kijai
ComfyUI I2V with Kijai nodes freezes at sampler step
Hi! I'm having an issue with workflows that include Kijai nodes for I2V generation in ComfyUI.
Using the native I2V workflow instead ( with patch sage attention KJ node set to auto ), everything works fine and I can successfully generate a video.
However, when I try workflows with Kijai nodes, during the sampler step everything just freezes.
I've tried these workflows:
https://github.com/kijai/ComfyUI-WanVideoWrapper/blob/main/example_workflows/wanvideo2_2_I2V_A14B_example_WIP.json
https://civitai.com/models/1818841/wan-22-workflow-t2v-i2v-t2i-kijai-wrapper?modelVersionId=2058285
but the result is always the same.
I did a fresh installation of ComfyUI portable (v0.3.49), installed Sage / Triton, and updated all nodes to the latest versions. Here are my current versions:
Working on a 4080 gpu (12gb) laptop with 32gb of ram.
Even when trying I2V workflows with reasonably small images (480 - 640 pxs) and just a few steps (4 - 8), as soon as it reaches the sampler, everything freezes :(
I get no errors in cmd.exe, it's just that the sampling process stays stuck at 0% and never progresses. I waited around 20 minutes and nothing happens. Even clicking "Cancel current run" does nothing, and I have to close cmd.exe. I also tried disabling these nodes, but it didn't help.
Can anyone help me figure this out? Thanks!
You are probably out of memory and you should enable the block swap nodes:
Hi Kijai,
Thank you for all the work you have put in.
I am facing the same issues as the OP. On my native Wan 2.2 I2V workflow, I am able to generate a video with resolution of 1088 x 1088 in 6.5 minutes flat, and 1 MP video in 6 minutes (101 frames, 14B FP16 models). Using your high and low noise 4-steps LoRAs.
I am running an RTX 5090 (32 GB) with 96 GB of RAM.
I have tried playing for hours and hours with your workflows and I am absolutely unable to get any of them to work whatsoever. The workflows always freeze at the high-noise sampling itself. This includes the basic I2V workflow, the LongCat workflow, and some others too. Somehow, I was able to get the Infinite Talk workflow going, but it needs 30 mins for a 10-second 896 x 896 video. I even tried the fp8 model for LongCat, but even that threw OOM errors or simply froze for even a 960 x 960 generation.
I have tried playing around with all kinds of offload vs. no-offload settings, block-swap, low-vram mode for LoRAs, etc., and I've tried many many combinations, but alas!
Do you have any idea what I may be missing here? I cannot get the workflow to work even with 960 x 960 resolution, which is a significant drop from the 1088 x 1088 native version that works in 6.5 mins.
There are so many awesome workflows that are being built with your wrapper, and I'd love to try them all out, but it is unfortunate that I am being unable to get them to work!
I know you must be very busy but any help would be great! Thanks for all your hard work again! :)
You are probably out of memory and you should enable the block swap nodes:
Hi Kijai,
Thank you for all the work you have put in.
I am facing the same issues as the OP. On my native Wan 2.2 I2V workflow, I am able to generate a video with resolution of 1088 x 1088 in 6.5 minutes flat, and 1 MP video in 6 minutes (101 frames, 14B FP16 models). Using your high and low noise 4-steps LoRAs.
I am running an RTX 5090 (32 GB) with 96 GB of RAM.
I have tried playing for hours and hours with your workflows and I am absolutely unable to get any of them to work whatsoever. The workflows always freeze at the high-noise sampling itself. This includes the basic I2V workflow, the LongCat workflow, and some others too. Somehow, I was able to get the Infinite Talk workflow going, but it needs 30 mins for a 10-second 896 x 896 video. I even tried the fp8 model for LongCat, but even that threw OOM errors or simply froze for even a 960 x 960 generation.
I have tried playing around with all kinds of offload vs. no-offload settings, block-swap, low-vram mode for LoRAs, etc., and I've tried many many combinations, but alas!
Do you have any idea what I may be missing here? I cannot get the workflow to work even with 960 x 960 resolution, which is a significant drop from the 1088 x 1088 native version that works in 6.5 mins.
There are so many awesome workflows that are being built with your wrapper, and I'd love to try them all out, but it is unfortunate that I am being unable to get them to work!
I know you must be very busy but any help would be great! Thanks for all your hard work again! :)
It's mostly been torch.compile problems, I sat down today to go through the code after learning some new things and I believe I figured some things out that should help, so try the latest update.