Kijai

/

LTXV2_comfy

like 230

You doing something wrong, in my tests some LTX look the same or even better

I am also struggeling with the image quality, everthing looks kinda blurry and the movement is soft drawn without texture.

Nah man, the image quality isn’t great. Even in the official website examples—especially around the mouth area—it comes out blurry. The motion is good, though. Overall, the image feels overly smoothed out. WAN still looks superior. I’m using both the official workflow and Kijai’s workflow, and the results are the same. I even tried the BF16 model.

So stick to Wan model, since you found this model blured and weak image quality to your eyes, no sense to say something it cant be changed ;)

@Aorora
in my tests the quality depends on:

• image size (bigger = better)
• sampler/scheduler try exp_heun_2_x0 with beta57 sigma
• frame rate (48 gives better quality)
  if you can handle it memory wise.

@erosdiffusion Thanks, bro. This is exactly the comment I was looking for. If this works, I’d really appreciate it.”

@erosdiffusion Are you using the base model or the distilled one? If it’s the base model, could you share the steps and CFG settings?

how do I set beta57 sigma?

i'm using distilled fp8 as i have memory constraints and the separate vae, text encoder from kijai.
the workflow i use is a modified version of the one from ltx nodes
beta57 comes from res4lyf package, the sigma curve is similar to he manual one that is in the flow but a bit more S like towards teh end. if find it gives a good balance between structure, coherence and details.
notice the exp_heun_2_x0 is twice slower than euler (if you want to spend the time)

dev + distilled lora seems slightly better quality but slower

Could also do a bit of a "2nd pass" with Kijai's implementation of FlashVSR that both upscale and refine (add details). And its ultra fast at doing so. ( 1 step sampler) (part of WanVideoWrapper nodes)

LTX-2 natively is capable of 4K i think, but most of us probably render video at a quite low resolution, so having a 2nd pass might not be a bad solution, if you cant use the LTX-2 latent upscaler to HD or more.

An example of a 720p LTX-2 with a 2nd pass with FlashVSR

(much sharper appearance)

Could also do a bit of a "2nd pass" with Kijai's implementation of FlashVSR that both upscale and refine (add details). And its ultra fast at doing so. ( 1 step sampler) (part of WanVideoWrapper nodes)

LTX-2 natively is capable of 4K i think, but most of us probably render video at a quite low resolution, so having a 2nd pass might not be a bad solution, if you cant use the LTX-2 latent upscaler to HD or more.

An example of a 720p LTX-2 with a 2nd pass with FlashVSR

(much sharper appearance)

Can you send me your revised WF, please?

I take back what I said. This is my best result so far, this model is really the best. I’m using Q8 GGUF at 720p, and it already looks great.

i'm using distilled fp8 as i have memory constraints and the separate vae, text encoder from kijai.
the workflow i use is a modified version of the one from ltx nodes
beta57 comes from res4lyf package, the sigma curve is similar to he manual one that is in the flow but a bit more S like towards teh end. if find it gives a good balance between structure, coherence and details.
notice the exp_heun_2_x0 is twice slower than euler (if you want to spend the time)

How should I set the noodles and how much denoise on the second pass? Naturally my first run resulted in just bad noise (both exp_heun and euler).
I tried to insert the basic scheduler but even with only in the second pass the result remains the same. Care to share your workflow?

Why 0.9 denoise ?
Set both times to 1. The wf is more or less the same from ltx nodes. I have just changed the loaders to use split files from kj and the dual clip loader as in kj screenshot. (Not at pc atm)

Also been updating to use the split files and got them working but now something seriously broke as I can't get it render even with original manual sigmas. And I had it first on full denoise and thought SDXL i2i logic that perhaps I shouldn't denoise the second pass fully but I guess this is a bit different.

Not sure why this started to happen, likely not related to sampler/scheduler as even old generations produce now noise.

I was trying to get Gemma3 GGUFs working by updating the nodes and loader.py in the GGUF custom node by City96 and I broke my comfy...

EDIT: got my comfyUI back, needed to merge the GGUF repo correctly via git and not just manually copy two files. Also basic scheduler and Exp_Heun works, now to some gens to see what works well