<h3 id="heading-193">Use weight 0.92 / don't use on diffuser backend on SD.Next</h3>Fares Fares (born 29 April 1973) is a great Swedish-Lebanese actor of Assyrian descent. Known for his performances in the video game A Way Out (2018), TV series Westworld (2018), Chernobyl (2019) and The Wheel of Time (2021)...I really don't like TI. I don't intend to train it anymore. It's hard to get to resemblance and it looks OKish at best compared to Lora.It can be worth to use together with another lora to "stabilize" the other loras changes and rigidness.For example: if I want Fares Fares to have a pink colored hair and a really long beard, the Fares Fares SD15 lora is hard at that because it knows Fares Fares don't have those characteristics (even using a Leco slider Lora). So using the embedding apparently breaks that a little.But as a trade off it will make the final result more "standard". For example with fantasy loras, like the <a target="_blank" rel="ugc" href="https://civarchive.com/models/139149?modelVersionId=153967">Venomancer</a>, the SD Lora alone it's way more flexible and play together nicely. It's more "fantasy like" than this TI alone or they combined.Also embedding + lora can SOMETIMES increase resemblance. Some other times it will get crappy and saturated. It's inconsistent but it happens, I'll post some examples.So for my recommendation is:TI alone: 0.92TI + Fares Fares Lora: TI:0.92 + Lora:0.80.92 Yes it's very specific. You can lower it if its too much saturated, but will lose resemblance. Use saturated color on the negative prompt.For this SD15 TI I did 20 (!) training versions, they all looked like crap, until I realized that <a target="_blank" rel="ugc" href="http://SD.Next">SD.Next</a> diffuser backend simply don't work well with safetensors embedding, I guess. SO don't use diffuser backend for SD15 things, it's not good. The diffuser backend is super fast but even with normal Loras, it gets weird.I feel like using the prompt keyword "Fares Fares" is not needed. But for using together with Fares Fares lora, might be a good idea to use it at "(Fares Fares:0.8)"There is no Keywords, use the name of the embedding:<pre><code>(Fares_Fares_TI_SD15:0.92)</code></pre>  Negative supporting prompts:<pre><code>glasses, saturated colors, dim lit, shadow face</code></pre> <hr />Try out my other SD 1.5, SDXL and TI here:<a target="_blank" rel="ugc" href="https://civarchive.com/models/179471?modelVersionId=201404">Fares Fares – Lora SD15</a><a target="_blank" rel="ugc" href="https://civarchive.com/models/178612?modelVersionId=200502">Fares Fares – Lora SDXL</a>Fares Fares – TI SDXL (will update later)<hr />The base checkpoint is the realisticVisionV51_v51VAE 3.97GB. I used it instead of the original 1.5 because this is not a broad concept. To get best resemblance you should use the best model for it, so use realistic vision 5.1.<h3 id="heading-194">Please use high-res fix to get better results</h3>CFG: 5.5Clip skip: 1512px DPM++ 3M Exponential (50 steps or more)DPM++ 2M Karras (25 Steps or more)DPM++ 2S a Karras Current problems with the Lora:<ul><li>Saturate colors at high weight</li><li>loose resemblance if below 0.8</li></ul> Some settings: Trained 768res. 177 pictures. Token name "Fares Fares". Initialization word "man".Because I did so many versions (20) and the best one was the second and Kohya setting on TI "Training comment" is not saved I don't know how many steps this was. I THINK it was 10 steps for best pictures, 5 steps for OK pictures. 2 different folders (sharing the same “Fares Fares person" concept).I used AdamW, constant, 12 epochs. Gradient accumulate steps at 1. Batch size 1.It would be around 16.620 steps I think.Hopefully you can leave some results and some comments. Any idea is appreciated. Thank you.

Fares Fares - Textual Inversion - SD 1.5

Weight it at 0.92. This was epoch 12 out of 20. v2.

Files with SHA256: 673817e01d63174ec764a911e222e8c5063b934ed474eb14e7a7f26bb42e6729