What sampler and scheduler you recommend?

Anima is very inflexible with samplers and schedulers if you dont pick the right one it can mess the whole thing. it took a while for me to find my favorite sampler/scheduler combo.

@Neon_signs seconding this, Volnovik is quite knowledgeable

@Neon_signs Do not have time to do a full test. Euler a beta/simple is a solid choice for both t2i and i2i, but I do not like it, adds to much random stuff. ER SDE Beta is really good at details and composition, but rather messy at higher resoulutions and img2img. My current setup is ER SDE t2i, then Euker A inpaint and upscale x2 tiled. Basically cleans up mess. https://civitai.red/images/131722318 This one is made that way. ER SDE upscale transforms noise into grid patterns with this prompt, they look denoised, but are exceptionally messy.

@Neon_signs Ok, another sleepless night, but I tested. Ones that seems working are Euler A and dpm++ 2s a RF, simple and flux2 schedule types. Both t2i and i2i. But I'll probably still go Er SDE flux2 and euler a for i2i for now.

@Volnovik its flux2 scheduler native or is there a comfyui node for it?

Euler A2 and Beta work well.

@HaloSkull it looked best at first. But upon further inspection it seems too "creative" for anima and with loras added on top it quickly starts falling apart.

@Volnovik I use it with style loras and it fixes a lot of bad hands at 30-50 steps. On base without the turbo lora.

I want Speed and Variety so this update fits in my alley... i dislike boring compositions.

It can do text - first attempt
https://civitai.red/posts/28926140

turbo lora is designed to enable cfg 1.0, because this skips the negative step for an instant 2x speedup, but this is good info too

@kai If I set cfg to 1.0 and turbo Lora strength to 1.0 it cut gen speed to 1/3 but breaks all other loras’ styles and compositions, literally unusable. That’s why I use turbo strength 0.7 cfg 4. That’s the sweet spot to keep full functionality of Loras with 2x speed up.

By the way turbo Lora as a distilled itself enable Anima to go beyond 1920x1080, I was able to gen up to 1.4x size at cfg 3 or 4 without seriously burning, 4K is still impossible though. But the fact that it can push beyond 2K means at 2K it actually gives more details than not using turbo. Image has a subtle fidelity upgrade.

Finally, using turbo will greatly alter Lora behaviour, it activates layers differently. As a result all Lora mixing strength receipe needs revision in order to get the same result before.

cfg 4 is too frying sometime. cfg 3 work better for me

cfg 3/4 and er_sde may cause limb errors, better use euler

@jack42 limb errors can be fixed with your own loras, and er_sde beta is necessary if you are doing intricate style. Euler A normal/simple is good enough for flat anime though.

The real problem for me is that turbo lora no matter how you tune it, gives very bad background compared to no turbo CFG 6, steps 30. All reflections on the windows of a highschool gone. I can only pull off perfect result for characters only. So I stop using turbo.

Still trying it myself - good suggestion, I find turbo good for testing, i set up lora strength a 0.87 (this is where for me changes images from 0.86 and 0.87 as a fine tune), 12 steps, euler, simple, cfg 1. your suggestion for me takes 10 seconds more and generates almost the same quality. hope this helps. Play with the fixed seed and change the strength of your turbo lora, when you find 2 images that are completely different at 0.1 change then that will give you the power to change composition quite a bit :)

You could try applying some negative weights of the anima rl lora. Might be the cause of it.

Now I can't unsee it....

You can use negpip to apply negative weights to completely get rid of that.

I can confirm after getting around to trying it that negpip like someone else recommended is the fix. Downloaded it via the comfy extension manager and added (sweaty:-1.2) to the prompt prefix. After that I like 0.2 more than 0.1, since it has a little more variety. If you want stricter control then maybe 0.1 is still better.

Explain to me what you think the difference between a "Turbo" LoRA and a "Lightning" LoRA is.
Unless you mean a literal Lightning LoRA that puts Lightning bolts everywhere in the image?

@scrunguscraig I mean Turbo is faster but like I'd like to see a Lightning LoRa for some quality changes if thats possible. And no im not talking about the Lightning bolt LoRas

@Zuutokari turbo lora reduse gen time for me from ~50 seconds to ~10s with my 4070, i doubt any faster results can be achieved without simplifying art-style

@aa110902 I'm not talking about the faster results I'm implying for better quality results in a sacrifice of slight time estimation

yeah I think turbo loras sort of 'neutralise' styles

also sometimes if speech bubbles with text are present - the tag "textless" helps

12 steps at double speed vs 30 steps, so over four times faster. Quality is subjective. I find it produces more coherent images with less issues, but it does push the style in a very "polished" direction as well as hurting variety. That can be improved by lowering the strength of the Lora.

Your best bet is to just try it and see.

Oh, I almost forgot. If you're using Comfy, install and use the Clip NegPip node (search for ComfyUI-ppm). It's invaluable IMO since negative prompts don't work with CFG 1.

@Silanda I use the base node "Zero Conditionning" linked to the positive one. Any advantage over this one?

idk, https://huggingface.co/dummy9996/anima-turbo-lora-extract/blob/main/anima_turbo_1-0.safetensors

@shoes22 thanks. Is it supposed to be that small? i guess ill test it out
Edit: its worse quality compared to the official Turbo v1.0