<h3 id="summary">Summary</h3>V2Converted the clips from my LTX dataset to images and captions to h3 syntax. Changed the trigger word for style to : "2d anime style, Edgerunners style." So don't use same trigger as LTX. Character descriptions and character tags are the same as with LTX version though.1950 steps checkpoint is good for overall style, use the 2250 steps checkpoint in the optional downloads area for better characters from the show though it may have more character bleeding. You can compare with the grid samples uploaded. Let me know your thoughts what's better. 1500 steps may also be good, I'll test that more later if I have time.This version is image only dataset, so no animation style or voices are included, maybe in next version. I don't see any degredation in the motion and this was trained using Ostris Adapter v3 on Musubi Fork.All samples most likely from 1950 steps; check the grid samples for character likeness in later steps version. Also I think its just better to use a ref image if you want character details 100% correct. It's just an H3 thing at this point.V1Trained from 354 clips total from the first 7 episodes of the anime Cyberpunk Edgerunners.The trigger words are:ANIMSTY, SHW_CYB.And at the end here I will show how to prompt the characters. It can do the main cast but there is some character bleeding, I didn't really make the point to be to recreate the characters, but it can do that. If I had enough clips of each character it would've worked better, but that wasn't my main goal here. I think this works well as a base for a modern anime lora. I will do that next probably (10-20 shows to give diversity to the dataset).The training data has blood, shooting, guns etc. from the show, but LTX doesn't tend to play nice with those. It's trained on video only (no audio trained). Its trained for t2v, but may work with i2v too. Its trained on 832x480 resolution with a max 121 frames.Use my example workflow for best results at 50fps.<h3 id="training-ltx">Training LTX</h3>I loaded the savestate from my 90s anime lora at 53k steps. So the lora had knowledge of how to do anime. I wanted to see if it could keep that information plus train a completely different style of anime. At 9k steps, I already had a good looking result (something usually that takes like 30k steps). So if you're training an anime lora, give it a try. Load from save state and train your lora (rank 64) using my anime 90s savestate as a base.<a target="_blank" rel="ugc" href="https://huggingface.co/comfyuiman/anime90s_ltx/tree/main">https://huggingface.co/comfyuiman/anime90s_ltx/tree/main</a>It should train just set these two: set "NETWORK_DIM=64" set "NETWORK_ALPHA=64"otherwise you will get an error. I use musubi fork with ltx 2.3 support.<h3></h3><h3 id="character-prompting-h3-ltx">Character Prompting H3 / LTX</h3>Character Tag: char_david Primary Outfit: A bright yellow jacket with a glowing cyan collar over a black t-shirt and dark pants. A gold cross necklace hangs around his neck. Appearance: A young man with dark skin, dark spiky hair with shaved sides featuring geometric lines.Character Tag: char_lucy Primary Outfit: A cropped white jacket with blue lining over a black, high-necked sleeveless top with red accents, black shorts, and thigh-high black stockings. Appearance: A young woman with pale skin, bright red eyes, and short white hair featuring a pastel rainbow gradient on the inner layers.Character Tag: char_maine Primary Outfit: A dark, unzipped vest over his bare chest, a thick red collar that emits a bright orange and yellow light, and dark pants. Appearance: A large, muscular man with dark skin, short blonde hair, massive cybernetic arms, and a metallic cybernetic lower jaw plate.Character Tag: char_rebecca Primary Outfit: An oversized dark jacket with neon green circular details and stripes, a dark top, and mismatched stockings. Appearance: A girl with pale skin, bright neon green hair styled in twin pigtails, and glowing red cybernetic eyes.Character Tag: char_kiwi Primary Outfit: A dark red jacket with a high collar over a dark brown garment, and a red mechanical face mask covering her nose, mouth, and jaw with horizontal slits. Appearance: A woman with pale skin, short blonde hair cut with straight bangs, and yellow sclera with intricate, glowing red cybernetic irises.Character Tag: char_dorio Primary Outfit: A brown jacket with a red and black collar over a purple shirt, and dark pants. Appearance: A muscular woman with short, light blonde hair shaved on the sides, blue eyes, and prominent cyan cybernetic lines tracing her cheekbone and jaw.Character Tag: char_pilar Primary Outfit: A dark, high-collared, sleeveless top. Appearance: A tall, lanky man with messy dark hair, a metallic jaw, long, metallic cybernetic arms, and a yellow cybernetic visor over his eyes.Character Tag: char_julio Primary Outfit: A dark vest over a light teal shirt with a white, frayed collar, and dark arm coverings. Appearance: A man with spiky brown hair and a black headset over his left ear.Character Tag: char_gloria Primary Outfit: A bright yellow jacket with a high collar, revealing a blue and white garment underneath. Appearance: A woman with dark red hair tied back, green eyes, and a small rectangular red mark on her right cheek.Character Tag: char_katsuo Primary Outfit: A dark suit jacket, a white collared shirt, and a dark tie. Appearance: A young man with light blue hair styled in a bowl cut with longer sections on the sides.Character Tag: char_faraday Primary Outfit: A dark maroon suit jacket, a light blue collared shirt, and a dark tie. Appearance: A man with short grey hair, a prominent jawline, and a cybernetic implant with three vertically stacked cybernetic eyes on the left side of his face.Character Tag: char_doc Primary Outfit: A light-colored sleeveless top and dark pants. Appearance: An older man with a pinkish skin tone and a bald head. He wears thick, translucent green, rectangular glasses over glowing yellow eyes that feature a distinct, circular mechanical iris design.

Cyberpunk Edgerunners Style Lora - H3 / LTX 2.3

Converted the clips from my LTX dataset to images and captions to h3 syntax. Changed the trigger word for style to : "2d anime style, Edgerunners style." So don't use same trigger as LTX. Character descriptions and character tags are the same as with LTX version though.1950 steps checkpoint is good for overall style, use the 2250 steps checkpoint in the optional downloads area for better characters from the show though it may have more character bleeding. You can compare with the grid samples uploaded. Let me know your thoughts what's better. 1500 steps may also be good, I'll test that more later if I have time.This version is image only dataset, so no animation style or voices are included, maybe in next version. I don't see any degredation in the motion and this was trained using Ostris Adapter v3 on Musubi Fork.

Files with SHA256: ab18826e69e586a6af362ed3878dfd7525e8c3249cd237d1076ca489449d07cb