Cowabunga, dude! Dive headfirst into the nostalgia of the 1990s with this LoRA, meticulously crafted to capture the iconic and totally tubular style of the Teenage Mutant Ninja Turtles animated series! If you're looking to generate images bursting with the vibrant energy, dynamic poses, and slightly gritty charm of the classic TMNT cartoon, then this LoRA is for you.
This isn't your average LoRA – it's the result of a complex training process. We utilized high-quality cartoon scans, rigorous upscaling techniques, and trained multiple LoRAs. After careful selection and merging of the best performing models, we achieved a LoRA that truly captures the iconic style of the 90s TMNT animation.
Recommended samplers: dpmpp_2m/sgm_uniform/beta, heun/beta, deis/beta
Generation with 14-16 steps (with 8 steps Lora), Also look at examples of prompts in images and make your own variations based on them.
Please Support if you like this model: Buzz / Likes / Add post or Review / Follow - Thanks!
Description
5670 + 11340 steps
FAQ
Comments (5)
This looks terrific, I can't wait to try it, great job!
The results are pretty good stylistically, with a huge caveat: The trigger word is insufficient, especially if you are trying to show something other than the actual characters. You have to include the additional "The artistic style appears to be a classic cel animation style, reminiscent of the 1980s Teenage Mutant Ninja Turtles cartoon" language that you see in the prompts for the example images.
I sound like a broken record, but I am begging people to understand that when you include language like this in your training captions, you are actually making your LoRA worse not better. When you include something in the captions, you are kinda/sorta telling the model "This is NOT something you should learn as necessarily intrinsic to the LoRA alone, this is something you should produce if these words are included in a prompt."
Anything you repeat in most/all captions essentially becomes part of the necessary trigger.
Ok. I posted an example of image where none of this is present. You can not use triggers at all and choose the style. Specify the cartoon at the beginning and describe the scene, this will also work. Perhaps some of the examples will give more details, but no more.
Main triggers: TMNT, Teenage Mutant Ninja Turtles, and some names.
@PrimeDM It may take a few minutes but I have posted two sets of images to the LoRA page. I used your exact settings and did the five same sequential seeds. In addition to your settings being suboptimal for coherence and prompt adhesion, you can see that the results are less faithful to the style without that additional text that it appears you trained on.
The LoRA is still capable of producing good stuff, but because of how it was trained, it doesn't reach it's full potential until you add the additional words.
@RudyBaga thanks! I think it has more to do with the dataset and scene description, plus training is not on a full model. I have similar artifacts on different models, I don't always get the desired result right away. But as you can see even from your images, the style is preserved and at the same time you don't write whether it's a cartoon or a photo or a drawing. And apparently you don't use the 8 steps lora. Unfortunately, we had to make several versions of the loras to get a more or less stable result. Therefore, there were no manual edits to some of the models texts and some character names in the dataset were also omitted. So I agree that the first sentence in the descriptions could have been shortened.
Anyway, I think we can do it better in the future.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.