from experience I can tell you that the cosmos-dmd LoRA is good for mixing with the others


so the real User Preferred Speed Distillation setup is some bizarre mix of them all :' )

I love RDBT, but I hate how bad it is with styles. If the prompt is somewhat short like 30 tokens, it basically ignores the artist tag completely, and it seems even more prone to do so if you prompt something that is NSFW. But even with longer prompts it's still a major issue, aswell as things just looking very washed out.

Maybe it would look a lot more faithful to style if I went crazy like 1000 tokens or something, but I shouldn't have to just for the sake of getting the style I want. I have no idea why RDBT does this when style is basically the most important factor when genning.

It's still quite fun to use if you don't care about style and want something that looks more visually pleasing (other than washed out colors) faster with what appears to be a bit better quality.

I might use it every now and then for casual fun SFW memes or cute stuff. But the instability is just insane. I had a relatively simple prompt where it had "photorealistic" tag and it worked as expected, then I add in a couple NSFW tags and boom, SD1.5 style lol.

Now if only anima could do some basic furry concepts. I really hope they decide to include E621 into the training data, or at least make another version with it in.

There's so much damn potential with Anima, and to think we're only at a preview version, not a finished model or a finetune of the finished model.

Almost the same

CFG 1.1 will probably work fine and allow negative prompts. You could also try NAG or one of the other no cfg negative methods.

There's Normalized Attention Guidance, NegPiP (in comfyui-ppm), or making your own silly setup by running the turbo LoRA at a lower strength and compensating with https://civitai.com/models/2364703?modelVersionId=2684678
some portion of "anima_preview_rdbt_v0.6_cfg_distilled_only" on top

@VeerGeer Tried NegPip, 'Anima' object has no attribute 'text_processing_engine

@VeerGeer "Normalized Attention Guidance" don't work for Anima

@chaosleges sounds like the text encoder isn't being loaded in as expected

https://civitai.com/images/128368192
a workflow with it working looks like that, at least with the nodes from this repository;
https://github.com/pamparamm/ComfyUI-ppm

I'm using this one: https://civitai.red/models/2426853/anima-preview-workflow

Could you share your workflow ?

这个可以不变采样步数，提高CFG的同时降低lora权重，可以达到速度和画面的平衡，我目前用着3.5CFG，12步，0.4的权重，时间缩短到原来的一半以下，画面损失并不多

Same here, less details, less color, sometimes limbs are not complete but significantly faster gen speeds. Hands on hip, but no forearm just floating wrist and palm on hip. CFG1, 12Steps.

Had i started with turbo i would have been satisfied but after experiencing what Anima is capable of, I'm just too spoiled to be satisfied with turbo. Perhaps I'm doing something wrong, perhaps its just what it is.

The hands looking good is a result of high accuracy on anima, this basically gives anima a mini version that is trying to replicate anima in the lightest way possible, instead of using all the data in the big model. This makes it less complex, which means it takes less steps to be more accurate. This comes at the cost of the model needing to essentially listen to itself more than it listens to your prompt, and thus being less capable at creating more complex, abstract concepts that aren't common in the data the model was trained on.

@CocoLarge Ohhh... I... still don't know if I get it. So the smushing of the Turbo lora into fewer steps results in better consistency, at the cost of constrained creativity? So it's a choice between more complex generations that allow for errors, or more rudimentary ones that don't do anything that might result in errors?

@Derpression Not exactly. Because the turbo lora makes the model want to reach a final image quickly by simplifying data, rushing towards the training data and then asking itself if the final image looks like the training data. This method of asking itself if the image looks like the training data conflicts with cfg and can cause problems. Which is why the instructions for turbo/lightning/dmd2 ask for you to use cfg 1 (cfg 1 is equal to not having cfg turned on). The problem is, cfg helps the model follow your prompt more closely. Errors aren't really a part of this equation. Errors can either be a thing that just happens randomly, or the result of not having enough steps (aka low accuracy).

The simplifying of the data results in the model knowing less.

Yes unless you use NAG (Normalized Attention Guidance): https://github.com/BigStationW/ComfyUI-NAG-Extended/tree/main

@AnimaXx i guess the workflow that im currently using might be a challenge for me to edit the nodes, rip. . .

You can set it to 1.5, and it will still work and look largely the same (although, from what I understand, long negative prompts might not work reliably?). If some parts of your negatives still get ignored, pushing it up to 2.0 might help. Beyond that, you might start noticing some fried contrast.

If I understand correctly: the point of making a turbo lora that works on exactly CFG 1.0, is that it runs twice as fast as the negative attention step is skipped by comfyui (because the negative prompt doesnt do anything on cfg 1) @Derpression 

@kai That's true, although I feel like going down to 12ish steps alone already cuts down generation time enough to make the CFG difference negligible.

And I would also argue that this lora is worth using even if you aren't taking advantage of the turbo thing at all. I've been using it on regular slow 30-40 steps + CFG 2 generations, and it's been shockingly good at stabilising their outputs, reducing anatomy and composition errors.
Granted, I've learned that it comes with some constraints and limitations in regards to prompt adherence, but it's still been a net gain for my purposes.

I've been using Anima with 35 steps, CFG 4.5 before. With the turbo lora, I've been enjoying the massive speed up, but found that the results were less to my liking, especially without being able to use negative prompts, so I've tried around and found that turbo lora, 12 steps, CFG 2.0, works perfect for me. Takes twice as long as with CFG 1.0, but still a lot faster than with 35 steps before this lora came around.

During image generation, Condition, Uncondition, and CFG in each denoising step follow the formula: Uncondition + CFG * (Condition - Uncondition). When CFG=1, substituting into the formula shows that Uncondition is canceled out, making the Negative Prompt ineffective.

skil l isseu

Saying your images look like **** with no further clarification like the platform you are using, workflow, settings, etc is not going to get you a solution. You get back what you put in. Give us something to work with and maybe we can help you.

I had the same problem, even after copying the gallery's prompt, it still didn't come out as strong as the original image, which is weird...

weight 0.6, cfg 3, step,20