<pre><code>Arise!</code></pre>Main character from GameCube exclusive From Software (yes, that From Software) game Lost Kingdoms 2. It’s a fun game where you fight exclusively by summoning monsters you have limited control over from a deck of cards. Please post the pics you generate with this Head: f_s_t_g, high ponytail, short ponytail, blonde hair, green eyes, asymmetrical bangs, blue hair tie, Outfit: blue crop top, sleeveless crop top, black caplet, necklace, arm wrap, asymmetrical arms, navel, blue shorts, open shorts, black ankle boots (see version description for V1 tags) Negative: 3D, game screenshot Worth trying: Holding card, card, holding Trained on but don't expect it to work well: Runestone is likely unstable due to being a unique thing and limited TD.I had been curious for a while to see how good a LoRA could be made if 2D training data options was limited (though extant), but there was a 3D model that could be captured form a nigh infinite number of angles in a decent number of poses. I then saw dwnsty’s <a target="_blank" rel="ugc" href="https://civarchive.com/models/676075/holding-yu-gi-oh-card-or-concept-lora">Holding Yu-Gi-Oh Card | Concept Lora</a> and realized I knew a girl who would work perfectly for such an experiment and I actually liked enough to bother doing a LoRA for. This was trained off 6 official 2D art (loading screens, and some chibi promo comic. No cover art), 6 screenshots of the prerendered intro, 10 screenshots of gameplay, 7 screenshots of the model on a black background (I have the prequel webcomic, but it’s too blurry even for its res to be usable), and one screenshot of the runestone alone in a cutscene in hopes of improving its rendering (Total: 29 of character+1 of accessory alone)I’m kinda impressed how this turned out for a relatively complex character design: Even by epoch 2 she was recognizable, and epoch 3 got even more details I expected would be fairly tricky, and epoch 4 was near flawless (only runestone being missing). Main thing it struggles with are the laced open shorts and runestone. I can fully recommend training a Pony LoRA primarily off of clean screenshots of a model now (at least models of PS2/GCN era quality). Main issue is that 1: Whatever pose the model is in will become a default, so you either need to be able to pose the model, or just require specifying a pose in generation 2: It will really want to adopt the style of the few 2D images in the set, which it's not particularly well trained in (though the results aren't bad). This is easily fixed by loading a style LoRA.

Tara Grimface (Lost Kingdoms 2) / リズ (ルーン2)

First version Tags: Trigger: tara_grimface Head: green eyes, blonde hair, high ponytail, short ponytail, blue hair tie, hair tie, single earring, Outfit: necklace, crop top, blue crop top, sleeveless, bare shoulders, black capelet, red trim capelet, capelet, asymmetrical arms, arm wrap, nail polish, blue nail polish, midriff, belly button, navel, shorts, open shorts, belt, boots, black footwear

Tara_Lost_Kingdoms_2_リズ_ルーン2.safetensors

767500_training_data.zip

<pre><code>Arise!</code></pre>Main character from GameCube exclusive From Software (yes, that From Software) game Lost Kingdoms 2. It’s a fun game where you fight exclusively by summoning monsters you have limited control over from a deck of cards. Please post the pics you generate with this to the model page.Trigger: tara_grimface Head: green eyes, blonde hair, high ponytail, short ponytail, blue hair tie, hair tie, single earring, Outfit: necklace, crop top, blue crop top, sleeveless, bare shoulders, black capelet, red trim capelet, capelet, asymmetrical arms, arm wrap, nail polish, blue nail polish, midriff, belly button, navel, shorts, open shorts, belt, boots, black footwear Negative: 3D, game screenshot Worth trying: Holding card, card, holding Trained on but don't expect it to work well: Runestone is likely unstable due to being a unique thing and limited TD. I had been curious for a while to see how good a LoRA could be made if 2D training data options was limited (though extant), but there was a 3D model that could be captured form a nigh infinite number of angles in a decent number of poses. I then saw dwnsty’s <a href="https://civitai.com/models/676075/holding-yu-gi-oh-card-or-concept-lora" rel="ugc noopener noreferrer nofollow" target="_blank">Holding Yu-Gi-Oh Card | Concept Lora</a> and realized I knew a girl who would work perfectly for such an experiment and I actually liked enough to bother doing a LoRA for. This was trained off 6 official 2D art (loading screens, and some chibi promo comic. No cover art), 6 screenshots of the prerendered intro, 10 screenshots of gameplay, 7 screenshots of the model on a black background (I have the prequel webcomic, but it’s too blurry even for its res to be usable), and one screenshot of the runestone alone in a cutscene in hopes of improving its rendering (Total: 29 of character+1 of accessory alone)I’m kinda impressed how this turned out for a relatively complex character design: Even by epoch 2 she was recognizable, and epoch 3 got even more details I expected would be fairly tricky, and epoch 4 was near flawless (only runestone being missing). Main thing it struggles with are the laced open shorts and runestone. I can fully recommend training a Pony LoRA primarily off of clean screenshots of a model now (at least models of PS2/GCN era quality). Main issue is that 1: Whatever pose the model is in will become a default, so you either need to be able to pose the model, or just require specifying a pose in generation 2: It will really want to adopt the style of the few 2D images in the set, which it's not particularly well trained in (though the results aren't bad). This is easily fixed by loading a style LoRA.

Files with SHA256: 88b457c0a3698cd50d364479d0442621472ae47b8f037272d5242f6f1e78368c