Thank you for a great video. I would like to add some of my findings with I2V after exploring LTX (0.9.0): - With mp4 compression, for realistic video, optimal CRF value is ~ 30, 50 60 will degrade quality a lot and cause diverse and extreme motion. For art and anime, there is usually no need for detailed texture so if you want more motion CRF 50 will yield more diverse animation - For prompting, keywords like slowly, gradually will make smoother and more consistent animation but the output will quite in slow motion, so interpolate and increase FPS is a good choice. Add very slowly will make it even slower. - LTX is very good at landscape/ video without people. So I2V or T2V with detailed prompts most of the time will yield great results. - Don't include multiple or sudden camera movement keywords (e.g. shift, suddenly etc.) as it will create morphing animations (i.e. from frame 0 to 20 it has consistent video, then will suddenly shift to entirely different scenes) - For dancing/ more rapid character movement, I have good results with this kind of prompt "[Subject] begins to dance with graceful precision, his movements fluid and intentional".
I have tried LTX quite a few times already, overall has great experience with it. Though one thing I struggle with is camera movement, e.g. zoom out/ pan right/ pan left etc. doesn't seem to work. Have you found a consistent method for camera movement?
great channel! thx a lot but i have 1 error at the end of the flow: ImageConcanate: Sizes of tensors must match except in dimension 2. Expected size 111 but got size 105 for tensor number 1 in the list.
just deledet the compare nodes, since i dont need it now error is gone, also noticed if you tell in your prompt to move slowly and use 30fps + steps 40-50+ in sheduler you get a much better result-> still renders it quite fast, which is amazing on a 10 years old pc + rtx 3060 with 12gb ram :-)
models\llm\Florence-2-Flux-Large it can be download in here manually : huggingface.co/gokaygokay/Florence-2-Flux-Large/tree/main Or use Download Florenece 2 Loader node, auto download.
Nope. While it doesn't show up in the Install Missing Custom Nodes, it is searchable to install from Custom Nodes Manager and works 100% when installed from there.
hello, anyone can help me ? its getting this error: Missing Node Types When loading the graph, the following node types were not found Florence2ModelLoader
Anyone else getting cross-hatching in moving highlight areas like a flickering flame? Looking for solutions in DaVinci Resolve. Topaz upscale seems to ignore cross-hatching. I was wondering if I should optimize footage in Resolve before upscaling in Topaz? Anyone else working in this way?
Sadly, like with the LTX text to video it comes out horribly deformed. I put a 4 second video I made with Hunyuan, which turned out perfectly, of Snow Site petting a fawn and I put in the prompt "snow white petting a small red dragon with wings". It came out with Snow White's head being oversized and the dragon looking like a grossly deformed aardvark. LTX is just garbage, bro. It's fast, yeah, but at the sacrifice of anything looking remotely normal. Also, virtually none of the LTX workflows I found ever worked, usually because of something about "missing 1 required positional argument: 'attention_mask", which I have no idea what it means or how to fix it. The only Comfy video thing that ever worked properly and with good results is the Hunyuan one. LTX stuff will just drive a person insane.
The guy in the apocalyptic city wouldn't be able to walk with a knee like that 😆 Jokes aside tho, in curious why it ignored the boy turning 180 while walking away from the camera.
Hi, I got always this error "Error(s) in loading state_dict for VideoVAE: size mismatch for decoder.conv_in.conv.weight: copying a param with shape torch.Size([1024, 128, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 128, 3, 3, 3]). size mismatch for decoder.conv_in.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([512]). size mismatch for decoder.up_blocks.0.res_blocks.0.conv1.conv.weight: copying a param with shape torch.Size([1024, 1024, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 3, 3, 3]). size mismatch for decoder.up_blocks.0.res_blocks.0.conv1.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([512]). size mismatch for decoder.up_blocks.0.res_blocks.0.conv2.conv.weight: copying a param with shape torch.Size([1024, 1024, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 3, 3, 3]). etc"
Have you used something that gives greater control? Personally I find everything pretty much the same at the moment: just with different ways of getting to the same results.
ComfyUI-LTXTricks
github.com/logtd/ComfyUI-LTXTricks
Attached workflows of 2, LTX with STG and LTX Flow Edit V2V (Freebies):
www.patreon.com/posts/ltx-video-0-9-1-118605761?Link&
Thank you for a great video. I would like to add some of my findings with I2V after exploring LTX (0.9.0):
- With mp4 compression, for realistic video, optimal CRF value is ~ 30, 50 60 will degrade quality a lot and cause diverse and extreme motion. For art and anime, there is usually no need for detailed texture so if you want more motion CRF 50 will yield more diverse animation
- For prompting, keywords like slowly, gradually will make smoother and more consistent animation but the output will quite in slow motion, so interpolate and increase FPS is a good choice. Add very slowly will make it even slower.
- LTX is very good at landscape/ video without people. So I2V or T2V with detailed prompts most of the time will yield great results.
- Don't include multiple or sudden camera movement keywords (e.g. shift, suddenly etc.) as it will create morphing animations (i.e. from frame 0 to 20 it has consistent video, then will suddenly shift to entirely different scenes)
- For dancing/ more rapid character movement, I have good results with this kind of prompt "[Subject] begins to dance with graceful precision, his movements fluid and intentional".
Loras are also possible, the model really has potential. I believe they have now also announced official Lora support.
I know a guy even fine tune this base model. And it awesome! Next video talk about this.
@@AIBusinessIdeasBenji that sounds great! 🙂
Thanks for the video, What is the min vram for this?
Bro, your channel is going to give me a heart attack. LOL! You're a machine with these vids updates. Awesome work.
hehe.. thanks. Have fun on holiday!
I will continue with the web app development for Hunyuan Video Web UI. ;)
I have tried LTX quite a few times already, overall has great experience with it. Though one thing I struggle with is camera movement, e.g. zoom out/ pan right/ pan left etc. doesn't seem to work. Have you found a consistent method for camera movement?
It need a very detail prompt for camera panning. and version 0.9.1 does improve on prompt following.
Good luck ;)
great channel! thx a lot but i have 1 error at the end of the flow:
ImageConcanate: Sizes of tensors must match except in dimension 2. Expected size 111 but got size 105 for tensor number 1 in the list.
just deledet the compare nodes, since i dont need it now error is gone, also noticed if you tell in your prompt to move slowly and use 30fps + steps 40-50+ in sheduler you get a much better result-> still renders it quite fast, which is amazing on a 10 years old pc + rtx 3060 with 12gb ram :-)
Great video! One question. Where does Florence-2-Flux-Large install to?
models\llm\Florence-2-Flux-Large
it can be download in here manually : huggingface.co/gokaygokay/Florence-2-Flux-Large/tree/main
Or use Download Florenece 2 Loader node, auto download.
@@TheFutureThinker Thank you so much! Just going through this now. Cheers.
Which is better, this one or the Nvidia one?
Great vid ! We have V2V that to be run on PC friendly like AnimateDiff before!
Yes this is it. The next .
The LTX-Video Node is still broken in Comfyui Manager Installer. The only way is to install manually via Git URL.
Nope. While it doesn't show up in the Install Missing Custom Nodes, it is searchable to install from Custom Nodes Manager and works 100% when installed from there.
what is the minimum vram
hello, anyone can help me ? its getting this error: Missing Node Types
When loading the graph, the following node types were not found
Florence2ModelLoader
in my experience so far, LTX 0.9.1 has higher VRAM requirement than LTX 0.9, even though 0.9.1 is half the size.
Yes, more stuff to extract in the execution
Anyone else getting cross-hatching in moving highlight areas like a flickering flame? Looking for solutions in DaVinci Resolve. Topaz upscale seems to ignore cross-hatching. I was wondering if I should optimize footage in Resolve before upscaling in Topaz? Anyone else working in this way?
Sadly, like with the LTX text to video it comes out horribly deformed. I put a 4 second video I made with Hunyuan, which turned out perfectly, of Snow Site petting a fawn and I put in the prompt "snow white petting a small red dragon with wings". It came out with Snow White's head being oversized and the dragon looking like a grossly deformed aardvark. LTX is just garbage, bro. It's fast, yeah, but at the sacrifice of anything looking remotely normal. Also, virtually none of the LTX workflows I found ever worked, usually because of something about "missing 1 required positional argument: 'attention_mask", which I have no idea what it means or how to fix it. The only Comfy video thing that ever worked properly and with good results is the Hunyuan one. LTX stuff will just drive a person insane.
The guy in the apocalyptic city wouldn't be able to walk with a knee like that 😆
Jokes aside tho, in curious why it ignored the boy turning 180 while walking away from the camera.
I can't seem to 0.9.1 model to run on my 2080ti without getting OOM error
2080ti ? Have you emailed LTX ask why it happened?
@TheFutureThinker no, I didn't think to do that
Benji stays on it.
Always on it.
Hi, I got always this error
"Error(s) in loading state_dict for VideoVAE:
size mismatch for decoder.conv_in.conv.weight: copying a param with shape torch.Size([1024, 128, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 128, 3, 3, 3]).
size mismatch for decoder.conv_in.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([512]).
size mismatch for decoder.up_blocks.0.res_blocks.0.conv1.conv.weight: copying a param with shape torch.Size([1024, 1024, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 3, 3, 3]).
size mismatch for decoder.up_blocks.0.res_blocks.0.conv1.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([512]).
size mismatch for decoder.up_blocks.0.res_blocks.0.conv2.conv.weight: copying a param with shape torch.Size([1024, 1024, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 512, 3, 3, 3]). etc"
Check out 02:20 . VideoVAE error because you need the new VAE loader for the version 0.9.1 model VAE archecture.
@@TheFutureThinker Yes I have update comfyUi from manager, but still got error, maybe I will update in another way, I'll try, thanks
Try not to use ComfyUI manager to update if that is the case. Git pull.
CR 文本替换
“list”对象没有属性“replace”
Very hard to control the output result. This is curently not the game changer as you said.
depends how you use it, if you said control? like a video editor? it won't be. it's AI.
Have you used something that gives greater control? Personally I find everything pretty much the same at the moment: just with different ways of getting to the same results.