For those don't want to read: This is a distilled model, you need to lower CFG scale to 1~1.5.
RDBT [Anima]
See the main page for more info. (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.
This is a fully finetuned model, but in LoRA. It needs the correct base model to function properly. Don't mismatch base models. If you don't know how LoRA works or what this means, use the ckpt version.
Links to the LoRA base models:
AnimaYume (ym) (hf link) (civitai link) (int8 version).
Anima pretrained (b1,p3,p2,p1) (hf link)
FYI: "Merged" base model usually is NOT recommended. It's weights might be very messy. Probably merged different pretrained versions. A nightmare for all LoRAs, not just this LoRA.
Description
FAQ
Comments (19)
I just wanted to thank you for the work you put into these. 1.0 does great work. This actually works pretty cool as a bit of a aesthetic detail booster for a high res pass even if not used in the first pass, if you're willing to give it a bit more control (0.4+ denoise)
After some testing, I have to say I find v0.39 to be generally better than 1.0 at standard generation settings (so non-1.0 CFG and 25+ steps). A non-step distilled version of 1.0 would be really nice.
Same. It matches my experience with other turbo loras and checkpoints that the details fall off at higher CFG compared to non-turboed versions. However, I really like the proportions that come in with v1.0 so I'm torn. 😭
I'm guessing your images are oversaturated with high cfg? Typical things for distilled model, as the model is distilled with cfg 4 already
Try RenormCFG node. This can fix the latent overflow.
I prefer renorm_cfg 1.05.
@reakaakasky sry to bother and for the stupid question but when i used the node it gave me an error, i threw that to chatgpt and it says only works with lumina architecture. i used the default comfy node. also i used the node right before ksampler.
Edit: Tested in default workflow it works fine. For some reason doesnt work in my workflow.
Edit2: A lot of troubleshooting later, finally found the culprit. Batch size >1 will break it.
Is 1.0 for yume planned?
No, ym1f v0.43 literally is v1.0 (identical training settings and dataset)
v1.0 didn't update much, just version number, and re-trained on original anima base v1.0.
Do I understand correctly that I should leave the regular CFG at 1 and set a higher value in the RenormCFG node? Or should I put 1 in RenormCFG and a higher number in the regular one?
When you need CFG (CFG value in the ksampler > 1), you can add the RenormCFG node before the ksamlper node.
I recommend set the "renorm_cfg" value in RenormCFG node to 1.05. It's a "suppressor" to prevent oversaturated image.
@reakaakasky thanks!
Boss, is there some additional distill in these 1.0 versions in relation to the previous 0.39v? which would work better with the new Anima Turbo released recently?
A little bit higher distillation strength. v0.39 has some stability issues.
I just tested rdbt v1 16step +new anima turbo. I'm surprised I got a 8step model without any issues, seems to be even better than my v1 8step version...
Do you think that you could release the style trigger words that were in a couple of versions as style loras?
No. Sounds great but in reality those triggers are not specific enough (one word mapped to many styles, and the model can't learn them well.
Most of the time, it would be better to just stack style LoRA.
I noticed the checkpoint version is up to 1.2, is the lora version going to see those updates? Having it in lora format is really handy and I am not nearly smart enough to try to extract them.
ComfyUI has a node to extract LoRA and is very simple. All you need is the rdbt model and the anima base v1.
@ikekph5 I have tried that, of course, which led me to asking this here. Loaded both diffusion models, attempted to subtract the base model from the rdbt model with modelmergesubtract and saved it using the extract lora node. The results are always incorrect, though, covered in thick latent noise. Tried swapping the models on subtract in case my understanding of the order was wrong, still bad. Are there docs anywhere that demonstrate this working on anima models? I've only ever gotten these nodes to function properly with older architectures. I dunno what I'm doing wrong. 🫠
ComfyUI built-in LoRA extractor nodes are puzzles. I still don't know how to use them after I check the code, there is no docs at all, even in code.
@sansmia Use the LoraExtractKJ node from comfyui-kjnodes.
@reakaakasky Thank you, that node is infinitely easier to use. I had actually tried to use it before but it crashed comfyui. I just realized it was that dynamic vram optimization comfyui added a while ago, temporarily disabling it let the node complete and now I have my lora. Cheers. :)