Spaces:
Running on Zero
Disconnects
This was a great space but lately there are so many disconnects or LoRA fails that it's nearly impossible to generate videos without constantly reloading or waiting or generating videos that either completely ignore prompts or are heavily artifacted. Kind of a pity.
The problem seems to be getting worse. Some LoRAs don't seem to be working at all. I've spend at least 30 Minutes of GPU quota trying different settings and strengths, but nothing works. Additionally the disconnects and "token errors" are getting more frequent. It takes way too much time, money, and effort to generate one unusable video.
Don't know, what happened to this space (most of dagloop5's spaces actually). They were near perfect a week ago.
From what I understand, some of the LoRAs are designed for the regular LTX 2.3 checkpoint while others are made specifically for Sulphur/10Eros, to which the current version of this space uses a version of the 10Eros checkpoint with a distilled LoRA integrated. I made this space and others for LTX 2.3 because my device doesn't have the specs to run LTX 2.3 via Comfyui, and they're pretty much for my personal use (hence the 'testing' name). I've seen other spaces for LTX 2.3 and its finetuned LoRAs/checkpoints use some sort of Comfyui backend with greater success, but I am not a coder, and have no idea how to tackle that. What I do know is that having adjustable LoRA strengths combined with checkpoint preloading has been trouble, and the strange cache system is the only thing that I have tried so far that comes close to working, as it creates a new transformer with the LoRA strengths infused on CPU (I think) separate from video generation in a manner that doesn't consume ZeroGPU. As far as the disconnects go, I have no idea what causes them other than its likely linked to the LoRA cache features; it wasn't really a problem until the way ZeroGPU spaces work on Huggingface was updated ~2 months ago. I might eventually try to create a space or two for LTX 2.3 based on a more recent commit of LTX-2 (all of ones since around March seem to lack the preloading components), but again I am not a coder, so if you have any recommendations lmk here.
Thanks for the answer. Well, same problem here: not a coder, at least not on that level. I'm trying to learn as much as I can from wherever, but doing this from scratch and with very limited time seems almost impossible.