Speed Boost Flux, SDXL, LTVX or HunyuanVideo with TeaCache & WaveSpeed!

[@fluffsquirrel]

7:28 So few differences in fact that I can view it stereoscopically on my phone and I see no interference between either eye. It almost looks like a 3D image! Also these demonstrations are so cute, I love the subtle nod to your channel name when you test these models!

[@NerdyRodent]

;)

[@ZelpinAI]

Great tutorial. Thank you. ❤❤

[@michaelbayes802]

wow this is great. Just tested on SDXL (I know I'm old fashioned) if you do a text to image and then an image to image and put the wavespeed on the 2nd step there is virtually no difference in the image quality but a big improvement on the speed.

[@michail_777]

Thank you so much, these are incredibly useful nodes!

[@MilesBellas]

Faster generations = faster choices for later upscaling.😁

[@Oxes]

Gonna check this out in depth later at work🔥🔥

[@Quantum_Nebula]

What if you pass in the hyper-flux you mentioned at 6:32 before the wavespeed rather than after?

[@moviecartoonworld4459]

Do I need to install Sageattention and Triton?

[@mariopadilla1445]

@@moviecartoonworld4459 I’ve been having trouble with getting triton to be recognized. Wavespeed still works waster than without triton but if you can get triton working it will be slot faster

[@bause6182]

Only if you want to compile the model

[@NerdyRodent]

Only if you want to use Hunyuan. For sdxl, flux, etc you’re good!

[@moviecartoonworld4459]

@@bause6182 Thanks for letting us know!

[@moviecartoonworld4459]

@@NerdyRodent Thanks for letting us know!

[@pdjay8912]

좋은 정보 감사합니다. 그렇지 않아도 4070ti에서 Flux와 Hunyuan생성이 너무너무 느려서 힘들었어요. 잘 활용해 볼께요 :)😊

[@boricuapabaiartist]

I’m back to training anim diff motion loras for sd15, are sd15 models working with anim diff using speed wave to gain some speed not using lcm but rather with high step count?

[@NerdyRodent]

Ah, wasn’t aware it worked with sd1.5 too!

[@MyComyui]

it would b great

[@boricuapabaiartist]

@@NerdyRodentactually I haven’t tried to see if wave speed helps speed up sd15 or sdxl anim diff inferencing, thought you had tried it, but will try it out and if it’s not supported I’ll put in a request on the repo

[@michail_777]

@@boricuapabaiartist 🤣

[@K-A_Z_A-K_S_URALA]

share the node!)

[@MilesBellas]

In VFX....15 second renders = amazing ! Faster turnover reduced rendering. Using two machines = normal for VFX

[@ZelpinAI]

May I ask what GPU you used?

[@bgtubber]

I don't see WaveSpeed in the Custom Nodes Manager. Did they remove it?

[@岩田邦夫-v8k]

Can WaveSpeed, Hyper-FLUX LoRA and FLUX.1-Turbo-Alpha be combined to speed up upscaling processes ? Also, Can you batch upscale a large number of images in any folder ? How do you make the name of the original image carry over to the upscaled image and naming ?

[@NerdyRodent]

Yup - with the caveats shown in the video, of course!

[@ArrowKnow]

Awesome! Thanks

[@NerdyRodent]

👍🏼

[@hansolo9546]

could you tell me if you have your workflow uploaded somewhere thanku

[@KriGeta]

will this work with SDXL?

[@NerdyRodent]

You can Speed Boost Flux, SDXL, LTVX or HunyuanVideo - and now also SD3.5!

[@KriGeta]

@@NerdyRodent I have RTX 2060 8GB Vram only, can I use teacache or wavespeed?

[@NerdyRodent]

@@KriGeta yup!

[@KriGeta]

@@NerdyRodent Trying to run Teacache but there is no option for SDXL in model type

[@juanjesusligero391]

Oh, Nerdy Rodent! 🐭🎵 He really makes my day! ☀ Showing us AI, 🤖 in a really British way! ☕🎶

[@Snaaaaaaaaaaaaake]

I want full version of a song!

[@AI_ART_MASTER]

THANK YOU MR.NERDY RODENT

[@v33tay]

WaveSpeed is actually cool thing, but it feels like it only works with some samplers (euler, dpmpp_2m, maybe some more). Other samplers does not affected by the node and have the same as base generation speed.

[@Huggingfaceuser]

Better to get a used 3090 for $800 or a new 5070 Ti for $750? I'm not sure if a 4090 would be worth it unless it dropped under $1,000 at the very least.

[@Movieboy777]

i dont really see any performance improvements on windows.. am i missing something? are there any other things i have to have installed?

[@NerdyRodent]

If you try with a cache value of 1 you should see a difference

[@AB-wf8ek]

I haven't tried myself on Windows yet because I've heard you need to install triton. I know there are pre-built whl files out there, but I keep hearing horror stories of changes to the Windows environment causing conflicts with other tools. I've also heard people running WSL instead and installing everything in there, but seems like an added level of complexity. I'm left undecided what to do.

[@NerdyRodent]

I don’t think it needs triton, but yes - I’d certainly upgrade to Linux for anything AI!

[@AB-wf8ek]

@@NerdyRodent Ah, I see. I think the issue is with the torch.compile node which offers additional optimization, but would require triton.

[@BBZ101]

so 8gb can run it right

[@philippeheritier9364]

I don't think that Comfy-WaveSpeed working without Triton. i have this error " KSampler "CachedTransformerBlocks' object has no attribute 'img_mod' I already try to install triton but it's a nightmare. Anyway thanks for this Nice tutorial

[@NerdyRodent]

I don’t think it does, but “pip install -U triton” will install the latest version for you!

[@philippeheritier9364]

Thanks for answering, it's not working for me

[@DaveTheAIMad]

One issue, when you go to install wavespeed it asks what version then gives a huge list. but unlike other nodes that do this annoying step...it does not say "latest" and if you choose the default "nightly" nothing happens. what version did you install?

[@NerdyRodent]

I always use nightly ;)

[@DaveTheAIMad]

@NerdyRodent ah. I am having to reinstall my pc fully tomorrow (failed attempt to switch to Linux lol) so will use nightly when I get windows installed. Side note.... Deleting a Linux partition stops your dual boot from booting, oops lol. I am planing to use anaconda and the GitHub version of comfyui, seems an easier way to be able to add pip dependencies when needed. Hopefully that's not madness :)

[@FusionDeveloper]

I don't think wavespeed works on a 1080 ti.

[@DaveTheAIMad]

@@FusionDeveloper no idea, I use a rtx 3090

[@AI-Rogue]

Thank You!

[@SpinAngelo137]

anyone else having trouble with the LTX demo workflow? "SamplerCustom backend='inductor' raised: FileExistsError: [WinError 183] " a file can not be written that already exists" I'm so off with windows.....i think i will install linux, it's so annoying always.

[@DarioToledo]

So Wavespeed ain't suitable for a 4gb VRAM card?

[@NerdyRodent]

?

[@DarioToledo]

@@NerdyRodent I mean, I have a 3050, I can't use wavespeed?

[@midnightstories5695]

Crusty 3090 indeed !!!

[@magimyster]

Unfortunately, I have to rely only on my CPU😅

[@lennoyl]

thanks for the video. I will try it. I can't manage to make Huyuan works (I tried everything for a week now, got only a lot of error messages) but maybe it will help me to make LTXV work better (it works but only once, it crashes at each second rendering. something wrong with the RAM management maybe) Argh KZbin put back that f**bip**ng automatic translation. that's awful. I understand English so I don't need that bad translation. Fed up losing my time with YT bugs/behaviour (once it's the dark mode, now, again the translation...) they are lucky they have no real competitor

[@dwainmorris7854]

Everybody's sucking my feet the main problem with flux is not to speak it's the accuracy when you're putting image to image we need accurate Fiux not faster Flux

[@ronbere]

wave speed fonctionne très bien. teacache réduit la qualité de l'image par contre

[@Smirnoff67]

Est-ce que tu sais si ça apporte des gains sur des models SD1.5 ?

[@ronbere]

@ du tout, j'utilise flux

[@unsound-methods]

I have not tested it with Flux yet, but with SDXL I get pretty serious degradation in quality. I mean I can get close-to-vanilla experience when cache value is set low, but then there is no time win. And when I increase the cache value - yes, render times are faster, but quality suffers drastically (blur, noise, artifacts). (WaveSpeed tests)

[@unsound-methods]

it almost feels like it converts SDXL to a bad quality turbo version or something.

[@unsound-methods]

but yes, NR - one of the top most quality ComfyUI channels. Thank you for your work and effort.