Please make ltx2.5 10eros，eager for it
please
Maximsobolev275 has tested it with LTX 2.5 and he reports that it does work quite well.
Maximsobolev275 has tested it with LTX 2.5 and he reports that it does work quite well.
Distill and DMD LoRAs work as well, and honestly, I like the results even better than with the official 2.5 distill LoRA.
If you use all those old tools you still just have a lesser version of the 2.3 model that's a bit faster. There was something about comfyui not yet supporting the correct implementation of half of the model's experts with DFR. But anyways, the old data doesn't carry forward well and Sulphur needs to be retrained on it to use all of the new features. With a full retrain and DFR integrated it'd be useful, but I don't control the training.
I have a temp one up and running: https://huggingface.co/spaces/ibyteohdear/LTX-2.5-Multi-Pipe-10Eros
I've got a decent LTX2.5 pathway in local comfy as well using extracted LORAs from 10Eros 1.2 and 1.5, but using all the proper LTX2.5 tools
Results are decent enough, posted some examples on Sulphur discord. Overall quality is significantly better than 2.3 (physics, prompt adherence, guidance...), but as TenStrip said, the NSFW part is still subpar until a proper retraining of the Sulphur base is done over the actual 2.5 transformer.
I'm also waiting eagerly to this. LTX2.5 is superior in every way to 2.3, WHILE also being faster, quite a feat if you ask me. And even if it's not as revolutionary as H3 has been, it's much more resource and time friendly on average-level local machines.
I've got a decent LTX2.5 pathway in local comfy as well using extracted LORAs from 10Eros 1.2 and 1.5, but using all the proper LTX2.5 tools
Results are decent enough, posted some examples on Sulphur discord. Overall quality is significantly better than 2.3 (physics, prompt adherence, guidance...), but as TenStrip said, the NSFW part is still subpar until a proper retraining of the Sulphur base is done over the actual 2.5 transformer.
I'm also waiting eagerly to this. LTX2.5 is superior in every way to 2.3, WHILE also being faster, quite a feat if you ask me. And even if it's not as revolutionary as H3 has been, it's much more resource and time friendly on average-level local machines.
That'll change when Sparse Attention gets debugged. But, time will only tell.
How to prompt ltx2 so it actually does what's described in the prompt instead of saying it out loud? I feel like it's mocking me every time.