Mochi 1 In ComfyUI Tutorial Guide - Generate AI Video Using Native And Wrapper Nodes

[@TheFutureThinker]

Mochi 1 Text2Video Workflow - Able To Generate Multiple Seconds (Freebie) www.patreon.com/posts/115533849?

[@Relinked-Media]

Great video! Can you do a All in one installer?

[@TheFutureThinker]

@@Relinked-Media Will try it and see

[@AgustinCaniglia1992]

It's amazing how comfyUI keeps being allways updated and includes so many AI tools.

[@motionau]

Awesome work from Kijai on this one. Optimising Mochi so we can run it on consumer gpus.

[@TheFutureThinker]

Looking forward to img2vid weights release.

[@gjohgj]

Thanks for this! Very curious about the video2video flow:)

[@TheFutureThinker]

Mochi Edit? This one kzbin.info/www/bejne/inyXk2qrna2aoc0si=7dO7alISVpEgITUt

[@darksushi9000]

3090 here rendering 7 seconds of video in 35 minutes all day long zero crashes

[@darksushi9000]

Some more info: Ryzen 7950x, 32GB RAM using 17GB RAM, using 15.6GB VRAM. Diffusion model is preview fp8 scaled and vae is preview bf16

[@Donzo89]

No way, I was disappointment when he mentioned 4090. Hopefully my 4070 ti super 16gb can handle it.

[@technobabble77]

Runpod would get you there for a few bucks if you want to play with it

[@darksushi9000]

@@Donzo89 just about. Pretty sure I've seen a way to get it down to 12GB VRAM

[@kait3n10]

I read that you can use tiled VAE to overcome the OOM crash. People got it working on RTX3060 12GB! edit: Nvm, I saw your final section. Btw thanks for the tutorial!

[@TheFutureThinker]

Yup, started from raw workflow to tiling optimze 👍

[@Nibot2023]

can't wait for the img2vid portion

[@TheFutureThinker]

Once it releases, i will make a automate flow to create documentary, or narrative base videos. Hehe

[@AndikaKamal]

@@TheFutureThinker 👍

[@FusionDeveloper]

I want a ksampler that shows a correct accurate preview of frame by frame, as it generates. You can go down to 320x320 and as low of length as 13. You can choose 480x480 to save some time instead of 848x480.

[@theskyspire]

Maybe we didn't see a dragon, because you said "dargon" on the prompt.

[@TheFutureThinker]

Yea the first one. updated: here's the correct one : x.com/AIfutureBenji/status/1854781439365333412

[@MilesBellas]

Purz did a live review, before Halloween,where he rented out a H100 for a few dollars and got some interesting results quickly.

[@TheFutureThinker]

After trim down, the model does lose some quality in Comfyui

[@insurancecasino5790]

Once again. Amazing video. You don't have an image to vid interpreter? Just an extension right?

[@TheFutureThinker]

Yes, just the extension

[@GreenAppelPie]

Excellent!

[@aprismaaprisma4090]

I get in all workflows, even with the all in one this message: KSampler meshgrid expects all tensors to have the same dtype, Any idea?

[@TheFutureThinker]

Your comfy need update

[@Secantmonroe]

i tried mochi1 last month with RTX 3090 24gb it took 1 hr to produce a 6 s video , not usable , u still need to 3 retries to get the desired result \

[@dracothecreative]

Hey all, so how do i get the mochi vaeloader and mochi decode?

[@WiseOwlLearning]

which model do you recommend for RTX 4090?

[@TheFutureThinker]

Fp16

[@eduardomoscatelli2775]

How to do imageToVideo?

[@eliassuzumura]

You're a GOD

[@TheFutureThinker]

No I am not. And I only have one God , he is watching us. 😉

[@ryansenger937]

Can this do image to video? Or video to video?

[@FusionDeveloper]

Yes, there are workflows for it now, I think it is called "MoChi Edit".

[@Dinesh-x6l]

Image to video works?

[@TheFutureThinker]

Img2vid model weights coming soon. The current version i2v, but it have v2v nodes already. It is weird, but is it what it is.

[@SUP3RMASSIVE]

I got a 4090 and can only generate 1sec vids. Like u if I go higher it freezes on vae decode.

[@TheFutureThinker]

Yes me too, Watch till the end you will find the answer

[@getmonie393]

can I use models from last week? I have the 2- 20 gig models, gguf's, and vae. Theyre huge so I was hoping but I think it has to be new one?

[@getmonie393]

also, I'm on m2 max 32 gb so will that be an issue as well?

[@TheFutureThinker]

I tried the last week models (I guess you are mentioning the one from KJ Mochi Wrapper model) And last part of the video i mixed the VAE from there and using Native node for Sampler and Model Loader

[@TheFutureThinker]

For Mac, I haven't try it. Since 4 years ago , my iMac and Macbook, then I stop using Apple products.

[@KananaEstate]

Is it only 4090 that can run Mochi locally? Any other lower spec possible?

[@TheFutureThinker]

Recommend GTX 4090

[@giuseppedaizzole7025]

wow...this looks amazing..can this make img2video? thanks

[@TheFutureThinker]

There's video2video for Mochi, but no img2vid yet.

[@giuseppedaizzole7025]

@@TheFutureThinker ok...thanks

[@J-ld9cl]

Does it need more VRAM or RAM?

[@TheFutureThinker]

Don't know, if you have 100GB VRam and Ram or you have 1GB on each?

[@DaveTheAIMad]

so you need a 4090? a 3090 wont do it?

[@TheFutureThinker]

You can try. The 4090 was the model ComfyUI.org tested with.

[@FedorBP]

RAM is the same. Should work, but slower.

[@DaveTheAIMad]

@@FedorBP I did give it a try earlier, 6 minutes on the default settings, a little over a second of video... Still awesome we can run it on local machines. Just need a way to use our own starting images now :)

[@JohnVanderbeck]

gah my 4090 is in my sim rig, the AI rig only has a 3090ti :(

[@TheFutureThinker]

1 , 4090 is good enough to run this. No worry

[@JohnVanderbeck]

@@TheFutureThinker yeah but my point was the 4090 isn't in my AI machine, only a 3090ti which has been plenty until now. The 4090 is in my gaming/sim rig :D

[@rageshantony2182]

I tried a 10 second video with RTX 6000 Ada 48 GB. After 20 mins sampling, then it entered VAE, the VAE burst even with kjili tiled decoding ending the generation in vain.

[@TheFutureThinker]

10 seconds. Nice try. While the model is supporting 5 seconds video clip.

[@rageshantony2182]

@@TheFutureThinker I set the length as 241. But I unable to get the result due to VAE crash

[@TheFutureThinker]

The max length I did was 129. It was using the original model weights, not the Comfyui version

[@zGenMedia]

If you do not have a powerful GPU just watch... do not touch.

[@MilesBellas]

Wow.....

[@TheFutureThinker]

Local + video

[@amigoface]

can it run on 4070 12 gb ?

[@TheFutureThinker]

Yes but be patient with the loading

[@amigoface]

@@TheFutureThinker cool, after the loading , is the generation relatively quick in your opinion ?

[@TheFutureThinker]

@amigoface it feels like the speed I usually did in AnimateDiff for 15 seconds video.

[@amigoface]

@@TheFutureThinker ok thanks

[@MilesBellas]

Dual RTX cards with NVLINK needs support.

[@TheFutureThinker]

How about A6000 ?

[@Guus]

⁠@@TheFutureThinker I run it on A6000. 4 second video takes 5 minutes. Do you know if I can run 2 gpu’s at the same time? And how?

[@MilesBellas]

@@TheFutureThinker I have dual RTX A6000s in the main machine.....