>>1086510 *next week after Loss 3/4

>>1086363 Cute! >spoiler

I'm thinking of making a character LoRA, it's an OC from some guy on discord, does the style of the images of the dataset need to be consistent? Do all images have to be the same style?

>>1086573 >does the style of the images of the dataset need to be consistent? Do all images have to be the same style? No. In-fact for certain things its recommended they not be.

>>1086587 Realized I should expand this when not tired: When you train a LoRA, you're actually teaching the output patterns and stuff that's consistent between your selection, with some filtering and sorting by tags. If all of a LoRA's training data is of a single style, the result will adopt that style (especially for the character, but the BG too to some degree). For characters this is generally perfectly fine and often preferable if you're doing their native style (indeed, a lot of animated characters are totally unrecognizable outside their native style) and lessen the effect of any style LoRAs applied (essentially it's wearing layers) and especially styles just from tags. This was a bigger issue in SD1.5, but some styles are still pretty extreme and make it much harder/impossible to combine with other stuff. Many of these can be isolated by hitching them to style tags the base model knows ("3D" is a big one. If you make a LoRA using primarily 3D training data and don't tell it that training data is 3D, everything you make with it will be a weird pseudo-3D, but if you say its 3D it works fine. Same for "realistic, photo (medium)", "chibi", "sketch", "pixel art", and a few others). It will really be an issue if you try teaching clothing, poses, etc. using a single source (properly tagged "realistic" aside) because your result will be as much a (bad) style LoRA as it will be one for clothes (etc.). (One caveat is that if a character has warped proportions (e.g., very big head) as part of their style, those tend to carry over and look strange in styles incompatible with such. Pic related)

https://civitai.com/models/1405786/il-loss?modelVersionId=1621579 Part 3 of >>1084560

Lady Maria, the Hunter, the Doll and her squid friend showing their mad skills with some insane stunts and crazy tricks.

>>1086782 >FromSoftware presents BloodBoard

>>1084560 Really like that Dagoth one

>>1085383 Do the needful and post the anus sirs.

>>1083656 >>1084033 >>1083907 >>1084176 >>1084758 >>1084778 >>1085383 I wish there had been a way to save him and his purple selfussy.

What model and lora is this Cirno?

>>1086782 Lady Maria, the Hunter, the Doll and her Hunter*

>>1086870 Better Days [checkpoint]: https://civitai.com/models/623891?modelVersionId=1143094 VAE: https://civitai.com/models/140686 IL_Neg [embedding] https://civitai.com/models/1304409?modelVersionId=1133058 Couldn't find all the information since metadata was wiped, but this should be close. >Positive Prompt: dynamic shot, deep contrast, soft volumetric fog, subtle particles, rich color tones, detailed eyes, large pupils, detailed pupils, thick eyelashes, shiny skin, detailed background, big eyes, parted lips, simple teeth, from below side, looking down at viewer, foreshortening, flat color, anime_coloring, anime_source, (artist:rariatto_\(ganguri\)), 1girl, solo, (cirno, touhou), short, skinny, pale skin, cyan eyes, cyan hair, short hair, dark blue hairbow, large bow, baggy white short bloomers, white shirt, blue dress, barefoot, toes, toenails, indoors, kitchen, inside refrigerator, legs dangling, open refrigerator, food in refrigerator, juice bottle, milk carton, <Negative Prompt: IL_Neg, body paint, watermark, patreon watermark, signature, heart pupils, three fingers, four fingers, six fingers, three toes, four toes, six toes, (small head), (1other, 2girls, group, multiple_girls), depth of field, out of focus, blurry, shoes, teen, adult, tall, dark skin, Important tag for the style is (artist:rariatto_\(ganguri\)) which works on other checkpoints. Here's the actual artist on danbooru for comparison: https://danbooru.donmai.us/posts?tags=rariatto_(ganguri) No LoRAs were used. Model knows this character & style without using a LoRA. Here's two other Cirnos, same model, but they're a mix of rariatto & gerph artist tags. Something like this: (chibi:0.5), thick outlines, (artist:rariatto_\(ganguri\), artist:gerph:0.4), big eyes, round eyes, simplistic, (flat color), hard shading, toon \(style\), Again I cannot recover the exact information since metadata was wiped, but I take a lot of notes.

>>1086803 Dagoth Ur is combines well with lots of stuff

Realistic Filter [LoRA]: https://civitai.com/models/1003513 Feelin Boujie [LoRA]: https://civitai.com/models/1326743 CFG Scale Boost [LoRA]: https://civitai.com/models/99619?modelVersionId=947620 <lora:Feelin_Boujie_Illu_Edition:0.8>, fb-illu, <lora:realistic filter [IL]:0.5>, <lora:cfg_scale_boost:0.7>,

me after AI roleplaying sex with wild animals (for seventeen hours)

Part 4. https://civitai.com/models/1405786/il-loss?modelVersionId=1624934 Previous parts >>1084851 >>1084560 >>1086780 Don't know why it's not capturing the pose. It did fine in the epoch pics during training. Should be a break from totally new LoRAs for a week as next 7+ days will be Pony versions of recent creations, but I've already got data for 8+ LoRAs assembled that I just need to clean (most of which already have that done), tag and train.

>>1086888 (checked) Thank You very much! Have an egg as a treat.

>>1087065 Noice. >>1087051 Looks alright, so it wasn't a total loss. I dislike how the cards turned out here, but I got too tired to keep working on it & this is my second attempt trying to gen this idea so I pushed it through. 0.6 weight https://civitai.com/models/1228761/il-vivian-james-afterlife-empire 0.9 weight https://civitai.com/models/1405140/ghibli-midjourney-style-illustriousxl 0.9 weight https://civitai.com/models/1343806/kagenashi-minimal-shading-illustrious 0.3 weight https://civitai.com/models/1299949/cloud-collection-background-sdxl-illustriousxl

>>1086732 This is the character I want to make, there's some material, this is her "default" attire, but she has also been drawn on different outfits and also different styles, what should I do? She definitely has that big head problem you mentioned

>>1087083 Just another question, Although I did use the correct prompts listed, sometimes I get esoteric looking cirnos and other times I get pictures that barely have the artstyle on point. How do I fix?

>>1087103 There's about a dozen different reasons that could be it. Check CFG, check denoiser, (euler a is my go-to). Comfyui also has differences between A111 and similar because A111 has a set of default 'randomization' that breaks parity. But it shouldn't make shit like that pic. Could be graphics card. My shit makes everything over-saturated so I have to do shit to lower saturation. Sasunga AMD (or something. Found it out when I duplicated another dude's setup over IRC. I can only assume its AMD because so much else AI AMD shit is broke.)

>>1087089 qt rat.

>>1087103 Yeah, I dunno how Better Days is outputting that. Looks a bit overcooked. Other settings I've used: Euler A Align My Steps 35 Steps Hires Upscale 1.5X at 15 Steps & 0.3-0.4 Denoise Clip Skip 2 Discard next to last sigma is disabled (thought this setting was called Discard penultimate sigma but I dunno) 1344 x 768 resolution (I usually crop & re-scale later to get 16:9) I have the --disable-safe-unpickle flag in my launch settings since some embeddings don't work without it, but that setting can allow the execution of malicious code or whatever so don't use anything that isn't vetted I guess. I may have been using DPM++ 2M Karras or something as well. It's difficult to recall. >>1087112 Well for starters, I don't use Comfy. I'm using reForge which has similarities to A111, yet is significantly faster. There is now another version of Better Days with a Baked VAE. I'm not using the Baked VAE version. The author created the Baked VAE version because people were getting results that were de-saturated & lacking contrast by generating without a VAE. So maybe try the Baked VAE version of Better Days? I dunno.

>>1087128 why did you make her a rat?

>>1087103 Are you using the recommended resolutions? Share your settings in case it's something obvious .

>>1087131 So, the problem was that I was mostly using image2image but normally using normal image generation made it better. Also the steps improved the quality a lot. Thanks dood.

>>1087386 That looks closer to what I was getting. I also gen'd this Cirno back in February using the Janku v2.0 checkpoint, I think: https://civitai.com/models/1277670?modelVersionId=1469829 There was a v2.1 update & now there's a v3.0 apparently. I may try Janku v3.0 after it leaves early access. For a while, I tried Koki'o Illu v2.0 [checkpoint]: https://civitai.com/models/1135145/kokio-illu But recently, I went back to Better Days checkpoint since I still saw really neat images being created using it.

>>1086331 Was there anything done with this model? I recall some AI art done by Coco himself during the early days of AI, but not much else.

Is there a name for this character archetype?

>>1087592 Yankee

>>1087194 At first I thought she had a mouse nose. So, naturally she grew a huge white cock.

>>1087691 She is supposed to be a cloud

>>1087714 Now make her naked.

>>1087714 My mistake.