Juggernaut Moss DMD2 Lora

作者：KandooAI

A huge thank you to RunDiffusion for sponsoring the compute for Juggernaut Moss. Without them none of this would exist! I'm currently looking for a job! Full-time or long-term contract work in model training, finetuning and image generation. If you like what I do, reach out on LinkedIn or send me a mail at KandooAI@web.de :) Juggernaut Moss DMD2 LoRA on HuggingFace Hello everyone, here comes the Turbo for Juggernaut Moss! :) This is my own DMD2 LoRA, made specifically for Juggernaut Moss. I didn't want to just grab one of the existing speed LoRAs, I wanted one that really fits Juggernaut, so I trained it myself. With the LoRA you only need 6 steps at CFG 1 instead of 30 steps at CFG 3, so you get your images about five times faster, and they still look like Juggernaut Moss. Not 100 % the same, the Turbo has a bit of its own character, but very close. Please keep in mind: this is a Beta. It was my first try and honestly it already turned out very usable, but it's not perfect yet. Hands and anatomy can go wrong here and there (an extra finger or a weird arm in busy scenes), different seeds give you more similar images than the full model, and it likes to play it safe with centered compositions. I'm already working on the next version, which will be trained on BF16 with an extended dataset. If you find weak spots, let me know in the comments, all of that goes straight into the next version. The LoRA was made for Juggernaut Moss and that's where it looks best, but it also works on plain MOSS-Image Base. If you don't want to mess around with a LoRA at all, there's also Juggernaut Moss Turbo, a checkpoint with the LoRA already baked in. You'll find it as its own version on the Juggernaut Moss page.

How to use it in ComfyUI Put the LoRA into ComfyUI/models/loras/ and add a "LoraLoaderModelOnly" node between "Load Diffusion Model" (with Juggernaut Moss) and the sampler. That's it. There's also a ready-made workflow on HuggingFace that you can just drag into ComfyUI. Recommended Settings: LoRA strength: 1.0 (0.9 works too) Steps: 6 (use 7 for complex scenes with several people or hands doing something) CFG: 1 Sampler: euler_ancestral, Scheduler: simple (dpmpp_sde with beta is just as good in my blind tests, but twice as slow) Avoid: karras (blurry) and dpmpp_2m with normal or beta (burnt) Negative prompt: leave it empty, it doesn't do anything at CFG 1 Res: around 2 MP, e.g. 1344*1792, 1792*1344 or 1536*1536 Juggernaut Moss version: int8 is the recommendation for this Beta

Prompting Same as with Juggernaut Moss: describe the image like a photo you're explaining to someone who can't see it. Who or what, what's happening and where, the light and the camera. Short prompts work, but detailed ones work much better. One thing is even more important with the Turbo: write every word you want to see in the image in quotes, and if there's a spot where text would naturally show up (a sign, a label, a shirt, a book cover), give it some text too. Otherwise you'll get gibberish there. Just writing "no other text" doesn't work.

And as always: Apache 2.0, just like Juggernaut Moss and MOSS-Image. Do whatever you want with it, a simple shoutout is all I ask. :) Have fun with it and I'm looking forward to your pictures in the Gallery! :) If you like the LoRA, please leave a Like. In the end, that's what helps me the most as a creator on CivitAI.