The is a concatenation of the original DMD2 Speed LoRA FP16 accelerator with six LoRAs that I've trained. It serves the same purpose, (create a great looking image in few steps) but the output is quite different from the original. Not saying it's better or worse, just different, and I think some of you may like it.
Use LCM sampler (any scheduler will do), and CFG from 1 to 1.8.
Works on SDXL, Pony and Illustrious models.
Description
v4 (previous versions will remain unreleased)
FAQ
Comments (6)
Can you be more specific in what those 6 loras are about? it could help people to decide to try this out.
LoRAs that enhance contrast, improve (in my opinion) composition, and randomize faces. But instead of me describing, it's much easier to look at comparison images I've uploaded or just download and try, since it cannot be used in onsite generation anyway, as CivitAi doesn't support LCM sampler.
@billem ok thanks for the explanation. Good work! have a nice day please.
I feel like it's probably overall worse, and you can see it in the example. People might get confused by there being stuff like more boob or a little extra detail here and there, but those differences can also easily change with different seeds. But what I noticed overall is details tend to be less sharp, with a bit of dysfunction in scene layouts. For example, in the street pic, it has way too many lights, and doesn't make sense when it comes to reality. The problem with baking in LORA, specifically in DMD2, is that those LORA will then become what's prevalent, removing the diversity from DMD2 and making its adherence worse overall, and more glitchy because of it leaning more heavily on the added LORA's, which would be limited themselves. At that point, using the LORA's separate from the DMD2 would make more sense. And with many checkpoints having built in LORA, conflicts can easily arise there too, where the output is negatively affected due to the sheer number of LORA being utilized at a single time. By baking it into DMD2, the problem itself becomes inherent. I would like to see a successor to DMD2 though, just not sure what would be needed to make it better. DMD2 on Flux or even just old SD 1.5 would be interesting too.
I'm not really sure I understand what you are tying to say. Yes, different seeds can alter the image, we all know that, and that's why I added comparison images with the same seed. Also, this is not a merge, or bake as you put it, it's a concatenation of LoRAs, so the original DMD2 LoRA remains intact, with additional guidance on top towards the style that I like - that is why I did not explicitly state that this is better or worse than original DMD2 LoRA - just different. You may think it's worse, I will disagree. It's all subjective. And obviously this Lora is not to be used with a base model that already has a DMD2 LoRA merged in. Flux has different accelerating methods, and SD1.5 should have been long forgotten by now.
Works very well with 8 steps. Also with Loras characters, What a speed generation ! Thx