New v5 workflow crashing where v4 worked fine
Hello.
Just downloaded the new v5 basic workflow for 10Eros. First pass with it is WAY slower than the same exact setup with v4 (about twice slower), while the LTX Tiled Sampler step in the second pass always crashes with OOM, (even on 512p low resolution 10s videos, whereas the same exact prompt and seed finishes correctly for 1024p 20s videos with the v4 workflow).
For reference 32Gb system ram and 16Gb 4080 RTX gpu.
Reverted back to using v4 and everything is dandy now, but given that v4 is not available to download anymore, not even on the /old folder, this is something you might want to check.
If needed I can provide logs and sample images, but not openly here given the NSFW nature of the prompts and initial frame image.
(Might be a memory management issue with some newer loader node of v5 or some new Lora, as memory usage skyrockets as soon as the ltxvideo model is being loaded even in the first pass to almost x1.5 times the memory usage on the same step of the v4 workflow, and the Tiled Sampler on the second pass always ends up paging to the SDD when loading, and eventually crashes by OOM, whereas the same exact initial setup on v4 never reached full memory and never needed page file)
There's an issue with v5 where the sizing controls aren't actually hooked up. Any image you add is getting rendered at full size on upscale pass and first pass may be huge. The width and height sliders are supposed to go to the first resize_v2 node I only recently noticed they were disconnected.
Instead of connecting the mxTools i did the following with great results,
added ScaleImageToTotalPixels and connected that so that it keeps aspect ratio and handles a resolution for the specific image that my machine can handle, for megapixels its basically your base resolution you want to use for example i want 576p so my megapixels = (1024*576) / 1000000 = 0.589824, the other resize nodes will take care of making the resolution work with 32 multiple, also did some experiments with a single pass only and the results are really good
Anyway take my recommendation with a grain of salt, what looks good to me can be subjective, you can experiment or not and just use what the OP recommends
There's an issue with v5 where the sizing controls aren't actually hooked up. Any image you add is getting rendered at full size on upscale pass and first pass may be huge. The width and height sliders are supposed to go to the first resize_v2 node I only recently noticed they were disconnected.
Also, now that you mention it I see that the width/height controls of the Mask subnode of the top "Force first-pass audio" node are also unhooked, (probably for the same reason the size are not linked to the first image upscale). I believe those ones do nothing, but I will wire them too as they were on v4, anyway. Will sort this out and report how it goes.
Ok. After attaching the size nodes to the upscale pass using the v4 wiring as reference the new v5 is working fine. Thanks, and sorry for not noticing it before. Quality seems a bit worse than v4 though overall. I might have had bad luck with the seeds, will keep cross checking just in case.
Also noticed that the last Video Combine VHS step does not save the output. nothing fancy, but you might want to also check this
Also noticed you disabled the JoyAI_Echo lora that was added on v4. Again, is this intentional?
I believe I got some good results messing with its strength before, but not sure if it's just placebo effect or not.