H3 latent upscaling always results in OOM
retained text
I see lots of people saying how they love upscaling H3 latent and how effective it is, but I can't run it, no matter what.
I have a 16:9 video at 0.2MP (14 seconds) I'm trying to upscale. I've tried upscaling to multiple resolutions, with 0.98MP being the highest I've tried, down to 0.4MP.
I have 5060 TI 16GB VRAM, 48GB RAM, Ryzen 9900X3D.
I tried everything, it seems - using multiple attentions (comfy kitchen, SLA, VSA), H3 FeedForward, tried raising the reserve-vram flag to 3.5. I use obvpm's workflows (you can find them on GitHub) and tried reducing the length of the window (in seconds) {basically, the lower the length, the less VRAM it requires}. Tried using his Headroom node to give more VRAM headroom.
I'm using LBH-123 latent upscaling (with fp16 v1 3d model, something like that), I understand there are multiple upscalers out there but I haven't tried others yet.
I don't know what the problem is. Could it be a CUDA/pytorch version or something? Do I need to work on 9:16 videos instead?document fetched 2026-10-13 00:42 · canonical https://reddit.com/r/comfyui/comments/1x3st5n/h3_latent_upscaling_always_results_in_oom