AI Video Generation - General Discussion, Tips, Tricks, Frustrations and Showcases

Brought these cuties to life using Wan 2.2.

Images from this post
https://www.akiba-online.com/threads/photorealistic-ai-generated-images.2123060/page-47#post-4928279

Video
https://mega.nz/file/bJci2SCY#DA4s1PCWfZJW2bkH9KaF8R82cpPrxk1G9czEZFG8-1w

columbussnake said:

Incredible! Truly inspiring.
Thank you Casshern2 for your work and for sharing.
Is the entire model A.I. generated or just specific parts?
Thanks again!

Click to expand...

Hey, there, sorry for the late reply, was under the weather for a bit. All AI using Forge and an initial model for the text-to-image named cyberrealisticPony_v125 then resized/upscaled using a model named lucentxlPonyByKlaabu_b20 for image-to-image

Death Metal said:

@Casshern2 I just saw this thread. I've recently started playing with ComfyUI and various workflows. I have been looking for something that would produce good image-to-video results, and so far, I've only got some success with WAN 2.1. It seems like you've been using WAN 2.2, right? Are you using a workflow that integrates both high noise and low noise diffusion, or a single WAN 2.2 model? From my experience, WAN 2.1 is competent but not great at adhering to text prompts; been using a CFG value of 1.5, as suggested in the post I downloaded the workflow from. Anyway, I still have a lot to play around, but I'd like to try my hand at a working WAN 2.2 setup. Can you share some more insights of your workflow and the models you are using currently? Maybe I missed the post if you already did.

Thanks!

Click to expand...

Hey, there, sorry for the late reply, was under the weather for a bit. Good deal on trying ComfyUI, hope you're having fun with it like I am. It can seem like a lot, going strictly by the visuals of it, but once you use a workflow or two as-is getting the mentioned models in place then watching a YouTube video here and there, it starts to get easier to manage and change. Don't be afraid to move things around for things to make sense for you, it's 100% flexible as you know, so all the workflows out there are just laid out however the author did it. I tend to move things around to make them try to fit on one screen.

I highly recommend watching this video from start to finish, multiple times maybe on one screen while you follow along on another if you have two. Real easy to follow; the guy's voice is friendly (probably AI generated but maybe not, I don't hear pauses or mispronunciations like you can on AI voices) and he takes you from start to finish creating a workflow yourself. This is what gave me better confidence and melted away any thought of "I can't do this stuff".

Wan 2.2 Image to Video for AI Commercials Comfyui GGUF

I haven't been active because of a stomach issue plus work, but as I'm typing this I have video a being generated with this image, it will most likely be done before I finish here.


At any rate, video generation just finished. It was the first of the night so it included all the models loading, 30min in total, including model loading, so basically around 16min or so. Subsequent runs are faster. I'll be genuine and post what came out, they're not all golden, it can take a while to get what you want, or just to avoid odd things and anomalies that can and will pop up. She's a bit all over the place because I put too many actions in the prompt. Something I picked up to avoid slow motion videos after reading a question from @kharo88 back in August. I'm sure if I remove something, she will behave haha.




EDIT: Wow, it actually took a bit longer to generate the new one (around 19min) without her hop in the prompt. I did notice it hung a bit on the prompt node, but not sure if that contributed to the length of the run that much. Oh well. What I learned to do pretty quick was not watch this paint dry, I either listen to music on YouTube or watch something and check back.

Here are some others I queued last night, same prompt and workflow, just fed it different images. You'll find even though they all had the same prompt, they won't all do the same thing. Probably a seed thing, I have it sent to Random after each generation, but even then based on the image being used it won't match actions exactly.