Kijai

/

LTXV2_comfy

like 391

I am using the updated VAE but, i am getting black screen only. it was working with previous VAE.

I am using the updated VAE but, i am getting black screen only. it was working with previous VAE.

did you update the KJNodes? It requires latest version of those nodes to work correctly

the new VAE version tends to increase contrast compared to the old one. similar to adding the detailer LoRA, it cannot retain the ref image's contrast.

to kJ: could you please make the old LTX2_video_vae_bf16.safetensors version available again as well? I replaced it with the new version but I'd like to retain the old version as well. for photorealistic content.

old vae (retains original contrast/saturation):
https://drive.google.com/file/d/1BlPbJxQRyLZuq4MO0SmZOnzHMqQ9oaOi/view?usp=sharing

new vae:
https://drive.google.com/file/d/1pPQKql77ST9zNFNZ_Majv8IL9qKBPEgw/view?usp=sharing
the wall is now with greenish tint. the character has oversaturated highlights/etc
the LTX-2 itself tends to create oversaturated shiny plasticky stuff (see the OP video). and using new vae only worsens this effect. it screams "I'm generated with AI". not to my liking. :)

I can fix oversaturation with post processing but I'd like to have an option to retain original image style as is by using BOTH new and old vae. :)

I am using the updated VAE but, i am getting black screen only. it was working with previous VAE.

did you update the KJNodes? It require latest version of those nodes to work correctly

sorry, i didn't. i've updated KJ Nodes, now its working.

with new VAE, close-up shots only looking good, in medium shots face got ghost effect, no detail even with detailer lora.

the new VAE version tends to increase contrast compared to the old one. similar to adding the detailer LoRA, it cannot retain the ref image's contrast.

Looked quite the same as ref image to me, but i only tried a few

(that being said, its not an objection to using the old vae ;-) but just to say had very nice results with the new)

using VAELoaderKJ in KJNodes version 1.2.5 or later for video vae, the colors are better.
but slightly saturated compared to the ref image still.

new vae + KJ loader:
https://drive.google.com/file/d/10xSV7JiUd1qHjvF8dZ_jQxSapPosj3an/view?usp=drive_link

the contrast/color difference is very subtle but its there. just saying.
overall, tradeoff between details increase and some saturation drift from the ref is acceptable
with new loader there is no need to have 2 vaes.

Note that if you use a VAE loader that does not load metadata, it does turn the results greenish and introduces artifacts, or it can completely break too, this is what I get on this image:

LTX-2 is a huge leap forward concerning facial audio+image to video generation and performance. its the first time I'm able to generate videos with adequate performance on my RT 5060 ti 16GB.
thx to all the LTX-2 developers and Kijaj's contribution :)

LTX-2 has huge potential, I hope we'll get SCAIL support for LTX-2 one day

LTX-2 is a huge leap forward

echoing that. I can generate HD videos right out of the box, something I had no chance at prior, and at quite a fast speed. PLUS audio ;-)
I thought this would be a model i would have to render at low res, and wait for 1 hour for each video. Thats not the case lol

(and havent even started with trying own audio inputs, pose and depth maps, first last frame etc.. its a bit of a jack of all trades seemingly)

(still love my Wan too of course)

Btw, 'your' old vae is available it seems. More choices ;-)
https://huggingface.co/Kijai/LTXV2_comfy/tree/main/VAE

at the moment, pose control LoRA does not look great compared to Wan SCAIL or even 2D-bones Wan Animate.

Btw, 'your' old vae is available it seems. More choices ;-)

Yeah I've redownloaded it just in case.

i echo the speed for the length and resolution. finally out of 120 frames territory without tricks and at lightining speed. That said, beyond a single person doing simple stuff (all examples above), that speed seems to come at a hefty price. It doesn't seem to have the faintest idea of physics and how things should interact with each other when more is happening. that's why it needs extreme miromanagy prompts. In a way you replace the workflow tricks for longer videos (wan remix+svi pro=amazing!) with prompt "tricks". for now.

I updated the KJNodes, but the new VAE doesn't work for me. However, everything works with the old VAE.

I updated the KJNodes, but the new VAE doesn't work for me. However, everything works with the old VAE.

+1
I use distilled Q4 model and distilled embedding connector in the workflow. using old vae was fine (not best) but trying this new vae resulting patterned noise like yours

can we have fp8 version of vae, please?

I updated the KJNodes, but the new VAE doesn't work for me. However, everything works with the old VAE.

Your update has probably failed as that's the kind of results you'd get when it's not updated.

I updated the KJNodes, but the new VAE doesn't work for me. However, everything works with the old VAE.

Your update has probably failed as that's the kind of results you'd get when it's not updated.

I got the same result, with a lot of patterned noise

I updated the KJNodes, but the new VAE doesn't work for me. However, everything works with the old VAE.

Your update has probably failed as that's the kind of results you'd get when it's not updated.
Thanks, it's working now. There was a problem with my workflow.

I saw lightricks updated their model today
https://huggingface.co/Lightricks/LTX-2/blob/main/ltx-2-19b-distilled.safetensors

Not sure if that matters, maybe Kijai already extracted the good vae and their update was just same fix

I saw lightricks updated their model today
https://huggingface.co/Lightricks/LTX-2/blob/main/ltx-2-19b-distilled.safetensors

Not sure if that matters, maybe Kijai already extracted the good vae and their update was just same fix

They just replaced the broken VAE in the distilled models with the correct one. The dev VAE was always correct, which is the one extracted here currently.

I updated the KJNodes, but the new VAE doesn't work for me. However, everything works with the old VAE.

Your update has probably failed as that's the kind of results you'd get when it's not updated.

I got the same result, with a lot of patterned noise

After the correct update of KJNodes, use VAELoaderKJ to load the new VAE, don't use the usual VAE loader node, this fix my pattern noise issue