the only thing i can say , LTX2.3 is like trash next to H3
thats amazing, 15sec video 4minute with 5090
very easy and accurate about prompting
also able to read the photo story board
very close to seedance 2.0
thank you
i feel so bad thrashing my beloved ltx, but yeah, this is far superior and times are worth it
but so slow
I feel one big problem LTX has is the cross temporal method. So when the model spent time going backwards, it made things weird and the physics was not progressive or correct. It's cool that LTX has competition so they can strive to improve themselves.
我下载了H3之后做的第一件事，就是把ltx2.3的模型、lora和节点全部删除了。它没有丝毫存在的必要，但是我依然留下了wan2.2作为补充。
thats amazing, 15sec video 4minute with 5090
very easy and accurate about prompting
also able to read the photo story board
very close to seedance 2.0
thank you
Can you tell me if this model can generate someone facing away with hands on knees? (LTX is pretty trash in this concept). Is it good with physics, like water and ect? It would be great know it, because I cannot test this model right now. I'm waiting for gguf stuff :/
thats amazing, 15sec video 4minute with 5090
very easy and accurate about prompting
also able to read the photo story board
very close to seedance 2.0
thank you
Which model Checkpoint are U using?
yes, there is the problem. 15 seconds. It is better than 5. But a lot less than the minute or so ltx can do easily.
I wonder if h3 can be stitched up with consistency as wan2.2 can be with things like svi pro. under certain conditions wan+svi it can do literally minutes with being consistent. Because i really don't feel like going back to 15 seconds again..
wan2.2 as video model of course also vastly superior (incl easy to use) to ltx2.3, but lacks the whole audio thing.