Edit:10/22 : Trained a qwen version check release notes for detailsPlease consider to donate or sub on my <a target="_blank" rel="ugc" href="https://ko-fi.com/tazmanner">ko-fi here</a>(all funds go right back into making more loras)This is a character lora used to generate the character Raven from the TV show Teen Titans (2003) and is intended to be specifically in the style of the show. I will create another lora sometime in the future which is more flexible for just Raven without style included. Its trained on the T2V model so it will also work for I2V as well.Necessary Trigger words: TeenTitans Style. RAVTTRecommended Strength: 1.0(See below for more info on prompting)I don't use "RAVEN" as the trigger because it generates birds...I trained this on 100 epochs, and #20 I found the best result, but try a few others and compare yourself:Difference between epochs:-Epoch 20: Recommended as it is has more emotion and expression but slightly off style-Epoch 37: While not bad, I think there is a bit stiffness. Style is matched fine.-Epoch63: This has very stiff faces but really good fit for the style of the show.<h2 id="dataset">Dataset</h2>An even dataset of 87 images of Raven with screen captures from the 2003 show. There is 5 or so screenshots from on model fan creations which are NSFW as well.Resolution 512x512<h2 id="main-trigger-words">Main Trigger Words</h2>Necessary Trigger words: TeenTitans Style. RAVTTOptional Trigger words for style reinforcement: animated, anime (but maybe not needed)I usually chuck "TeenTitans Style." at the start of the prompt and refer to her as "RAVTT" through the rest of the prompt.Use "animated" or "anime" to trigger animated style, and use "live action" to get real life versions, though the training data doesn't have any cosplay photos so it might not work, I will release another lora in the future for this functionality.To get her outfit use this phrase in the prompt: dark navy long-sleeve leotard, purple hooded cloak, red gem belt, gold clasp at collar, black ankle bootsYou can also specify "hood up" or "hood down" and might be good to mention with hood up "her face partially obscured by her hood's shadow" to get the mouth and eyes only showing. Her cloak can be changed to various colors like in the show (pink, grey, yellow etc.)To get her appearance use this phrase as well in the prompt: short angular dark purple hair, red oval gem on forehead, pale lavender skin,The gem on forehead and skin color often need to be prompted.Bunny Raven can also be prompted but I havent tested it:For funsies I added some images and captioned Raven as a bunny.Replace RAVTT with BunnyRAVTT trigger, maybe add this description for good measure:A white rabbit form with black muzzle and visible buck teeth. She has long white rabbit ears with pink interiors. She wears her dark blue hooded cloak, the hood is up.<h2 id="training-info">Training Info</h2>Trained on runpod on I think an L40S for like 3 days straight.Default settings except:LR 2e-5, Repeats 5Epoch 20: Steps 4500Epoch 37: Steps 8500Epoch 60: Steps 13,500(Epoch 100 had 23,000 steps FYI...)Attached is all the captions and an example workflow under "training data"Also all epochs can be found <a target="_blank" rel="ugc" href="https://huggingface.co/datasets/comfyuiman/RavenTeenTitansLora/tree/main/data">here</a> along with training dataI took screenshots using VLC from the show and cropped and resized all images using Birme website. Then I fed the images in batches of 5 to google's gemini 2.5 pro (it's amazing at captioning images). I used <a target="_blank" rel="ugc" href="https://civarchive.com/models/1615071/civitai.com/user/seruva19">seruva19</a>'s prompt as a base and used that with gemini to get all prompts done except NSFW ones which I either doctored the photo to make it sfw so gemini would caption it or I just did it by hand. I wanted the captions to be detailed so that you could have flexibility in changing her outfit, design etc. but I wanted to learn a little bit about style and also what happens when you overtrain. So I kept the dataset to only screenshots from the show for the most part and I tried to separate the trigger words. So if you don't prompt "RAVTT" and describe a random person it can make that character and the environment in the style of the show. But I learned it has its limitations since all women tend to have raven's style too. SO you should either caption all the characters or just describe them generically if going for style. And separate lora for character. All the other titan maincast is in the training data but captioned generically so they can be prompted but I think it wont be great. I did some of starfire but it looks not right... sometimes green animals look like beastboy as that animal.Example caption TeenTitans Style. RAVTT. RAVTT stands looking to her right with her arms to her sides in a dark room. RAVTT, with pale lavender skin, short dark purple angular hair, and a red oval gem on her forehead, wears her dark blue hooded cloak; the hood is down, resting on her shoulders, fastened with a gold and red circular clasp. She wears her black long-sleeved leotard, a gold-colored belt with three red circular gems, and dark blue cuffs on her wrists, each with a gold circle and red gem. Her pale lavender legs are visible. Her facial expression is neutral. Her dark blue cloak billows slightly to her left. The setting is a dark room with indistinct, dark blue and grey angular wall structures in the background. A portion of a dark, metallic, geometric structure is visible behind her on the left. The Camera Composition and Perspective is a medium shot, at eye-level.<h2 id="why-did-i-make-this">Why did I make this?</h2>This was more of a learning experiment. I also like the character and the style of the show, and I wanted to try an animated character lora (this one was made before my Bowsette one). I also wanted to see if I could do style + character in one lora. And I trained it up to 100 epochs (22K+ steps) to see if I could understand at what point it was overtrained or not and in what ways. I learned that it kept the style really well but her facial expressions were frozen. Really early epochs like 10 or so had a hard time getting her animated vs live action, so it was cool to see the style part of the training take over. I still think it requires a prompting of "animated" to get things where you want. Some positive take homes were that you really don't need that many images if they're captioned well for it to learn style, and you don't need videos if motion isn't a big part of the style.<h2 id="big-thanks">Big Thanks</h2>As always <a target="_blank" rel="ugc" href="https://civarchive.com/models/1615071/civitai.com/user/seruva19">seruva19 </a>Ghibli and Red Line lora post along with training data have been a constant inspiration and source of knowledge for me.Banodoco discord for always answering my questions on trainingKijai for his amazing nodes and advise on using them.

Raven Teen Titans (2003) Lora Qwen / Wan 2.1 14B T2v / I2V

-Trained on the same dataset as Wan2.1 (though slightly changed the trigger to "TeenTitansStyle." while keeping "RAVTT" trigger.-Trained to 18K steps, I kinda messed up and trained on the wrong dataset but it fixed itself after I updated to the correct dataset.-This is just kind of a learning experiment, the triggers and prompting are the same as wan 2.1 version. This lora though will ONLY give you Raven.-I want to take this learning and use it to make a real Teen Titan's Style lora, with potentially all the characters working properly it in it. If you don't mind raven being the characters, this lora also can do the style sort of.

Files with SHA256: ef73cb2b2f73a39fd532fd51c6c5ec3e4a5511e6b04b31def630af5885fde752