles.

Files with SHA256: 3576ae9275a69d859f34ce476c6a5c59695b8383a64676ddc2f1d792340fcc50