As a person with a career in tech I go into these tutorials knowing that I am gonna have to jump through hoops or technicalities that aren't mentioned. This went so gosh darn smoothly just following your instructions. many thanks for this.
Exactly my experience too as most of these tutes if more than a week or two old won’t work as comfyui and basically everything gets updated very quickly while many of the dependents nodes / code is abandoned.
11:58 "Does not look like a dragon". It's literally the Toothles from "How to Train Your Dragon" with red eyes!!! I can't belive that someone does not know this character!
also the soldier scene at 16:00. like wtf is this guy living in a cave, does he not a fps shooter when he sees one? im convinced this channel is some AI robot voice BS channel
Previous leader in open source video was CogVideoX 1.5 from The Knowledge Engineering Group (KEG) & Data Mining (THUDM) at Tsinghua University. This model from Tencent looks much better. And it doesn't matter that it's from China since the model is a "safetensor" file so just the neural network weights, and the code to do inference is open-source from whomever you want like the GitHub user Kijai
Using examples he says is his that are the exact same examples I saw other YT people use yesterday, because they are actually the examples on the website from the people that made the program. Looks like somebody is playing the money game. Look what YT has become.
I'm pretty sure the AI advancements in this decade will result in the end of video proof being accepted as legal evidence of anything (on its own). Like, there's gotta be a point where it's literally impossible to tell real and fake apart, even by experts.
Ultimately this is also very very early. I suspect by the end of January we will have multiple refined models that will run on different gpu’s. Source has been out less than a week and we already have a first community build. I can’t wait to see what happens over the coming weeks. I personally use ai for vfx and composites. I think the goal will be tp have models running on 8-12gb gpu’s and at 720p. There’s plenty of video upscalers already available
@@theAIsearch I think the most logical step will be smaller more specialized models like we have with Img generation. Id like to see one specifically for photorealism, then for animation, then for 3d animation. Obvioly there would be nsfw variants for thos who ummm. . .need it lol. That being said i'm really interested in the lipsync mode. That could be its own dedicated model.
was about to write it^^ exactly, and cause "How to Train Your Dragon" is an animated movie I would say it has done this part pretty good, even if dreamworks isn't disney :D
That anime scene was insane. It really looks drawn and especially - Framerate is like animation, not like this "Gatcha game style"-Kling look. That's really impressive
another tricky prompt is to ask the AI to generate a turnaround video. sometimes the face changes after the turn; or the model acts like an owl; or it seems the head is rotating but it only show the back of the hair.
The requirements are.... "An NVIDIA GPU with CUDA support is required. We have tested V100 and A100 GPUs. Minimum: The minimum GPU memory required is 11GB. Recommended: We recommend using a GPU with 32GB of memory for better generation quality."
The estimates from the model creators tend to be pretty conservative. I've heard of people getting it working on 8GB with the right settings but expect to wait 10+ minutes for a single generation.
@@KManAbout how many generations will you get in 1 hour? 4 seconds like the premium ones out there? ive worked out you getting around 100 vids for $30 per month
Both CogVideoX 1.5 5b and this can run on lower-end consumer GPUs with quantization and tiling. The quality may be worse and inference takes longer, but 12GB GPU isn't that far out of reach for many people
it's sometimes better, sometimes worse. I think a lot of it might come down to sheer randomness. run the same prompt a few times and you might get something entirely different, with either model. comparing videos based on a single attempt is actually kind of silly - these models are very random and might create you something with each attempt that you happen to either love or hate.
holy moly, using this with 45vram rn (on lightningai, L40S gpu on the free tier) the quality is really next level, now I'm getting interested in video generation, we're finally getting out of the weird "embryonic" phase, like when we had stable diffusion 1.4 also, free and uncensored? 😯wow
@@HikingWithCooper yes, free tier. You download the model there, free up to 100gb of storage. You need to register with phone number to get the free mothly credits
I just need something that works well for AI filmmaking. So, it has to have a good ‘Image to Video’ tool, camera controls, a start frame & end frame feature, video to video, video inpainting, lip-syncing, the ability to have multiple consistent characters interacting with each other, etc.
I think the more you can specifically correlate a new, abstract, innovative idea to existing ideas while providing differentiating features..it will probably be more likely to provide results that match the picture in your mind.
Based on the understanding of what it's "foundational model" was trained upon. As with all communication it is important to KNOW YOUR AUDIENCE..Understanding where they come from..thier frame of mind..or..the dataset it was trained and...weighted...on. Is important more than all else. As it is with humans.
7:30 Kling AI is definitely the best out of what is shown here. Ultra realistic details on the horse and environments, looks exactly like Id expect from a Hollywood production, the only giveaway is the infinite sand/fog around the feet
In the soldier combat dip and swing prompt. I think if you look carefully at each result. You can see the average Ai interpretation of "dip and swing". It shows up..order of execution wise..towards the end of the clis
That looks like toothless from How to train a dragon which was made by DreamWorks Animation. The others look like a generic dragon. So this does look like a dragon firm, a 3D animated film. Just not Pixar.
Hi there, when can you make an updated video with the ehnace-a-video node, and IP2V (If it's working) to hold us over until I2V comes out? As always, great video!
How much VRAM do you need to create still images / photos? Can you provide an image, and have the tool change parts of it (such as a different face, or different skin tone / color, or different hair style, or different weight (make the person skinnier or heavier))? I would like to try this Hunyuan, free, open source software. I have a 4070 ti with 12 GB of VRAM. Based on our host's demonstrations, videos would be very short and somewhat low resolution, with only 12 GB of VRAM. But how about still images or photos? Also, once an image is created, can it be tweaked? For example, tell the tool to have the person smile, or wink, etc? Will each tweak require more and more VRAM?
Has anyone gone through all the code to verify that this is not some kind of Honey-pot/Trojan/Backdoor scheme? Why would a Chinese company just give this away? How does this benefit them?
I mean whats the difference to a US company doing this? Maybe they just wanna show they‘re not falling behind and become as relevant in the ai industry as possible Or it‘s just a backdoor by the ccp 😂 guess we‘ll never know
oh it's definetly a backdoor alright, but you prob should be safe if you run it on a VM you don't care about, and in the case this can leak trough host, maybe it'd only work on windows hosts, so consider using a linux host or smth but i doubt this is a backdoor, like it just why would it be? the least they do is sell your personal data
Why are people worried about "CCP backdoors" when its OPEN SOURCE 😂 why do i get the feeling 99% of the online "techbro" fanbois have Z-E-R-O clue when it comes to basic software terms?
I am starting to think that one guy was just determined enough to not read any books that he invented AI text to image generators for the sole purpose of being able to directly watch the books and never ever read again.
The Pomerianian puppies generation did follow your prompt. One puppy is the sous chef teaching the others who are not yet chef's hence they do not have the chef's hat or uniform, they're new students. Case of garbage in garbage out, you didn't specify they all needed to be wearing outfits.
I can't even get Midjourney to produce a static image when I use the term "running away". It always renders subject 1 running towards subject 2. or both subjects running with their backs to the camera. I wonder if you try different variants like "dragon chases princess" or "princess flees dragon" or something similar?
Very impressive indeed, though there are few clips shown here that don't at least have a bit of "hmm... feels off" with clear signs that it was at minimum assisted with A.I. Passing most of these off as legit camera recorded vids would be idiotic at best. But damn this is insane and amazing
Downloaded both models. I cannot choose bf16 quantization in Comfy for big model. Only fp8. Both models are generating without errors, but the video is black.
27:51 i get a missing node types but more than just the one you got mine says its missing HyVideoModelLoader HyVideoSampler HyVideoVAELoader DownloadAndLoadHyVideoTextENcoder HyVideoTextEncode VHS_VideoCombine HyVideoDecode is there any way to fix this?
For me, it's because I didn't restart python main.py before refreshing. It didn't work like in the video, I had to re-run the command before refreshing
It's looks like they used very big amout of data of films, anime and cartoons which tenct can showing in China as training data, so sometimes it literally showing fragments of animation from cartoons or anime in results.
I tried 3 videos including examples provided, exact same models and settings, and i get videos of only a black image and nothing else. Nvidia 4090 here. EDIT: Updating comfyui bat file as well as from manager didn't fix issue, but running update_comfyui_and_python_dependencies bat file fixed it for me. others on github said a fresh install also works, but i seen no reason to do a fresh install when i just installed comfyui only weeks ago.
@@HikingWithCooper working great on my 4090, i'd still use smaller sizes shown in this video to give you some room to generate videos of a couple seconds longer, but yea it blows my mind how well it's doing considering all the other local ones i tried in past have been absolutely terrible. I am not sure it's something i'll personally play with for long, it's kind of a cool thing to try out then move on. but i have been impressed with what it has done so far with my prompts. It is still hit or miss of course, like it may take a few minutes to generate a video by default settings then maybe 25% of time you get something fairly decent, 75% it's okay but with glitches, and like 5-10% of time you get something really cool. I only been playing with it for a few hours.
Great video! I do have one question, though: is it possible to segment the project and use just one of its features? I'm particularly interested in the image-to-animation/lip-sync functionality, as it seems like a great alternative to Hedra. Additionally, would focusing on just that feature make the model lighter? I’m not a developer, so I’m not sure if what I’m suggesting is feasible or even makes sense.
If you want to use local models like this, you need to do image to video starting with the text as you want it to appear and then modify it to video using the AI. Not expected to do it. Start to finish.. 13:22
My 16gig vram almost works but gets an OOM error on the final step. Any advices on how to tweak it a bit more to save on ram/vram? [edit] Reducing the frames works. Should've watched the whole thing.
There was definitely a few impressive generations from Hunyuan.. I'm not denying that! But in my opinion, overall, Kling still performed the best out of all of them. But thank you for taking all the time and presumably money to put this research together. If it's the only locally compatible generator out there, I may still try it!
It's good to be suspicious, but the models are "safetensor", which is a format that's merely neural network weights. The inference code is open-source. That said, people should never trust random ComfyUI extensions and their Python dependencies, so maybe being suspicious is smarter.
@@Ddotkay Im more confused about what this technology is actually good for. I think all the hype about these techs is just that, hype, hype over nothing. This may end up going down in history as the ultimate vaorpware
Thanks to our sponsor Abacus AI. Try their ChatLLM platform here: chatllm.abacus.ai/?token=aisearch
video to video wont work for me i have rtx 4080.. it give me length error
that was the "train my dragon" dragon, exactly.
🤣LOL i create better animations on my channel than any of this AI generated JUNK!
Is it working on apple M3 Max with 128G ram... can it work on mac?
when I click on "try" it turn to a Chinese version.. Any advice?
i hate spons-
WTF is going on? USA have close source censored models and China is dropping open source uncensored models.
lies coming to the surface
Probably spyware. Trying to catch people to use it so the Chinese can spy on you.
I wouldn't trust a Chinese model with a 10ft pole. I don't care how good it does, I know there's gotta be a catch.
@@armartin0003
Lol .... look who is saying that! A brainwashed US individual who is wearing an underwear that was made in China! 🤣😅
china : now who is the real evil?
As a person with a career in tech I go into these tutorials knowing that I am gonna have to jump through hoops or technicalities that aren't mentioned. This went so gosh darn smoothly just following your instructions. many thanks for this.
@@joehick8102 the more quickly you follow a tutorial the less chance there is that some part of the process has been updated and broken the tutorial
Exactly my experience too as most of these tutes if more than a week or two old won’t work as comfyui and basically everything gets updated very quickly while many of the dependents nodes / code is abandoned.
And you don't mind that the AI code you just installed on your PC is made in China? You should seriously reconsider.
@@majorhavoc9693 Open source means that you can check the code to see if there are malicious programs, genius.
Thank u, as a person of similar caliber that is nice to know.
2:54 What's really impressive here is that it actually changes the position of her chest to match her taking in a deep breath.
I was going to say the same, impressive!
11:58 "Does not look like a dragon". It's literally the Toothles from "How to Train Your Dragon" with red eyes!!! I can't belive that someone does not know this character!
However, it still loses points because Toothless is a Dreamworks style dragon, and the prompt was Disney/Pixar🤣
at least he kinda recognized Po
yeah blows the mind
also the soldier scene at 16:00. like wtf is this guy living in a cave, does he not a fps shooter when he sees one? im convinced this channel is some AI robot voice BS channel
Yeah, it would seem as though the AI model was trained on a lot of DreamWorks content
Thanks
"crushes everything" and then proceed to demonstrate how literally every other model is better.
LOL! Exactly what I was thinking!
crushes every open source?
Previous leader in open source video was CogVideoX 1.5 from The Knowledge Engineering Group (KEG) & Data Mining (THUDM) at Tsinghua University.
This model from Tencent looks much better. And it doesn't matter that it's from China since the model is a "safetensor" file so just the neural network weights, and the code to do inference is open-source from whomever you want like the GitHub user Kijai
Looks much better than the old open-source king, CogVideoX 1.5 5b
Using examples he says is his that are the exact same examples I saw other YT people use yesterday, because they are actually the examples on the website from the people that made the program. Looks like somebody is playing the money game. Look what YT has become.
When you said uncensored, well ... I will use this for research purpose only.
"Doesn't look like a dragon" that is literally Toothless from How to Train Your Dragon.
That's because this guy is actually an A.I. and it doesn't get the dragon description very well...
It's because Toothless does not look like a dragon.
00:40 i know what you thinking 💀
I am just thinking about physics😂
@@HikingWithCooperproof?
ok what the sigma
@@HikingWithCooperso what would be the current industry leader? Asking for a friend
@@amrazaq I suppose this for video. Still not great but it’s only 2024.
that's actually insane! the video model looks so unreal!
fake bot paid by this channel cause no comment only like bro just stop you propaganda with ugly chinnesmodel not free and not in 2k hd
I'm pretty sure the AI advancements in this decade will result in the end of video proof being accepted as legal evidence of anything (on its own).
Like, there's gotta be a point where it's literally impossible to tell real and fake apart, even by experts.
9:28 i love the way the bottom left one just spits out all of the spaghetti
Ultimately this is also very very early. I suspect by the end of January we will have multiple refined models that will run on different gpu’s. Source has been out less than a week and we already have a first community build. I can’t wait to see what happens over the coming weeks. I personally use ai for vfx and composites. I think the goal will be tp have models running on 8-12gb gpu’s and at 720p. There’s plenty of video upscalers already available
yes, things move so fast! can't wait to see what updates we get next
@@theAIsearch I think the most logical step will be smaller more specialized models like we have with Img generation. Id like to see one specifically for photorealism, then for animation, then for 3d animation. Obvioly there would be nsfw variants for thos who ummm. . .need it lol. That being said i'm really interested in the lipsync mode. That could be its own dedicated model.
@@theAIsearch Are you going to do a comparision of all the open source models available?
god daym, 8-12 gb vram for just 720p videos? i aint doing nothing with my rtx 3050 4gbvram
@ lol I mean I got my 3060 for $300. Honestly they are really fair priced now. Mines a 12gb. Plus you can upscale 720p easily with like Topaz or Krea
Eta on image to video? This is the most important one that changes everything.
Kijai's plugin for CogVideoX has image to video. CogVideoX, even 1.5 with 5b param is much worse than this though
Agreed. I'm looking forward to this as well.
7 minutes to generate
@@Rucky8887 minutes for what? Image to video has been released?
@@thays182 also wanting this bad.
11:32 it does look like a dragon, just specifically toothless from how to train your dragon
(Edit: timestamp was wrong)
was about to write it^^ exactly, and cause "How to Train Your Dragon" is an animated movie I would say it has done this part pretty good, even if dreamworks isn't disney :D
Only much much scarier. Even scarier than all the other dragons. It was directly staring into my soul :|
True - thx for sharing! Admittedly, I have not watched it
@@theAIsearch Shame on you. Go watch it. But i liked only first chapter
What is this time stamp? 😂😂
It looks like kling is still way better, and it impressed me way more than the other AI.
That anime scene was insane. It really looks drawn and especially - Framerate is like animation, not like this "Gatcha game style"-Kling look.
That's really impressive
another tricky prompt is to ask the AI to generate a turnaround video. sometimes the face changes after the turn; or the model acts like an owl; or it seems the head is rotating but it only show the back of the hair.
For a second I thought this was a suspicous advertisement for a second because of the thumbnail 😭🙏
2025 is going to be a freaking wild year for AI and AI generated content.
21:24 lmao I wasn't expecting Po to appear as a result of this prompt 🤣
Umm... The title of the video says it "crushes everything", but it performs worse in almost every one of the Text-to-video generations...
well it was compared to not open source paid generators that are older and so more trained?
I think it’s useful for the audience to appreciate that if you did those prompts 3-5x you’d get very different results from them all.
Thanks for your hard work. Great video. In your opinion, in a long run which would you suggest between Kling and Hailuo?
The requirements are.... "An NVIDIA GPU with CUDA support is required. We have tested V100 and A100 GPUs. Minimum: The minimum GPU memory required is 11GB. Recommended: We recommend using a GPU with 32GB of memory for better generation quality."
The estimates from the model creators tend to be pretty conservative. I've heard of people getting it working on 8GB with the right settings but expect to wait 10+ minutes for a single generation.
Cloud compute is a thing. You can rent a GPU like an h100 for 5 dollars an hour.
@@KManAbout how many generations will you get in 1 hour? 4 seconds like the premium ones out there? ive worked out you getting around 100 vids for $30 per month
Both CogVideoX 1.5 5b and this can run on lower-end consumer GPUs with quantization and tiling. The quality may be worse and inference takes longer, but 12GB GPU isn't that far out of reach for many people
@@KManAbout Then it's not free is it...
Kling is still better by a long shot but for a free program, it's impressive.
the uncensored part is what hooks me up
@@Psychopatz have you tried it?
kling is censored, no boobies in kling
it's sometimes better, sometimes worse. I think a lot of it might come down to sheer randomness. run the same prompt a few times and you might get something entirely different, with either model. comparing videos based on a single attempt is actually kind of silly - these models are very random and might create you something with each attempt that you happen to either love or hate.
its really impressive
13:30 this is what I found the most amazing out of it all. That looks like it was actually animated by a studio. Incredible.
Text to video, video to video but where's the image to video?
not out yet
Where is video to video? All I can find is text to video on the website.. and they’re only 5 seconds long
holy moly, using this with 45vram rn (on lightningai, L40S gpu on the free tier)
the quality is really next level, now I'm getting interested in video generation, we're finally getting out of the weird "embryonic" phase, like when we had stable diffusion 1.4
also, free and uncensored? 😯wow
@@HikingWithCooper yes, free tier. You download the model there, free up to 100gb of storage. You need to register with phone number to get the free mothly credits
Yeah, I'm also interested in this lighting AI thing. How's the free tier work?
I hope you'll make another tutorial on the hunyuan picture to video after its releasing. You always explain things really well.
I just need something that works well for AI filmmaking. So, it has to have a good ‘Image to Video’ tool, camera controls, a start frame & end frame feature, video to video, video inpainting, lip-syncing, the ability to have multiple consistent characters interacting with each other, etc.
Eureka moment for Avatar Music Video!!!
Thank you Admin❤❤❤❤
how come 500k people agreed on watching a 39 minutes ad
13 billion parameters, not too shabby
I think the more you can specifically correlate a new, abstract, innovative idea to existing ideas while providing differentiating features..it will probably be more likely to provide results that match the picture in your mind.
Based on the understanding of what it's "foundational model" was trained upon. As with all communication it is important to KNOW YOUR AUDIENCE..Understanding where they come from..thier frame of mind..or..the dataset it was trained and...weighted...on. Is important more than all else. As it is with humans.
Until it sports an onboard universal "wild human" translator.
Hunyuan can work on Google collab, or free colud source like kaggle environment, please try on that
7:30 Kling AI is definitely the best out of what is shown here. Ultra realistic details on the horse and environments, looks exactly like Id expect from a Hollywood production, the only giveaway is the infinite sand/fog around the feet
Yo bro keep it up the video is fire!
Cia spying vs ccp spying, you can choose
@@captaintroll7294I'mma take the ccp one
@@captaintroll7294 Why not both?
When do we think the image to video workflow will be released?
You know how the RUclips algorithm works, putting naughty thumbnail 😂
🎉 great review video🎉 you did a great job comparing these
thanks!
Are we sure we want to run a chinese owned a.i on our computers.
Yes
it's free for a reason, they can earn by selling your data :V
How's it any different from running Western Owned AI? You think your precious NSA isn't spying on your ass? XD How CUTE.
@@sinfulgrace One doesn't sell your ID and Password to russian kids, other does.
As opposed to... ClosedAi? 😂😂😂
i died at the terra cota dancing lmfao
Where, time stamp?
@@ShifterBo1 I am only 5 minutes in and already saw it. Too busy to go back but you'll see it pretty early on
it was pretty camp too.
And this is free... that's just nuts! Will need to give this a try.
It's free because it's from china. Keep an eye on your bank account if you do any banking on that PC.
@@majorhavoc9693 Good point, that's why I use VMs
@@majorhavoc9693 its open source...
@@majorhavoc9693 lmao.
11:55 -- actually it is the dragon from "how to train dragons" a dreamWorks movie universe.
I was going to comment the same
In the soldier combat dip and swing prompt. I think if you look carefully at each result. You can see the average Ai interpretation of "dip and swing". It shows up..order of execution wise..towards the end of the clis
That looks like toothless from How to train a dragon which was made by DreamWorks Animation. The others look like a generic dragon. So this does look like a dragon firm, a 3D animated film. Just not Pixar.
They did him dirty
True - thx for sharing! Admittedly, I have not watched it
What’s most impressive to me here, 2:42 is his chest when he inhales. 🤯
不得不说，这个视频的创意真是满分！
Wow MiniMax is pretty amazing.
7:49 I think you might be underestimating just how many Pomeranian chefs are out there tbh
I noticed there is no Image to video after following all directions. Well, at least I made a video.
Hi there, when can you make an updated video with the ehnace-a-video node, and IP2V (If it's working) to hold us over until I2V comes out? As always, great video!
the image to video(dancing) making the wrinkles in the dress appear as she moves , is incredible.
I wasn't convinced with the terracotta warrior clip, i needed the anime girl clip as well to be sure 😂😂
that dragon is the "train my dragon" dragon, exactly.
so basically every other ai model coming out every week "crushes" all the previous ones?
the new iPhone is the fastest iPhone ever LOL
Apparently it's a game changer and will blow your mind.
its been like that for years
No one: The camels: 0:12
no one: muumipahpa 4 hours ago:
No one: IN-pr3lw 18 hours ago:
wtf is going on here
here on going is wft
wtf is going on here
Dragon clearly ripped off of How to Train Your Dragon.
Unbelievable, but when is it likely we would also have image to video feature with Kijai's version of Hunyuan?
HOW IS IT FREE???
Cause you spend $4k on the local computer you run it on. So it's "free"
How much VRAM do you need to create still images / photos?
Can you provide an image, and have the tool change parts of it (such as a different face, or different skin tone / color, or different hair style, or different weight (make the person skinnier or heavier))?
I would like to try this Hunyuan, free, open source software. I have a 4070 ti with 12 GB of VRAM.
Based on our host's demonstrations, videos would be very short and somewhat low resolution, with only 12 GB of VRAM. But how about still images or photos?
Also, once an image is created, can it be tweaked? For example, tell the tool to have the person smile, or wink, etc? Will each tweak require more and more VRAM?
Has anyone gone through all the code to verify that this is not some kind of Honey-pot/Trojan/Backdoor scheme? Why would a Chinese company just give this away? How does this benefit them?
I mean whats the difference to a US company doing this? Maybe they just wanna show they‘re not falling behind and become as relevant in the ai industry as possible
Or it‘s just a backdoor by the ccp 😂 guess we‘ll never know
oh it's definetly a backdoor alright, but you prob should be safe if you run it on a VM you don't care about, and in the case this can leak trough host, maybe it'd only work on windows hosts, so consider using a linux host or smth
but i doubt this is a backdoor, like it just why would it be? the least they do is sell your personal data
@@chibisayori20conflicting statements
Why are people worried about "CCP backdoors" when its OPEN SOURCE 😂 why do i get the feeling 99% of the online "techbro" fanbois have Z-E-R-O clue when it comes to basic software terms?
US companies do something: zermg! 🥴 😍 ❤️ Capitalism 🗽
China does something: 😡 😡 😡 THEY MUST BE UP TO SOMETHING, THOSE DEVILS
Even AI learned how to pull crowds with thumbnails
Tried this on my 4GB VRAM laptop... It's as bad as it sounds
I am starting to think that one guy was just determined enough to not read any books that he invented AI text to image generators for the sole purpose of being able to directly watch the books and never ever read again.
The Pomerianian puppies generation did follow your prompt. One puppy is the sous chef teaching the others who are not yet chef's hence they do not have the chef's hat or uniform, they're new students. Case of garbage in garbage out, you didn't specify they all needed to be wearing outfits.
Where is the learning g to cook part?
11:56 it made the dragon toothless from how to train your dragon :O
Thanks for another great comparison and installation guide!
The dragon in "Princess runs away from dragon" is clearly based on Toothless from How to Train Your Dragon.
I can't even get Midjourney to produce a static image when I use the term "running away". It always renders subject 1 running towards subject 2. or both subjects running with their backs to the camera. I wonder if you try different variants like "dragon chases princess" or "princess flees dragon" or something similar?
Of course try variants, like princess stands in front of dragon then runs towards camera. Try synonym generators like fleeing or being chased by
Things like putting in "frontal view shot" helps
Very impressive indeed, though there are few clips shown here that don't at least have a bit of "hmm... feels off" with clear signs that it was at minimum assisted with A.I.
Passing most of these off as legit camera recorded vids would be idiotic at best.
But damn this is insane and amazing
Downloaded both models. I cannot choose bf16 quantization in Comfy for big model. Only fp8. Both models are generating without errors, but the video is black.
My Vids are black too.
@@Strakin did you figure out the problem?
@@petejorative No sorrry. Maybe something about the torch installation.
I can't even generate anything, it keeps saying it can't find the text encoder but idk wtf any of that means
Thank you 🙏 but i want to know how to make a movie, or short movie with consistent characters?
27:51 i get a missing node types but more than just the one you got mine says its missing
HyVideoModelLoader
HyVideoSampler
HyVideoVAELoader
DownloadAndLoadHyVideoTextENcoder
HyVideoTextEncode
VHS_VideoCombine
HyVideoDecode
is there any way to fix this?
For me, it's because I didn't restart python main.py before refreshing. It didn't work like in the video, I had to re-run the command before refreshing
Same problem, except I could find and install the VHS_VideoCombine, any else not in the Manager/custom nodes list
oh sh*t... :) run the .bat file as Admin, something couldn't update w/o rights...
@ ok I’ll try Ty
It's looks like they used very big amout of data of films, anime and cartoons which tenct can showing in China as training data, so sometimes it literally showing fragments of animation from cartoons or anime in results.
I tried 3 videos including examples provided, exact same models and settings, and i get videos of only a black image and nothing else. Nvidia 4090 here. EDIT: Updating comfyui bat file as well as from manager didn't fix issue, but running update_comfyui_and_python_dependencies bat file fixed it for me. others on github said a fresh install also works, but i seen no reason to do a fresh install when i just installed comfyui only weeks ago.
looks like you need to update torch: github.com/kijai/ComfyUI-HunyuanVideoWrapper/issues/55
ima try this now thanks glad to see im not the only one
it is fixed
@@HikingWithCooper working great on my 4090, i'd still use smaller sizes shown in this video to give you some room to generate videos of a couple seconds longer, but yea it blows my mind how well it's doing considering all the other local ones i tried in past have been absolutely terrible. I am not sure it's something i'll personally play with for long, it's kind of a cool thing to try out then move on. but i have been impressed with what it has done so far with my prompts. It is still hit or miss of course, like it may take a few minutes to generate a video by default settings then maybe 25% of time you get something fairly decent, 75% it's okay but with glitches, and like 5-10% of time you get something really cool. I only been playing with it for a few hours.
@@zengrath,can it work without internet? also, i have OLD GPUs can i RUN them with 2 3070TI + 1 3070? thanks
Super natural... may you need to adjust your empathy sensors...
@1:31 my mind is absolutely blown at the ability to use stick models to make a single image dance🤯
I don't care that much about that stuff...
uncensored!
Well, i might give it a try 😅
Image to video, then we'll talk
isn't 1:39 doing that?
Not yet. At the end it says it’s not released yet. Image to video was a released example, but looks like we can’t use it yet.
try hotshot, its in 2k hd and the best since long time
i cant wait for my crush to do what ever i please :)))
Ese whey@@NostalgicMem0ries
Great video! I do have one question, though: is it possible to segment the project and use just one of its features? I'm particularly interested in the image-to-animation/lip-sync functionality, as it seems like a great alternative to Hedra. Additionally, would focusing on just that feature make the model lighter? I’m not a developer, so I’m not sure if what I’m suggesting is feasible or even makes sense.
Arent you supposed to have 60gb of vram to even get to 720p?
Watch the whole video
@@pengiPenguin We did. Anything less creates a 2 second horrible quality gif pretty much. so YES You do need a beast of a GPU
@@HikingWithCooper thank you that is exactly the answer I was looking for
If you want to use local models like this, you need to do image to video starting with the text as you want it to appear and then modify it to video using the AI. Not expected to do it. Start to finish.. 13:22
11:32 They did toothless dirty
It struggled with the soldier request. It looked like it pulled it straight from an old version of “Call of Duty”.
Amazing, let's hope this opens a new era of free and democratic models
12:25 maybe tell the placement of dragon and direction of running for the princess
That moment when the model is so good the GPU can't handle it
that was my first thought, i had heard for a local install you need some industrial scale gpu to get results
I think you should have tried a tilt shift video
My 16gig vram almost works but gets an OOM error on the final step. Any advices on how to tweak it a bit more to save on ram/vram? [edit] Reducing the frames works. Should've watched the whole thing.
14:00 WOW !!!!
THIS IS THE BEST ANIME STYLE !
I went to high school with a white guy named Will Smith. That wasn't him.
There was definitely a few impressive generations from Hunyuan.. I'm not denying that! But in my opinion, overall, Kling still performed the best out of all of them. But thank you for taking all the time and presumably money to put this research together.
If it's the only locally compatible generator out there, I may still try it!
it's really crazy how nobody is talking about the book Nifestixo The Hidden Path to Manifesting Financial Power, it changed my life
itadori and megumi sneaking in at the top right clip at 13:57
Ah yes. Free. Tencent. Nothing fishy here.
xd
It's good to be suspicious, but the models are "safetensor", which is a format that's merely neural network weights. The inference code is open-source. That said, people should never trust random ComfyUI extensions and their Python dependencies, so maybe being suspicious is smarter.
There isn't. Ignore at your own peril.
for text i have learned to use . and on the wall there is written : Subscribe . works with flux for example
Can this also do image to video, ? thanks for the great tutorial, but without image to video, its no good for me.
i'm 3 days into these open source models and my mind still cannot comprehend how this is legal
why?
Why do you say that?
@@Ddotkay Im more confused about what this technology is actually good for. I think all the hype about these techs is just that, hype, hype over nothing. This may end up going down in history as the ultimate vaorpware
@@aegisgfxwrong. Major disruption for advertising and marketing.