@cinemaVerse I wonder if this kind of process can survive interpolations and upscales, as the two parts of the image are different

Ah, I see. Thanks!

@cinemaVerse Thanks for getting back to me so quickly. In reply to your three points:

1. Do you have a recommended upper resolution limit that you've found Wan2.2 can support and works well with your lora/workflow? I'm happy to have generation time take longer if the results are crisper and/or more coherent.

2. I wasn't aware that Wan2.2 would support the doubling of the frame count to 161. I seem to recall (from when I used Wan2.2 regularly) that it used to create a tight loop of events if you went over 81 frames, and could end up ignoring sections of your prompt. I'll give it a try, as you suggest. Hopefully a nice way to get smoother motion.

3. The "Stereo Stabilize" node: I'll definitely try this one out. Do you consider it to be a node that can provide improvements "for free" or does it have known side-effects (e.g. softer image, reduced 3d depth)?

Also, I've been testing out adding other loras into the mix. To be honest, I assumed it would cause the whole thing to collapse, ruining the 3D effect but, so far, it seems to be working OK. I'll test it a bit more but the initial results are very promising.

@cinemaVerse I've had a good play with this lora and workflow now and am consistently getting worthwhile results. It's a bit hard to get used to the Wan2.2 slow speed (per second of video generated) after using Minimax H3 and especially LTX2.5 but it's just such a damned reliable model for not screwing up.

I gave up on the stereo node as one of its settings was making the brightness/contrast zoom up and down throughout the video and the other one was making a small improvement to the depth effect at the cost of lots of weird artifacts along the edges of any moving person/object.

Your advice regarding pushing the frame count up (and effectively doubling the fps) does seem to stabilise the jittering effect - still visible in some videos but not as bad. The predicted downside is there - the old Wan2.2. trick of wanting to return back to the first frame again by the end of the video. Not really a problem if your video is something simple like a person turning around, though no amount of prompting will get them to stay facing away at the end of the video. And of course it more than doubles the time it takes to create the video.

Nevertheless, this would be my advice to anyone serious about using this lora to make 3D SBS videos: Don't bother with any extra stereo nodes; seriously consider putting the frame count up from 81 to 161 frames; do experiment with using additional loras - you may find they don't impact the 3D effect negatively at all.

Thanks again for this fun/useful lora and workflow - and, of course, I'll be keeping an eye out in case you manage to get Minimax H3 to work with 3D. The good news is that it does seem to have some native understanding of 3D SBS so maybe it just needs a nudge in the right direction.

@bennyboy_77 You're getting faster speeds than Wan out of H3? For me wan takes 5-10 minutes to do what Minimax takes 30+ minutes to do. And that's with "turbo"...

I reckon per pixel produced I'm getting quicker videos out of H3, yes. Not sure if it's because I've got a lot of RAM (128GB) and a fair bit of VRAM (24GB) available. LTX 2.3 and 2.5 are scarily quick compared to either but I find that they're only useful in a few circumstances (usually where better sound is the priority).

@bennyboy_77 That makes sense. VRAM probably the biggest help there I think, I can only fit a chunk of the model on my 16GB 4060. Ram is only 64 but it never seems to fill.