LTX 2.5 I2V basic workflow on ComfyUI is stuck forever if try to generate 10sec video
Hi, I just tried LTX 2.5 I2V, works great and fast for 5sec vids (like 2mn30 or so without Sage Attention), I have a 3090, 128Gb RAM, 150GB disk space but if I try 10sec (in medium res) it gets stuck forever in the decoding phase just before the video appears. On the console it's exactly like a successful 5 sec video, the only thing missing is the line "Prompt generated in 254 seconds" that you see in green.
Am I alone ? Any idea about what the issue might be ?
Go into the subgraph and change the VAE Decode (Tiled) node's tile size from 768 to 512.
I hope you realize that you only have 3090, it's not the best one to test the new video model.
That is the reason you get stuck.
I had the same issue, but with a 5090 32GB + 192GB DDR5 RAM—meaning the LTX 2.5 would freeze while decoding VAE in ComfyUI.
Changing 768 to 512 solved the problem, but it's quite odd—the entire generation process before decoding VAE takes 1.5 minutes on my card, but then the actual decoding takes another 1.5 minutes (with SageAttention enabled).
I'm using FLF2V, not Image to Video, at a resolution of 720x1280 with a video length of 15 seconds.
Go into the subgraph and change the VAE Decode (Tiled) node's tile size from 768 to 512.
Thank you, it works.
I hope you realize that you only have 3090, it's not the best one to test the new video model.
That is the reason you get stuck.
I do know that it's not the best, but DAK25 has a 5090 which is also not the best but much better than a 3090 and he had the same issue.
Mine got stuck too. I have a 5070ti but I adjusted these and it goes pretty fast. I fixed it by changing:
temporal_size: 64
temporal_overlap: 16
I didn't do the 512, but it works nonetheless.
Currently we have 2 decoders:
The convolution decoder is the one you all know from LTX-2.3 — it works well and produces decent results.
The diffusion decoder is completely new and produces much higher quality output than the convolution one. That quality comes at a cost in performance and VRAM: it operates much closer to pixel resolution than the convolution decoder does, so it has to process roughly 500x more tokens for a 1080p, 121-frame video. If you're running the distilled pipeline at 1080p, the diffusion decoder will end up dominating your total generation time.
Adjusting tiling is a good way to make the diffusion decoder workable on consumer hardware. As a general rule: bigger tiles decode faster but carry more OOM risk; smaller tiles are safer on VRAM but slower, since more of the compute goes into the overlap/halo regions between tiles rather than final output.
We're continuing to optimize the diffusion decoder's performance, so stay tuned for updates.
The diffusion decoder is completely new and produces much higher quality output than the convolution one. That quality comes at a cost in performance and VRAM: it operates much closer to pixel resolution than the convolution decoder does, so it has to process roughly 500x more tokens for a 1080p, 121-frame video. If you're running the distilled pipeline at 1080p, the diffusion decoder will end up dominating your total generation time.
Adjusting tiling is a good way to make the diffusion decoder workable on consumer hardware. As a general rule: bigger tiles decode faster but carry more OOM risk; smaller tiles are safer on VRAM but slower, since more of the compute goes into the overlap/halo regions between tiles rather than final output.
We're continuing to optimize the diffusion decoder's performance, so stay tuned for updates.
Have you tested int8convrot of the diffusion VAE decoder? I've been wondering if it would reduce memory requirements and speed things up. Maybe the impact on image quality would be too drastic, dunno.
I've been trying to find a int8convrot quant of the model from the HF community but it doesn't seem anyone has made one yet.
Both dev and distilled int8_convrot versions of the model are here: https://huggingface.co/Lightricks/LTX-2.5/tree/main/diffusion_models.
Both dev and distilled int8_convrot versions of the model are here: https://huggingface.co/Lightricks/LTX-2.5/tree/main/diffusion_models.
I meant the VAE model. They are BF16 only.
That is correct.
At this point we didn't release any official int8_convrot checkpoints of the diffusion VAE. We are exploring all the options to optimize the VAE with minimal quality loss.