As far as the frame length like the link said it's pretty restrictive right now, the models are trained for 14 frames with regular SVD and 25 frames with SVD-XT.

The (untried) solution to that limit would just be to save a sequence and just use one of the looping nodes and a conditional to feed the last frame of the video the model generated back into another img2video to produce another 25 frames and repeat as desired.

I tried doing the feedback of the last image into another ksampler thing finally with an img2video... as long as your last frame was stable (sometimes it goes weird on you) it'll usually keep generating smoothly. If there's a ton of motion it might bounce back in a weird way or start doing something else but if the image is similar enough to the first it'll work. So keep motion lower maybe? It would be harder with txt2video since you can't easily inject the prompt conditioning into the conditioning the image produces. I just kept getting mismatched tensor sizes regardless of attempts to forcibly resize the conditionings. It's not predictable but it's fast enough that I don't mind. I feel like this process would work better with animatediff video2video where the image feedback from the first sampler probably wouldn't even be required so it could be set up as a loop more easily and allow for more optimal batch sizes for your GPU to be run with the same results even if you kept the length the same.

I used qq-nodes-comfyui's "image accumulator start" and "image accumulator end" to gather all the images produced, set count to 49, fed the first 24 of the batch from ksampler 1 into that, the last one into a clip setup for the second ksampler and then fed those 25 into the accumulator node which trips the count and sends it through the end node (needed a cr image list to batch, nothing else worked for some reason) to the animation output.

@shiroppo OK thanks👌