Weird Pattern Shift
Same issue here.
Did you update the spatial latent_upscaler to 1.1?
The speed gain is great but the quality loss is not worth it... Yes Im using the upscaler 1.1.
Did you update the spatial latent_upscaler to 1.1?
I just gave it a go too and SAME issue. Went from 55 secs for a 1280x720 gen to 45 secs, which is a nice time cut BUT... hope they fix the quality loss ;)
Same issue here.
I get the same behavior on very long videos (+30s) with the FP8 model. I have tried to solve it with more steps like I have done with WAN2.2 videos when I get odd artifacting, but at least in my simple testing that hasn't seemed to help.
Other than the quality loss, which I have too, in many prompts the outputs are in slow motion, on FP8 sometimes there are head spinning issues but NEVER slow motion... I do not have on the FP8 model generating 20secs in 1280x720.
I really like LTX2.3, being able to gen in 720p, 20secs and with audio is great I really do not want the slow motion problem from Wan 2.2 on LTX2.3. I hope it can be fixed otherwise in my opinion is worth waiting more time for each generation.
Did you update the spatial latent_upscaler to 1.1?
Yeah,.. but I think the issue is probably relative to me just prematurely trying to run NVFP4 with the same setup as the dev and fp8 models. There's probably other loras and encoders that have to run along with it. IDK... I'm slowly learning how to use all of this... painfully slowly.
There is an updated version of the NVFP4 model.. It mostly fixed the image quality issue. But I can't seem to get any significant render speed increase out of it vs. the FP8 model.
I have the same issue in comfyUI on a 5090. It's actually slower than FP8
I have the same issue in comfyUI on a 5090. It's actually slower than FP8
it is faster than FP8... the quality loss is the issue.
Gen time decreased about 4 to 6mins on a 5060ti and 5080
for anyone curious, coming from some one tweaking a 13gb fp4 (yes it works), this issue, with the cloudy/smokey shift, is from the lora injection. my 13gb distilled is stable and amazing until you add a lora.
so light tricks, if you have a thought on that, please let me know.
I only know 2 fixes atm. 1 is run all stages at 9+ steps. or add 0.01 noise to the latent after 3-4 step upscale pass and then single low sigma sample with no noise during sampling at the end.
its a noise issue, once you add a lora, the noise doesn't sample out in under 9 steps without some tricks.
examples of my 13gb distilled, no lora. all are 8 step first stage, 3 step upscale.
https://civitai.com/images/124089918
https://civitai.com/images/123710121
https://civitai.com/images/123710106
the model if you wanted to try it, again, doesn't work with lora's without heavy tweaking (also, no vae/text projection)
https://civitai.com/models/2445970?modelVersionId=2751189
Edit: Also, i do this all on a 4090, hints the need for a small efficient fp4.
I can't believe I spent a day using Gemini to get all this working on my 5090 just to have to be usable. Has anyone gotten Dev only to work without a Distilled LoRA at some percentage? You can run 50 steps and it still poor quality.
I'm trying to figure out what this means, which could be a fix or a lead.
https://github.com/Comfy-Org/ComfyUI/pull/12978
for anyone curious, coming from some one tweaking a 13gb fp4 (yes it works), this issue, with the cloudy/smokey shift, is from the lora injection. my 13gb distilled is stable and amazing until you add a lora.
so light tricks, if you have a thought on that, please let me know.
I only know 2 fixes atm. 1 is run all stages at 9+ steps. or add 0.01 noise to the latent after 3-4 step upscale pass and then single low sigma sample with no noise during sampling at the end.
its a noise issue, once you add a lora, the noise doesn't sample out in under 9 steps without some tricks.examples of my 13gb distilled, no lora. all are 8 step first stage, 3 step upscale.
https://civitai.com/images/124089918
https://civitai.com/images/123710121
https://civitai.com/images/123710106
the model if you wanted to try it, again, doesn't work with lora's without heavy tweaking (also, no vae/text projection)
https://civitai.com/models/2445970?modelVersionId=2751189Edit: Also, i do this all on a 4090, hints the need for a small efficient fp4.
Is it possible you're mixing up FP4 with NVFP4 models? From what I'm learning the NVFP4 have similar performance as BF16 with less VRAM and increased speed on Blackwells 50xx cards.
And regular FP4 is a further quantized version with only vram savings, at the cost of performance? I think the issue is with the NV versions not working with the LoRAs.
I'm happy to be wrong! Still trying to figure this out and this is where I'm at.
for anyone curious, coming from some one tweaking a 13gb fp4 (yes it works), this issue, with the cloudy/smokey shift, is from the lora injection. my 13gb distilled is stable and amazing until you add a lora.
so light tricks, if you have a thought on that, please let me know.
I only know 2 fixes atm. 1 is run all stages at 9+ steps. or add 0.01 noise to the latent after 3-4 step upscale pass and then single low sigma sample with no noise during sampling at the end.
its a noise issue, once you add a lora, the noise doesn't sample out in under 9 steps without some tricks.examples of my 13gb distilled, no lora. all are 8 step first stage, 3 step upscale.
https://civitai.com/images/124089918
https://civitai.com/images/123710121
https://civitai.com/images/123710106
the model if you wanted to try it, again, doesn't work with lora's without heavy tweaking (also, no vae/text projection)
https://civitai.com/models/2445970?modelVersionId=2751189Edit: Also, i do this all on a 4090, hints the need for a small efficient fp4.
Is it possible you're mixing up FP4 with NVFP4 models? From what I'm learning the NVFP4 have similar performance as BF16 with less VRAM and increased speed on Blackwells 50xx cards.
And regular FP4 is a further quantized version with only vram savings, at the cost of performance? I think the issue is with the NV versions not working with the LoRAs.
I'm happy to be wrong! Still trying to figure this out and this is where I'm at.
That is a good question, these are my first FP4's, but, I did build them with comfy kitchen, which says its all NVFP4, so your quess is as good as mine, but considering my output was identical to his video above, they are at least operating close to the same. and since lightricks released the dev model first, assuming these are generations with the distilled lora?
I don't have a blackwell so can't test performance difference, but on my 4090, my fp4's run same speed as the official nvfp4, mine just use a lot less vram.
I did fix my issue though, not at the model level because i couldn't and keep it at 13gb on gpu, but at the lora loading level, that noise is all low block, so i built a ramp into my lora loader, left block 0 at full since that is bf16 in my model, then block 1 at 0 and ramp up to block 10 at full, problem fixed.
come to find out, it also REALY hated the mixed at block 1 (fp8). yes, i went bf16 > fp8 > fp4, 46gb model stable at 13gb was not easy.
and like i said, my models run awesome with no lora, it was only lora that was the issue.
i did upload them here.
https://huggingface.co/MrReclusive/LTX-2.3-FP4/tree/main
ill add my lora loader soon