RDBT [Anima]
This page contains the "original LoRA files" for advanced users.
See the main page for more info. (ckpt version, usage, faq, int8 ckpt...)
See this page for update log and version info.
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Description
FAQ
Comments (12)
Just a suggestion when you do your next dmd2 lora but you might want to check this out: https://civitai.com/models/2466415/cosmos-predict25-2b-base-distilled-extracted-dmd2-lora
As it's a great DMD2 by polyx The creator also shared this link about how they created it:
https://arca.live/b/aiart/164898297
This is the Higginface link too: https://huggingface.co/hanzogak/Anima-Comradeship/tree/main/LoRA
As previously with preview 2. I have used both this Lora and your one together and it works really well.
that's a extracted lora from cosmos predicted 2.5. I was surprised that the lora actually worked without any problem on cosmos predicted 2.
According to their paper, they swapped te from old t5 to 17gib qwen2.5vl (17gib te + 4gib dit...) and changed dataset, captions, pre-train + post-train, balabala. Sounds like a very different model.
@reakaakasky So from my experience and testing, it does work pretty well on all of the preview versions. You want to use it around 1 to 0.7 strength and the Sampler of ICM or euler... You want to use around 8 steps and 1 CFG. The creator did say this however "since NVIDIA DMD 2 has a characteristic of producing faint results, it seems better to use it in combination with Anima-Preview2-Turbo" Lora. It does however work pretty well. Actually out the box even on its own but when you mix it with style and stability even you're DMD2 Lora It works really well as well and I've even checked with the LoRA Compatibility Analyzer node from this great custom node set https://github.com/ethanfel/ComfyUI-LoRA-Optimizer/tree/main
And it basically said it worked well together. I think your CFG distilled Lora works slightly better with it tho.
It is actually a really great LORA as it enhances anima and it doesn't really matter what preview version it is as it works well in all of them. This is the great creator of it https://huggingface.co/hanzogak
If you still want to use negative prompting with the speed boost from cfg=1, you can use the CLIP NegPip node from comfyui-ppm.
Do you have a workflow for that?
@brahianvalles https://civitai.com/posts/27850789 This was one with the left side image generated without any negative and the right side generated with (trees:-10). That weight might be a bit high but I was just kinda nuking stuff to test it out.
https://civitai.com/posts/27858159 Here's the same setup with trees only getting a -2 weight which seems to be around the sweet spot.
Version 0.24 is such a downgrade... Even version 0.23 actually performs better on the P3, even though it's not supposed to be compatible with it.
you mean "v0.24fd vs v0.23f dmd2" or "v0.24fd vs v0.23fd" ?
@reakaakasky v0.24fd vs v0.23fd
It's as if someone poured slop into the LoRA dataset (I know that's not the case), but the art looks more generic
if you use the old version, that problem isn't there, scenes become more detailed, angles more interesting, lighting more beautiful, and so on. Could it be that the "-40% steps" had such a negative impact on the result?
probably, I can't tell. p3 also has stronger styles. Looks very different than p2.
Dataset didn't change.
nvm, it's a bug in my code, v0.24fd will not fully denoise image. p3 speedrun failed.
dmd2 version is fine.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.