calcuis

/

ltxv-gguf

like 44

did you fix the seed (under sampler)? what is your prompt? ok, seems the same structure

Yes I have give you a simple workflow, you can see the difference.

thanks; according to the console statistics, the new set is faster than before; will look into the source file, see anything mess up

If you can't test my workflow here a quick video exemple.
https://www.youtube.com/watch?v=jCnH3kl1Sic

thanks for the video review and output comparison; just re-converted the original safetensors, the recent general.file_type is [1], but it should be [32] (seems irrelevant; since 1=F16; 32=BF16; which is no problem)

Good, thank you!

If any users are reading this message, I’ve made some updates to my workflow ltx model test.
This workflow tests both models 0.9.0 and 0.9.1, includes an experimental I2V method, and provides an example demonstrating how to extend the T2V mode video and how to faceswap.
If it’s helpful to you, that makes me happy.

Updated workflow with both VAEs 0.9.0 and 0.9.1:
https://www.mediafire.com/file/n56u2qyyzut1vxk/Dave_Gravel_LTX_TEST_PLUS_LONG_VIDEO_EXEMPLE_V3.zip

Edited:
This video show how to use my workflow test and how to faceswap with a custom face.
https://www.youtube.com/watch?v=9cfB9sQWLfg

seems need a different architecture vae to make the 0.9.1 works differently

Yes, I see now. I have updated my workflow to test both models with both VAEs and added a method to save the VAE from the original model. It seems to be working now. Thanks a lot!

Edited:
This video show how to use my workflow test and how to faceswap with a custom face.
https://www.youtube.com/watch?v=9cfB9sQWLfg

Version: 0.9.0 + vae 0.9.0

Version: 0.9.1 + vae 0.9.1

Good, thank you!

If any users are reading this message, I’ve made some updates to my workflow ltx model test.
This workflow tests both models 0.9.0 and 0.9.1, includes an experimental I2V method, and provides an example demonstrating how to extend the T2V mode video and how to faceswap.
If it’s helpful to you, that makes me happy.

Updated workflow with both VAEs 0.9.0 and 0.9.1:
https://www.mediafire.com/file/n56u2qyyzut1vxk/Dave_Gravel_LTX_TEST_PLUS_LONG_VIDEO_EXEMPLE_V3.zip

Edited:
This video show how to use my workflow test and how to faceswap with a custom face.
https://www.youtube.com/watch?v=9cfB9sQWLfg

Could you upload the VAE from version 0.9.1 please? The one in your workflow is "LTX_VAE_0.9.1_FBFLOAT16". The one posted everywhere from version 0.9 does not give good result with 0.9.1 GGUF model.

Hi PixelPlayer,
I don't really have enough space to upload a file of this size.
And it's quite simple to obtain on your own. You need to download the 0.9.1 model in its original version, place it in the 'checkpoint' folder of ComfyUI, load the model into a regular checkpoint node in ComfyUI.
At the VAE pin, you need to connect a 'save VAE' node, and this will save the model's VAE into a folder in ComfyUI, which you can then use with GGUF.

I created a special sampler node for the LTX model and noticed that the VAE doesn't seem to work well with the new updates to the tiled VAE decode node. This appears to cause several glitches or artifacts in the video.
Alternatively, the default settings of the tiled decode node might not be suitable for the LTX model.
When I use a simple decode, I experience almost no glitches or artifacts.

thanks @Evados , just tested; it works very good; taken couple seconds only to generate a short (less than one minute) with the q4_0; great!

hi @PixelPlayer , you should find the new vae here and the workflow; credit should be given to @Evados ; Merry Christmas and Happy New Year 2025 (is coming)!

It makes me happy! Happy holidays and a Happy New Year to you too. Thank you as well for the GGUF.

Thanks guys! @Evados for the explanation, @calcuis for the extraction. Happy New Year!

If you liked my first workflow, here's another version with a simple custom node I created to include a sampler refiner.
You'll find the workflow and the custom node in the video description.
Have fun!
https://www.youtube.com/watch?v=hGLNOXOypRo

I downloaded ltx-video-2b-v0.9.1-bf16.gguf this week and had same results output as ltx-video-2b-v0.9-f16.gguf. For 0.9 I used the vae downloaded from here, for 0.9.1 test I have the full model already, so I used the vae from ltx-video-2b-v0.9.1.safetensors just to be sure, same result. If I tested same WF with ltx-video-2b-v0.9.1.safetensors (full) then results changed.

I downloaded the full ltx-video-2b-v0.9.safetensors and can confirm a video made with this or ltx-video-2b-v0.9-f16.gguf are the same BUT 0.9.1 or at least the bf16 (I could see no f16 for 0.91)I will try one of the other maybe the q8 or or q4 they seem to be 1 day newer so maybe?

@Evados nice WF, a lot going on, but great. Not sure if you have one that was designed for just 1 model (no compare) think with the color coding I was able to strip it back to do just one. Also did you see any differences in the o.9 and 0.9.1 gguf?

Hi, In my last video showcased this workflow implemented for just one model. If you want to ensure you have the correct VAE, the best approach is to save it yourself from the original model. ComfyUI provides node tools for this. The difference between versions 0.9.0 and 0.9.1 is not always significant—sometimes, it’s just slight color variations, minor details, or occasionally differences in animation movement.

The VAE was not an issue it was the model 0.9.1 gave same output as 0.9.0 proved it should differ using the non gguf models and knowing how it should look. Not checked to see if any fixes since.

modified the original safetensors and quantized gguf again; uploading a new version; test the revised one; see any difference(s); thanks

I tested, output is a little different than the previous release but still is not the same output for same seed as the safetensor (I already have 0.9.1 from https://huggingface.co./Lightricks/LTX-Video/) but it still, well at least in my tests (tested several img2vid) is a lot closer to 0.9 outputs and not 0.9.1.
As a test I ran the revised Q8 paired with a 0.9 vae and then it gets almost identical to 0.9. safetensor
As a control I reversed the test to check the vae influence and ran 0.9.1 safetensor with a 0.9 vae and the output was still very close to the 0.9.1 (safe and vae) so the 0.9.1 gguf seem to have a much closer similarity to 0.9 than 0.9.1 for some reason., at least in my testing.

what did you test actually? please provide more details; as it might need to code it at the upper level for adjustment; btw, have you tested it like this:

or you could even disconnect the t5xxl and connect the original checkpoint clip dot; in that case, you might be able to find out the true difference is coming from the model itself or the clip or the vae or both, etc.

I will try and check it again this week. I just ran a test video with 0.9 normal and 0.9 gguf and they were almost identical but when I try the same with 0.9.1 normal and gguf it was not the same
Can somebody else check maybe my testing is off somehow.

we tested the new architecture; the difference was significant; you could test these two as well: based on ltxv0.9 here and based on ltxv0.9.1 here

what did you test actually? please provide more details; as it might need to code it at the upper level for adjustment; btw, have you tested it like this:

or you could even disconnect the t5xxl and connect the original checkpoint clip dot; in that case, you might be able to find out the true difference is coming from the model itself or the clip or the vae or both, etc.

I had checked this it was very similar way. I had an existing workflow it was an image to video with various injections and I used the checkpoint loader of 0.9.1 to load the vae as in that image you posted. I had done the same test previously with my 0.9.0 safetensor (full) and the gguff and the output was almost identical but with 0.9.1 it was not. I had been testing *8 but I deleted it so I only have the Q4 so here is that test.

With that simpler workflow I ran the test and the pixels are similar, not exact but close and the color is different (full was bright lush green, gguf was washed out more, maybe the way gguf works with a full vae?) .

MOre testing showed yes different, exact WF node by node posted here.

I revisited Image to video and sure enough here they do not match. Used the basic old I2V and used same settings

just tested; you could try these two; the results/outputs are not really identical video-0.9.0 and video-0.9.1; should match with the proper vae; btw, the old vae is not suitable for the 0.9.5 version; can use still but the output you won't be satisfied

Thank you for the added info, I was hoping it was just my method, maybe my GPU makes the difference more exaggerated I know each is different.

This was last I tried:

I have 12gb of VRAM so I should not need the PIG ones and don't they need extra nodes? Are they more reliable for consistency?

It is because 0.9.5 I wanted to get this test done again as I will archive 0.9.1 now and move to 0.9.5 checkpoint. I will probably still test the gguf for that when it comes. I am wondering if maybe the simplicity of the first version meant quantized versions were not as far different in results so not visible at a glance.