Видео

dude thats awesome! Youre blazing now! my new favorite app.. just incredible

​@@cognibuild yewspewed erased my reply here where i told you about my channel and my video that i put shouts out and links for your video and channel - i guess because i put the name of my video in the message ? is that stupid or what ? - yewspewed CONSTANTLY wipes my messages let's see if this message persists or gets erased

@@cognibuild the name of my vid is ' First AI ComfyUI / Hunyuan Video '

i noticed that too... will have to correct that next time. Turn up the volume though and its ok

true.. its definitely a new technology.. im excited for when they add new personalities

@@DM-dy6vn if you want to take me to the next level you can add an output knowed that saves all of the images used in the video.

thanks for that info

i gotta agree. its hitting next level. I love this stuff

@@cognibuild it’s going to be a game changer when we get the ability to write them personas like Silly Tavern. I just imagine the AI’s being a DM in a DND game. Like imagine having the narrator of BG3 now being the narrator of a story like that. It really opens up a whole new realm of possibility, plus even more accessible as this tech develops. The creativity that’s going to come from all this stuff will be astonishing.

16GB right? Should totally work! You can also try the WSL install with Sage Attention that I mention in another video

@cognibuild well then I'll give it a shot. This weekend I think. Most likely just gonna sub to your patreon for that installer.

@@Appreciation-Community awesome! Follow the instructions but if you need help hit me up Disocrd and I'll be glad to assist

that sounds like a workflow error.. look for sometihng on the left side of the workflow that says "bfloat" . .change it to to float16

@@cognibuild a new comfy install solved it, now it has "pytorch version: 2.5.1+cu124", old install had "pytorch version: 2.1.1+cu121". Cuda was the problem...

@@q6g36 ahh.. you have CUda 12 ?

@@cognibuild thats what the portable version of comfy comes with...

yes.. the initial load in WSL is PAINFUL.. however, once you invest in the initial load the inference speed is excellent. I usually load it up and do 20-30 generations at a time and it pays off

@@cognibuild In my case of RTX 3090, the generation time was not halved, but still reduced by 25%. I guess, this older GPU is not as fit for fp8 as 4090 is. I checked the read rate of mounted drives vs. native. The mounted suck big time. Wondering if it would make sense to have models in the "native" Linux folders

@@DM-dy6vntheres still hope! Go check out the GGUF models that just released!

merge?

Thanks!

@@DM-dy6vn thank you. Is that missing from the code I shared?

@@cognibuild The REG code in your blog is using Left Double Quotation Mark (“) and Right Double Quotation Mark (”) which are perfectly preserved if copy-n-pasted into Notepad++. I guess, they are translated into (normal) Straight Double Quotes if pasted in a more simple text editor.

@@DM-dy6vn hmm.. let me go see if I can change the way it dispays them. thank you

@@DM-dy6vn seems the best I can do is leave a note to remove formatting :< seems to be a restriction of my template., I'll have to look to switching that ast some point. again, thatnks for pointing that out

no.. however a friend and I have been working on an improvement to LTX image to Video that i should be realeasing soon

im glad to hear youre enjoying it. It really is a wothwhile hobby, Yeah, Hunyuan isnt going to work on a 6gb but theres a lot of other cool stuff you can do with images and even getting into chatbots. Let me know if you want help finding projects to work on :)

@@cognibuild you are right it probably won't work but worth a try and if something does go through with it I'll let you know so you can let your fanbase know cause there are people like me that can't afford it at the moment and sure they be like me very grateful for people like you that are willing to help others wish the world was like us how we all use to share and help others like the days we were grew up in.

hmm.. weird. Post the actual arror and i;ll take a look. Come find me on Discord

yes it awesome huh?! i use the Sage workflow in WSL so I dont have to mess up my Windows build changing stuff around.

thats awesome! And you can always try a lower quant as well and see how that works> let me know the results!

@@cognibuild I can fairly reliably create 50 frame videos at 768*512 using the BF16 GGUF. I run out of memory at 100 frames. Generation takes about 270 seconds. The Q8 model is close but slightly worse. It's 15-20% faster, and a bit more likely to get a bad video. I had one that wasn't even close. Q5 is worse and I'm not impressed. Now that I tried these I'm not impressed with the low vram non-quantized model. I'm not getting good quality video at all, buildings and such are pretty distorted, fall foliage looks like psychedelic trees. I also had to reduce the size to 640x480 All from the perspective of a casual user learning about AI stuff and with far too many AI toys to play with.

@@davidwootton7355 thank you for sharing your experiment feedback! I'm with you.. the Q versions are interesting in that they can load very fast on my machine, but tbh, my sage wsl build works great and I have no need to move from it. These quants ar better for people who cant get into the fp8 /bf16 versions.

That's exciting! It's because of the "quantizied" nature of the differnet versions. Think of it like precision decimals.. ie. you have the number: 1.23456789 which is pretty precice but takes up 9 slots .. what if we made it: 1.2345 instead? It's still the same number but not quite as precise. However it only holds 5 slots. Quants work in the same way. 517 seconds still seems long to me.Try going down to Quant 5 and see if you can get the speed up! Let me know the results!

@@cognibuild Yeah but I saw someone on an 8GB card not being able to run it in the comments so I'm wondering why I can with only 6GB VRAM, I'm assuming cause I'm on Windows and it shares some of the RAM with the GPU but I'm not sure I'll play with the other quants to try to find a good perf/quality ratio.

@@robertbencze8205 I'm not certain what you mean. LOL did I say that in the video? You definitely need Vb tools installed.. that video should lead you to a link that you can just type in your command prompts to download it

@@rogersnelson7483 that's good to know but that's a bummer. Try dropping it down to an even smaller dimensions (200 x 200) taking the frames down to 35. I'd be interested to see if it works

What you have to do is drag the image on top of your comfy screen UI screen. You can also right click on the json file beneath it and choose save as

@@cognibuild What I can't find is the json file, I can't find it to download, can you put the link please?

@@marcosgrima2237 comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/hunyuan_video_text_to_video.json

@@cognibuild I downloaded it but it doesn't work, the screen goes black and nothing can be seen

@@Nelwyn I'm with you where having a 4090, my sage attention in WSL build tends to work better. There's a distilled version that just came out yesterday as well and I'll try to get that up.... That might help your speed at the cost of quality

@@cognibuild I think it just comes down to frames and steps. If I disable the blockswap node then I get 49 frames in 2 minutes with 20 steps, but that's the limit of the vram. I've been settling on 225 frames in about 10 minutes using 20 double blocks as it seems like a good balance. I can't wait to get a 5090. Imagine what you could do with 32GB of VRAM!

@@Nelwyn drools in 32gb .. tbh, i want it to run bigger and better chatbots!!

@@cognibuild Same I love using 70B but it's SO slow.

@@Nelwyn i can get a 30b running at low quant. .but then at that point is like why bother

you can just keep reducing the tile size it uses less and less vram and i cant tell the difference. what version of the worked best on your 4080?

@@Baka_Oppai The tile size won't reduce more than to 128, there's only two settings available in the node, unless you can type ones in, I didn't try that. I also tried the Q8 model and it worked on my 3060 too by the way.

@@Alex_Niko_Y honestly if your other install is already working really good you're not going to gain much here. But if you need a speed boost this might help

And thanks to a Super Fan!!! :D <3

@@Chaz-x1i go to the video manager and update comfy. Close the server and start again

@@cognibuild Yup, that worked, thanks.

@@cognibuild Hey! What do you mean by video manager? Running into the same issue lol. Disregard, I just had to update comfy via the batch, custom manager update wasn't working. Thanks for the video!

@@loktar00 glad to hear it! btw, video manager is a game changer with comfy UI.. heres installation directions. real easy and I walk through how to do it in the video: github.com/ltdrdata/ComfyUI-Manager

@@Heisenberg238 nope just incredibly real text to video

Estimated release for an image to video version is in January

@@kingfirebone2000 thanks!

you know.. im unceratain. I'll keep my eyes open on the subreddits and if you hear something as well please let me know!

its becaues your dependencies arent being installed correctly. RUn this system checker and see if Git, conda, and python are installed. www.patreon.com/posts/automated-system-117200313

@@cognibuild I ran the checker and these are installed: Miniconda: Installed - Version conda 24.7.1 Git: Installed - Version git version 2.47.1.windows.1 Python: Installed - Version Python 3.12.4 CUDA: Installed - Version v10.1 CUDA: Installed - Version v11.8 CUDA: Installed - Version v9.0 Visual Studio Build Tools: Installed FFmpeg: Installed - ffmpeg version 6.0 What am I missing? Thanks!

@whatup2003 so the issue is that you need to downgrade your python to 3.10.11 on a side note you should probabloy get ridf of the CUDA 10.1 and 9.0 unles you know exactly why you have them (you shuld only have 1 cuda installed)

simply follow the directions in the video and you wil be fine

yes the issue is that yhou are using an AMD.. the trorch and CUDA is for NVIDIA. Im uncertain it will work, but try the install but use a normal versaion of torch (w/out cuda) and dont install the CUDAtoolkit

@@cognibuild but when using a normal version of torch, it gives the first error: AssertionError: Torch not compiled with CUDA enabled I'll have to wait if something appears in the future for the AMD/WINDOWS version. Thanks again for your attention and top quality tutorial

@@geniodestemido then it probably means that it doenst work with NVIDIA at all.. you can still try the ComfyUI version

Thank you so very much for your gift support! Super fan Weilandsmith! <3 I'll mention you in my next video :D