Here we have Lyndis or Lyn my favorite FE waifu.
Rath, Kent, Hector and Eliwood? Do you mean the retards randos strays heroic guys who sacrificed themselves in seemingly pointless combat to cover our retreat? They really shouldn't have even dreamt of putting their dirty hands on the Tactician's woman.
That mechanic of sword of fire is weird, almost like state sponsored Netorase(a type of NTR where the cuckold incites his own cuckoldry). So yes I consider the one who implemented that mechanic to be that particular type of deviant.
In my head canon the Tactician ends up with Lyn and Florina while the kingdoms go to war for his expertise. Any other option is the NETORARE sponsored option and will be disregarded with extreme prejudice.
Seems to work fine with: <lora:LyndisV3:0.7> lyndis,
I was able to isolate 3 outfits and a variant and also a couple of parts:
Blue_dress_side_slit_Pelvic_curtain_Sash_Short_sleeves_Belt: Her original outfit.
White_strapless_dress_Long_detached_sleeves_Blue_sash_White_choker_Bridal_veil: Lyn's wedding dress.
Blue_long_dress_Juliet_sleeves_Blue_choker_Headdress_Yellow_underbust: a blue party dress.
Blue_dress_side_slit_Pelvic_curtain_Sash_Short_sleeves_Belt_Blue_armor_Fur_trim: Variant of her original outfit with some chest armor and a pelt.
black_fingerless_gloves: The black gloves she uses with her original outfit.
brown_knee_boots: The black boots of her original outfits.
candle_bouquet: The sword bouquette candle thingy for her weeding dress.
As always I will be adding the dataset with the tag summary so people can experiment.
Remember, say no to netorare and netorase and yes to NETORI. Vote KNXO for eternal president of the universe.
Description
FAQ
Comments (8)
Hey for some reason I just can't replicate the very first image you posted with this lora. Everything is the same, resolution at 512 x 1024 and yet it's always different. I'm pretty new at stable diffusion so is there something I don't know about?
same prompt, same negative prompt, same seed, same cfg, same number of sampling steps (although it seems to be closer to yours the more steps I add idk why), same clip skip and same sampling method. VAE doesn't seem to influence anything besides colors of the final result. Can you help me?
If you are using the same model anything v4.5 with the correct checksum, the problem is likely the embeddings or Textual inversions, i use a couple of them in the negatives so you might be missing some of them
Checking the img info if it is the one i think, it seems to not have detected bad hands-5 https://civitai.com/models/116230/bad-hands-5 otherwise If you want the exact same image it might be due the ENSD that value can be configured in A1111 and in relates to the seed numbers and was inherited from the original Leaked NAI model.
Here's the image full metadata:
<lora:LyndisV3:0.7> lyndis, huge_breasts, standing, solo, Blue_dress_side_slit_Pelvic_curtain_Sash_Short_sleeves_Belt,, masterpiece, best quality, detailed face, detailed eyes, highres,
Negative prompt: art by bad-artist, bad-image, lowres, bad anatomy, bar censor, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, out of focus, censorship, censored, deformed, amateur drawing, odd, morphing, ((depth of field)), simple background, artist name, subtitled, ((split screen)), (multiple boys), patreon username, patreon logo, multiple views, bad hands, Missing vagina, Blurry faces, Blank faces, bad face, Ugly, extra ear, amputee, missing hands, missing arms, missing legs, Extra fingers, 6 fingers, Extra feet, Missing nipples, ghost, futanari, Extra legs, Extra hands, Extra arms, bad-hands-5, easynegative, (2boys)
Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 1186587369, Size: 512x1024, Model hash: 6e430eb514,
Model: anything-v4.5-pruned, VAE hash: 600345c503, VAE: clearvae_main.safetensors, Clip skip: 2, ENSD: 31337, Lora hashes: "LyndisV3: 2adf5bc4b635", TI hashes: "bad-artist: 2d356134903e, bad-image: 5b9281d7c6ad, bad-hands-5: aa7651be154c, easynegative: c74b4e810b03", Version: v1.6.0, Hashes: {"vae": "600345c503", "embed:bad-artist": "2d35613490", "embed:bad-hands-5": "aa7651be15", "embed:bad-image": "5b9281d7c6", "embed:easynegative": "c74b4e810b", "lora:LyndisV3": "b8336e05e3", "model": "6e430eb514"}
@knxo So I experimented quite a bit after your reply, but I still can't get the exact same, although it's much closer than it was before for sure.
Right now there's a 2 things that confuse me:
First, I have no idea what a checksum is and google isn't helping at all. And I couldn't find a setting for it.
Second, Are you saying that the resources associated with the image are not always the right ones and that there could be a mistake? Because I don't know if you're aware, but the anythingelseV4 checkpoint you linked, according to the description, is not an official version of it. So did you truly use the one you linked?
What I initially did (before I posted the original comment) was copy paste both the prompt and negative prompts and download everything that was linked as a resource used and made the settings the same.
After your reply, I changed the ENSD like you said. But then when it came to the negative embeddings I had a ton of questions. I had simply copy pasted your negative embedding, but when I went in "textual inversions" and clicked on my own embeddings, it added something to the prompt automatically. Specifically for bad image and bad artist. In your prompt, bad image is in the prompt as just "bad-image". While for me, when I click on it, it ads "bad-image-v2-39000" to the negative prompt. This is odd, since you're the one who posted the download link to this embedding yourself. So are negative embeddings activated only when you click on them and it inputs the specific text, or are they still activated if you vaguely have the same thing in your prompt? As for bad artist, instead of "art by bad-artist", it was "By bad artist -neg". I've never seen "-neg" before so it added more questions. Was I supposed to put this in the positive prompt instead (since "-neg" would render it negative") or remove it entirely? As for bad hands and easy negative, when I clicked on them, it removed the original text that was associated to it from the negative prompt. This seemed to imply to me that this text was exactly the one required. So if I understand correctly, if when you click on it, it ads text, it means it wasn't activated before but if it removes text, it means it was already activated?
So right now I don't know what I'm supposed to do. Am I supposed to stick to the exact negative prompt despite the odd things that happened, or am I supposed to add the things that are automatically added when I click on the negative embeddings. If so, in what order? Or am I supposed to replace your key words with the exact text that is automatically added when I click on them, even in the same order (so instead of being at the end, "bad-image-v2-39000" would be the first thing in the negative prompt). This is what I did for the following metadata, which is, as far as I can tell, the closest to the image I'm trying to replicate:
<lora:LyndisV3:0.7> lyndis, huge_breasts, standing, solo, Blue_dress_side_slit_Pelvic_curtain_Sash_Short_sleeves_Belt,, masterpiece, best quality, detailed face, detailed eyes, highres, Negative prompt: bad-image-v2-39000, By bad artist -neg, lowres, bad anatomy, bar censor, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, out of focus, censorship, censored, deformed, amateur drawing, odd, morphing, ((depth of field)), simple background, artist name, subtitled, ((split screen)), (multiple boys), patreon username, patreon logo, multiple views, bad hands, Missing vagina, Blurry faces, Blank faces, bad face, Ugly, extra ear, amputee, missing hands, missing arms, missing legs, Extra fingers, 6 fingers, Extra feet, Missing nipples, ghost, futanari, Extra legs, Extra hands, Extra arms, bad-hands-5, easynegative, (2boys) Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 1186587369, Size: 512x1024, Model hash: 6e430eb514, Model: anythingelseV4_v45, VAE hash: ab33ad55c6, VAE: clearvae_v23.safetensors, Clip skip: 2, ENSD: 31337, Lora hashes: "LyndisV3: 2adf5bc4b635", TI hashes: "bad-image-v2-39000: 5b9281d7c6ad, By bad artist -neg: 2d356134903e, bad-hands-5: aa7651be154c, easynegative: c74b4e810b03", Version: v1.7.0
I'm pretty confused right now and I tried everything I could think of with the negative embedding prompt, but it's still not quite the same.
I hope this isn't too much text or too hard to understand.
The embeddings are invoked with the name of the file. Civit renames them. So they will have a different name for you. You can either rename the file or change it in my prompt from "bad-image" to "bad-image-v2-39000"
A checksum is literally what the name implies, remember that programs are made by 0 and 1 so literally you sum the contents and create a unique signature. If the check sum matches it means it is the exact same file.
As for the data published by civitai... well not to put them down but it has had more bugs than an ant colony, when comparing what it displayed vs the metadata I quickly noticed it was missing the band hands 5 embed. I think all the others were there.
For the embeds having the -neg part, that's civitai renaming things to make sure you know it is a negative embedding, you could disregard it and use it in the upper prompt but it is trained to produce crap. so if it is in the normal prompt it produces crap if in the negative prompt it removes crap.
Looking at your prompt it looks close, it might just be the spaces in the by bad artist part.
@knxo Thank you so much for all the explanations! It's now extremely close to yours, so while I could try to figure out what's missing, it may not be worth it other than pure curiosity. The only lead I have right now is that my metadata doesn't show hashes I think? Also "b8336e05e3" is somewhere in your metadata while it's nowhere in mine, and that's linked to the lyndis model itself so it can't be wrong, can it?
As a last question, does the specific ENSD you have have any benefits? From what I understand, this value is simply a group of seeds, nothing more, so there can't be anything that makes one better than the other right? If so, why do you choose this one specifically? Is there something I don't understand or that I'm missing?
Regardless, thank you so much.
@italianhulk I think the difference in the checksum of the lora is because of the version of the civitai plug in i was using at the time if you check it says 1.6 while your is 1.7. I checked the checksum of the lora and my image has the sha-256 checksum while your has the autov3 checksum just different standards.
About the ENSD, there's absolutely no benefit for it. It was used after the NAI leak to validate you had a working configuration as it was a bit more of a wild west days. Normally you would put the prompt and seed of an evangelion asuka and check the pose to see if you could replicate it to use as a sort of baseline. That's how most of the known default negatives were discovered and collected.
@knxo Ah I see. Well thanks again for your help. I was mostly frustrated that what I generated never looked nearly as good as everything else I saw, so I tried to challenge myself to replicate something that I thought looked good, so your first lyn image. Turns out what was wrong was my negative embeddings, but now everything looks really great. It's awesome, thanks so much.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.