Turbo Lora
So I'm glad circlestone-labs/Anima have released an official turbo Lora and I'm just wondering distills both steps and CFG like other DMD2 LORAs.
These are some of the great distilled turbo LORAs/checkpoint I have been using in the past with previous preview versions and preview 3.
https://civitai.com/models/2466415/cosmos-predict25-2b-base-distilled-extracted-dmd2-lora
https://civitai.com/models/2356447/rdbt-or-anima
The cosmos predict 2.5 distilled Lora is interesting. This is from the creator and how they made it. https://arca.live/b/aiart/164898297
Both of these work really well and even works with different style LORAs. I'm yet to test the official one, but hopefully it works well as well.
Also preview 3 has been really good, especially at 1024 x 1024 resolution. As I've noticed the quality has drastically improved, plus it's given a lot more different results regarding creativity which is great. It's definitely progressed a lot since the first preview version. So well done to the team as always.
I hadn’t seen the Cosmos Distill LoRA, that’s a super cool idea. Definitely going to try that, it looks awesome.
The RDBT Anima model is great overall, very stable and works really well. The downside is that everything gets heavily flattened. You lose a lot of detail and texture, and styles that don’t align with the base aesthetic get almost erased. They’re still kind of there, but more as a vague “vibe” than something strong or distinct.
What I’m hoping for is something closer to how Flux 2 Klein 9B is structured, where you have a Base model for training, and then a non-base variant that’s faster and higher quality. Z-Image Turbo is another really strong example of distillation done right. It’s insanely effective, though it does make it harder to steer away from photorealism, so there are definitely tradeoffs.
I think this is 100% worth exploring. I’d assume the Anima authors are already aware of this direction and are probably considering something similar given where model development is heading. They recently started adding “-Base” to their releases, which makes me wonder if there’s an internal branch experimenting with aesthetic alignment or some form of distillation.
It’s probably too early to push for that right now, but once the base model is fully trained, and if there’s budget for it, I’d definitely love to see it (Authors are on a strict plan and budget from their Comfy-org "deal"). RDBT already made Anima way more approachable for my friend, who prefers SDXL-based models mainly for speed.
I hadn’t seen the Cosmos Distill LoRA, that’s a super cool idea. Definitely going to try that, it looks awesome.
The RDBT Anima model is great overall, very stable and works really well. The downside is that everything gets heavily flattened. You lose a lot of detail and texture, and styles that don’t align with the base aesthetic get almost erased. They’re still kind of there, but more as a vague “vibe” than something strong or distinct.
What I’m hoping for is something closer to how Flux 2 Klein 9B is structured, where you have a Base model for training, and then a non-base variant that’s faster and higher quality. Z-Image Turbo is another really strong example of distillation done right. It’s insanely effective, though it does make it harder to steer away from photorealism, so there are definitely tradeoffs.
I think this is 100% worth exploring. I’d assume the Anima authors are already aware of this direction and are probably considering something similar given where model development is heading. They recently started adding “-Base” to their releases, which makes me wonder if there’s an internal branch experimenting with aesthetic alignment or some form of distillation.
It’s probably too early to push for that right now, but once the base model is fully trained, and if there’s budget for it, I’d definitely love to see it (Authors are on a strict plan and budget from their Comfy-org "deal"). RDBT already made Anima way more approachable for my friend, who prefers SDXL-based models mainly for speed.
I think when Z image turbo was made alot of the open source image modells Kind of copy that as having a distilled version of the base model is really good as it makes quality prompted Adherence and stability much better. You do lose out on some of the variety but it's worth it for quality in my opinion.
I think when Z image turbo was made alot of the open source image modells Kind of copy that
Distillation existed at least as far back as SDXL, I think even 1.5 if not before. ZIT is not special, at all.