https://huggingface.co/nvidia/Cosmos-Predict2.5-2B/tree/main/base/distilled - https://huggingface.co/nvidia/Cosmos-Predict2.5-2B/tree/main/base/pre-trained = this LoRA
Description
rank64
FAQ
Comments (20)
this is a gem! Makes Anima preview 2 actually usable and works perfectly with my style LoRA.
BTW are you hosting this also somewhere on HF?
https://huggingface.co/hanzogak/Anima-Comradeship/blob/main/LoRA/Cosmos-Predict2.5-2B-base-distilled-LoRA.safetensors
I uploaded it to HF, but it is not a single repository. For more detailed information about this LoRA, please refer to https://arca.live/b/aiart/164898297 (Korean).
god so nice it is
this is the best thing since sliced bread. Actually makes the model very enjoyable to use.
yo my generation speed in anima is significantly slower than illustrious is it the same for you.
for those wondering: yes, this works perfectly well with Anima Preview 3 as well.
for those wondering: yes, this works perfectly well with Anima Preview 3 as well, even 4 steps!
Not a single word explaining what this Lora is about?
its in the name ? DMD2 = Turbo Lora?
For those wondering, is a turbo tool you can use 4 steps and get the a comparable result of 30 steps
Thanks, these posts are always so cryptic!!!
Gonna try this with Anima base-v1.0. Hopefully it will work 🤞
can anyone share the settings? such as steps, cfg, sampler, scheduler?
To use it, simply set the CFG to 1.0 and the Step to a value between 4 and 12.
It also works fine with Anima 1.0.
Anima Turbo LoRA is also available, so feel free to use whichever you prefer.
What is the best LoRA weight to use?
@FFrince Basically specify a weight of 1.0 when using them.
I’ve tried several speed-up Loras and checkpoints, but this model is still the best in the end, even on b 1.0. It would be great to have an update tailored for b1, but it still works well.
It’s truly amazing.
This is very useful, but there are some minor flaws. On the base model, compared to the official turbo v0.2, it has a greater impact on the artistic styles of other Loras, with slightly higher output monotonicity than the official turbo. However, this can sometimes be surprising. Looking forward to a better version of this Lora
I've tried many turbo LoRAs and checkpoints, but this LoRA is still the most compatible with other LoRAs even on B1, and it also delivers the best quality.
It would be even better if it were updated for B1, but thank you for sharing such an amazing model.
PRO TIP: This LoRa is amazing for HI-Res fix!!
It is extremely stable even at extreme resolutions. So even if you don't like how it alters your base output and prefer a different turbo LoRa please try switching to this one during the hi-res sampling. I can go as absurd as 2x with essentially zero artifacts.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.











