v10 is still the best version overall
v10 gives me really good nsfw prompting results for i2v. The same prompt (and i tried many variations) does not work in the mega versions for me, it just wont do what i am asking. I hope this will get improved, the only thing that works much better for me in the mega versions is First Frame - Last Frame Transitions.
Many of us have noticed the same thing, but it's been difficult to provide concrete examples. But we see what we see. Regardless though, what phr00t has done and is doing is amazing and we are very grateful for his work :)
Yeah, I really need concrete examples so I can have a "target" when testing new versions.
I've been working on improving my "Qwen Image Edit" model lately, because First to Last Frame videos are really powerful combined with Qwen Image Edit: quickly make your Start then edit it to make your End (or generate an End frame with the same seed), then animate between the two. Far better success rate in my experience. Can't do that with v10 I2V. So, I'd consider v10 the best I2V model, but that is all its good at (if that is all you need, I'm happy you are finding it useful). Personally, I'd much rather improve upon "Mega" to enhance its I2V use case without losing features I've come to rely on.
@Phr00t
Qwen Image 2509 is a two sided sword. It is great in changing angles/poses of characters. But on the other end it is only good with limited resolutions around 1MP. And there is the problem with image shifting in the result image. The second problem is a really complex one where it is part fault by the model itself and part fault by the confy ui nodes. But after a longer search and many wrong informations about how to fix the shifting/blurry picture results I stumbled on this reddit post that helped me to create a workflow that works with zero/close to zero image shifting: https://www.reddit.com/r/comfyui/comments/1nxrptq/how_to_get_the_highest_quality_qwen_edit_2509/
I already created a workflow based on this reddit post where you select just one front side image of your character and you get as much different angle shots as you like with just one excecution. I can upload this workflow if you are interested. I'm working on a similar worklflow that does the same but instead of different angles with different poses.
You're wrong, the best version was the V7 nsfw... I still use it now and there isn't a single ruined frame, it all depended on the type of wavimage you used...
You're wrong, the best version was the V7 nsfw... I still use it now and there isn't a single ruined frame, it all depended on the type of wavimage you used...
I want to try but where can i get v7 nsfw version?
You're wrong, the best version was the V7 nsfw... I still use it now and there isn't a single ruined frame, it all depended on the type of wavimage you used...
I want to try but where can i get v7 nsfw version?
I removed it because it was buggy, overfitted to motion you didn't prompt for and shifting faces. See threads like https://huggingface.co/Phr00t/WAN2.2-14B-Rapid-AllInOne/discussions/30 and https://huggingface.co/Phr00t/WAN2.2-14B-Rapid-AllInOne/discussions/39 and https://huggingface.co/Phr00t/WAN2.2-14B-Rapid-AllInOne/discussions/28
Newer models have been greatly improved and should be used instead.
I did more tests:
v7 / v10:
- good for nsfw motions
- characters are fidgeting/restless when told to not move and hold a pose
- bad at doing first frame-last frame transitions
megav5 / mega6:
- can't do any basic nsfw motions
- very nice smooth motions, no fidgeting
- really good at first frame-last frame transitions
is there a good keyword to prevent the fidgeting in the vXX versions?
I did more tests:
v7 / v10:
- good for nsfw motions
- characters are fidgeting/restless when told to not move and hold a pose
- bad at doing first frame-last frame transitions
megav5 / mega6:
- can't do any basic nsfw motions form my tests.
- very nice smooth motions, no fidgeting,
- really good at first frame-last frame transitions
is there a good keyword to prevent the fidgeting in the edgesvXX versions?
It is a double edged sword. v10 may have more motion, but it is still a bit strong and as you mention, you get it even when you don't want it. As a base model, I don't want to overfit anything as the prompt should be king. I haven't noticed significant problems with basic NSFW motions in Mega, but you do need to be more descriptive on the motion you expect to see. I'm not done improving though as I find ways to do so.
it is indeed and i am grateful for your work! quick question the mega builds are they 16 or 24 fps? as i understand it the Vxx versions are 24fps and mega 16?
it is indeed and i am grateful for your work! quick question the mega builds are they 16 or 24 fps? as i understand it the Vxx versions are 24fps and mega 16?
Mega is 66% Palingenesis and 33% SkyReels 720p, so I think it is a bit fuzzy. I usually get 16fps outputs with v6.
Great Phr00t teacher, thank you for your hard work