Comfy updated? Are you using Img2Vid model? Which node do this happen?

Yeah I have discord: goldencrow777

Or you can send diret message in here :)

Wow! That's interesting! Thanks for show me this! As much as I love what kijai do, sometimes his nodes cause a lot of problems for general public due to complexity of instalation, etc. But In this case I don't see a drawback! i'll add this for sure after I test. Thank you!

Edit: Fantastic! After testing it, it's amazing! I'll update my workflows as soon as possible with this feature!

@Mikusha Haha! Thanks! Well, I'm not a big fan of changing resolution manually as well. Hence I always endup making my own workflow and share i nhere for the ones who are "lazy" like us.

@kg_09 Not really a speedup, but makes the workflow simpler... Instead of divide stepst in 2 samplers, kijai node can divide the Steps with and without CFG in one single sampler.

Change, cfg, seed. You will get different results. Depending on the image, any I2V model can do deform original image to create a different video. It's all about config, prompt and what are you trying to create.. The time per interaction depends on the latent size you define, also the length of the video. My workflow come as default with 720p as large size, which is large for video generation. If you change it to 512, like most of the workflows around, the time will go down to 10 seconds per interaction. You just need to pay attention to inputs.

The GGUF models for now, If I'm not wrong, are only TextToVideo, so to run it you need to bypass the "WanImageToVideo" node and use the workflow as a TextToVideo instead of ImageToVideo. Once they release the GGUF Img2Vid, it will work. I just got ahead with the notes because this workflow is similar to a previous workflow I have.

@Mikusha Great. Just tested it! Update the "Unet Loader (GGUF)" package and it will work. They added wan support in latest version. And conect it in Power Lora Loader (The node the ChackpointLoader was conected).

@Kiefstorm AH. My mistake. This is from an old workflow. This one don't need this.

When you paste the path in "Load Video", it will start to play your video as a preview, just like in "Video Combine". If the video don't start to play, something might be wrong with the path or format. Do you have the error for me?
Once it load the video, you just need to run the workflow and it will work.

Ah. It was a Mistake I forgot to fix in this workflow. If you want to run it as a T2V you can just add a node called "EmptyHunyuanLatentVideo" and connect it to the sampler. You can get width, height and length from the old "WanImageToVideo" node.

Sorry. My bad.

@Meenouse It's possible to run it with 12GB. Use a GGUF model, Q4 or Q5. Start with small "large size". 512px or 480px. And only 3 seconds length. Once you're able to get a result, ou can raise the values until the limit of your GPU. I don't remember the settings I run it in a 4070. I can chekc it later and tell you.

Try to run it with original config. The text is just an explanation why the workflow uses 2 samplers. The second half of the sampling can be with cfg 1, which saves time.
You can't get any video from the workflow?

Well... This is very specific. I also had problems trying to animate a little cute grass animal/monster when I was starting. I tried to change the prompt sometimes. In the end it's more like try and error. And if you can find a Lora with the movement you need or something similar, Yeah... This is the way. It's all new so I don't really know. I'm trying just like you. Hehe.

Yeah. It's not necessary. In fact Wan works at 16 fps. I just forgot to remove it. lol

AH. it's a problem with "VAE Decode (Tiled)" If you have enough memory, replace for the original Vae Decode and the problem will be gone. I'll update the workflow later.

@drukware So, the Normal Vae decode, solved this problem for me. But If I'm not wrong, the basic Vae Decode also split the encoding in Tiles if you have no enough memory to decode it at once. It could be this.

About Start-End Frame, I still didn't touch it yet. I'm giving it a time until I make sure everything is stable, because we end up with a lot of trash files in HD after the whole revolution of the model happens, and I'm trying to avoid it. lol

Wan is made to generate 16 fps in general. So 16 fps usually will be the best way to go. After that you can interpolate frames to reach the FPS you wish for final result! :)

Thank! In the first sampler, change the "control_after_generation" from 'fixed' to 'random', and you will get random seeds.

Yeah. It's possible. I need to update this workflow. Once i do it, I'll add one.

@BubbleButtPrincesses Added interpolation to the workflow.

If you follow the instructions in the red Note you might be good to go. It's very simple to use.

@matriksAi A big red Node with Instructions at the left of the workflow.

@aimodelmaya Solution for what?

@aimodelmaya Ah! Cool! Like a Tutorial?

Not possible atm from what I know. Wan changes the color of the frames a little bit. It will only be posisble once we have a I2V that don't change the image saturation so much. I usually correct it in a video editor to join the videos.

Hmmm... You can replace the "Load Video (Path)" for a "Load Video (Upload)". It will work.

Just did! :D

Done!

@Santaonholidays It all depends on your PC... How much Vram, which model are you using, Large size of Image, Video Length, etc...

@Santaonholidays So, I also have a 4090. I use it with the original config in the workflow. For a rectangular image (5:8 or 9:16) you can use the large size as 720 and 5 seconds, and it might work with sage. To use Teacache I think you need to reduce this a little bit... You can also try to use some GGUF smaller model. Q6 do a fine job in my tests.

@Santaonholidays I'm not sure. I Don't remember the values to make teacache work. I suggest you to start trying to generate a video with 512 large size and 3 seconds. Then you go up until the GPU cannot handle it anymore... lol The teacache you can leave as default for the type of video you're generating. T2V or I2V have diferrent setting...

@Santaonholidays It's the one I uploaded. I just don't use TeaCache. It's using the config I use for 5:8 Images. But it's really tight. If I open a second program that uses even a bit of the GPU, the generation time goes up sky rocket. With this setting I usually close everything else, turn off internet and leave it generating 4 or 5 videos, so I can pick the good ones after it finish the work. It takes around 6 minutes for normal generation and 2~3 with fast lora. But If I use Vram with anything else, the time goes up like hours. So I stop the work, close everything else and start again.

@Santaonholidays I use Crystools to check my PC usage on comfy. So I can adjust the workflow to what my PC can handle easilly. Once you see the resources used on screen, you will have a better idea on how to adjust it. I cannot tell you so much because it really changes from PC to PC. The best advice I can give you is, start with small length and small "large size of image", and go up slowly, until you understand what your specific case can handle. Also try with different image shapes. Square ones are the worst, becuase it resize to maxSizeXmaxSize, which is the worst case scenario for Vram usage. I uploaded the workflow with the cofig I use in my 4090 for rectangular images. And as I said, I close everything else, so it don't get stuck. It really take the 4090 to the limit in this config.

@Santaonholidays It's the dream XD I wish I could afford one of those server GPUs made of Unobitanium... With Gigallions of Vram...

@Santaonholidays Haha! Thanks a lot! :D