skyreels I2V Final video output error

[Hazukiaoi]

when use 10 step, the final video look like:

And if use 30 step:

The add-on version is 2997d65

the source img is :

Is there a problem with my parameters, or is there a bug in the plugin?

[kwas-123]

I get the same, either totally static output or the character becomes a cloud.

[kijai]

The model sadly doesn't seem to perform well with anime/cartoon inputs.

[Hazukiaoi]

But when I tested this image with the online version, the results were excellent.

72afe2b0-f9ee-4d0b-9980-5e960d1d647d_w.mp4

[Hazukiaoi]

When I use a resolution of 544x960 and 50 steps, I can generate a normal video about half the time, but there's still a huge difference in quality compared to the official website's results.
and add-on commit at ecd60a6.

When I'm lucky, the results are something like this.

But most of the time, it looks like this.

[pftq]

Try changing attention from sageattn to spda. It'll be much slower sadly but I found the results to more stable then and less of the "airbrush" look.

[Hazukiaoi]

if use spda, 960*544 will be OOM.

[pftq]

Ah then you just don't have enough VRAM to run at full quality. That said, the online version doesn't appear to be the same model as the open-source one so maybe in the future there is more optimization/improvement in the offline one.

For context, I'm having to run at 100 steps on SDPA to get something comparable to the website.

[pftq]

One other thing I noticed is you can improve the image quality without increasing steps/VRAM by lowering CFG but it comes at the cost of less prompt adherence. It seems the default CFG of 6 is too high (you need more steps to sustain CFG 6). I did quick tests on your image here, same seed for consistent comparison while keeping the rest same in your setup like SageAttention and Torch.

20 steps and 3 CFG
(less VRAM/time to look good but less prompt adherence)
https://github.com/user-attachments/assets/2a073056-90a9-4f72-b59a-b8a09c49f757

100 steps and 6 CFG
(more prompt adherence but more VRAM/time to look good at high CFG)
https://github.com/user-attachments/assets/f3021c23-a205-4ae0-81de-c09dd6eab445

The camera controls are a bit difficult and a bit luck of the draw to just generate until a seed gets it right. You can also force it with the orbit camera LORA like this, but LORAs always degrade the overall quality ever so slightly (I would retry without LORAs first):
https://github.com/user-attachments/assets/3ac28252-7d29-44dd-bb92-4e16925ddd51

You can get the workflow settings for these by dragging the clips into ComfyUI as well. The clips are short to get the point across - if you increase the frames, you'll need to raise the steps to compensate as well (or lower the CFG, basically maintainer a higher steps-to-CFG). For 8 seconds (which seems to be the hard limit for this model), I've generally been needing to stay at least 100 steps.