Same here. V2 is giving me far better results for audio and video. Been playing with it for a few hours and it's not close to V2. Steps and everything else is adjusted for the 2nd pass and it's not close.

@Delavestra Just to confirm, do you get the ultra blurry stuff with the Final Video node? Giving far better result with v2 is an understatement since it makes it completely unwatchable.

I've managed to get decent results with V3, just not with the default setup.
Here's what I did if you wanna test it too.
1. Disable the Final Video node by selecting it and bypassing it (ctrl + B)
2. First Pass Preview node must have "Save Output" enabled
3. Use the Video Editor section to do an RTX pass Upscale and the result is similar to V2 (need to test more to see if it's actually better or worse).

EDIT:
It seems that the Toggle Disable Final Video works correctly on the second attempt (after I reloaded the defautl workflow and restarted, so no need to manually bypass the node). I still get a messy blurry video.

I am getting the same results. V2 results were great. But with V3, the final Render extremely blurry and distorted. I haven’t had too much time to test anything or change any other settings I was just using it default.

@nutmegzero439 It seems like it's a problem the author overlooked. You can follow my steps above to test without the final pass. Note that without the Final Video, your video will be low resolution. Put it as a second step in the Video Editor on the right side of the workflow and it looks good. Just extra steps tho. Will keep playing with V2 in the meantime until the author fix it.

@MrXin Same here, V3 gives a blurry mess and I"ve tweaked and dialed everything. still at 3 steps with OG settings on final pass. If I go to 1 pass at full size then the results are pretty good. What I really want is the RTX upscale and interpolation to all happen in the same flow. I was unable to get this to work, not due ot vram, but comfyUI's handling of videos. Super strange this is difficult to do. Is there any way to make it happen? I just added clear vram and cache, then did pull file name and selct file then upload from path, but that broke due to subnodegraph interactions that didn't always happen, really weird... Anyways, I'm trying to optimize for a 5090 setup so my wants are different, but hoping to get consistent results from either wf asap.

@MrXin V3 is a blurry mess too on final video. Width and height is 736x1280 (W x H)

@MrXin input resolution: WxH = 704x1280
Video length: I tested at 6s and 20s
Basemodel: ltx2310eros_beta.safetensors

Basically used everything as defaulted by your WF upon download. Changed the input, the positive prompt and the seed only. I tried without any LoRas as well. Hope this helps!

this is a final pass scheduling problem on the ltxvscheduler node on the final pass
the workflow default (steps 3, max_shift 4.00, base_shift 2.00, stretch true, terminal 0.00) is bonkers aggressive (resolves to sigmas of something like 1.00, 0.74, 0.0, 0.0 - yes the final denoise does nothing) i have no idea what its tuned for.
i tried the following settings and it eliminated the blur but doesn't preserve enough movement from the first pass:
steps 4, max_shift 1.0, base_shift 0.5, stretch false, terminal doesnt matter with stretch false
specifically this deviates a lot from the first pass in terms of movement but detail is good.
still figuring out some proper settings but its bedtime now, maybe @MrXin can shed more light on it all i'm just a noob
i'm testing some gentler settings (steps 4, max_shift 0.35, base_shift 0.15, stretch false) overnight and can report back in the morning if this isnt solved

@Fellaitio @Delavestra @Melty1989 fyi see above post

@MrXin My image size is 576 x 1024, I am not using the RTX upscale because I have an issue with the NVDIA Control Panel.

pytorch version: 2.9.1+cu130

Set vram state to: NORMAL_VRAM

Device: cuda:0 NVIDIA GeForce RTX 3080 : cudaMallocAsync

Using async weight offloading with 2 streams

Enabled pinned memory 14711.0

working around nvidia conv3d memory bug.

Using pytorch attention

Python version: 3.12.10 (tags/v3.12.10:0cc8128, Apr 8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]

@MrXin It might actually not be your fault, ComfyUI can be the one to blame since we all have different installations :)

I'll keep an eye out for your next versions! will happily try them. In the meantime, your v2 and another WF from someone else are really good enough!

Thanks for your time!

Thank you very much for the workflow, it is awesome ! About the blur problem guys, be sure to have the correct path for the ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors in the [Load distilled lora ] node and it will be ok. :)

thats a good point, @MrXin what version of comfy are you on? if you can give a specific commit thatd be great
i already had to patch a number of stability issues in the ltxv nodes locally, i think they're all still pretty early in development so its very likely theyll change internally

@kalifa7548 Care to confirm path? Mine is at ComfyUI_windows_portable\ComfyUI\models\loras\

i dug into this some more:
LTXVScheduler on the final pass might not be the most appropriate. it always starts with a sigma of 1.0, effectively throwing away the entire image from the first pass. the official lightricks workflow uses manualsigmas for the final pass, as this workflow's v2 did, but we observed issues with movement reduction in that case.
i think the answer here lies in more careful setting of manualsigmas values for the final pass. i'll experiment more with it today.
edit: just used manualsigmas with 0.85, 0.7250, 0.4219, 0.0 with some good results - motion appears to mirror the first pass but with improved detail. i'll see if i can figure out refining them

@seaborgiumthemad Doing good work friend! I saw this same node in the backend but I have not the slightest idea of what those numbers mean :D So I just disable it and I'm getting good result. Can't wait to try a second pass with my updated first pass :)

@Fellaitio if you've done any Img2Img work then its kind of like stepping down the denoising, e.g. doing a generation at 1.0 denoise strength (making a new image), then another one at a lower denoise str (keeping some small amount of the original image while introducing a bunch of new latent noise to solve), then another at an even lower str, etc., gradually reducing down until you get a completed refined image.
you want to use the big aggressive 1.0 str (i.e. from the ltxvscheduler node) for the first pass so it can generate fresh content, then use refinement (i.e. not 1.0) oriented denoising for the final pass.
try making a manualsigmas node and put those values from my previous post in and connect the sigmas pin to the sigmas input on the sampler final pass node (where the ltxvscheduler was previously connected - make sure you do this on the final pass, not the first pass) and let me know how it goes.

the real mystery is why the ltxvscheduler on the final pass seems to work fine for some people but not for others. it doesn't look like there were any significant changes to the node's code in comfyui since the node was added, i doubt it's hardware/driver related but i really don't know

@seaborgiumthemad Damn, that's a lot of great info... for someone more knowledgeable at this than me :D

@seaborgiumthemad Hi there, thx for info on Sigmas. I am getting very good results on V3 with 1 pass at 10 steps and high starting dims and no downsize in the video process. My overall motion is still a bit constricted, trying to get more out b/c I know it can do more re motion and 'loosening up' what's happening. I'm wondering if you could comment on the Sigmas node in V3 1st pass (my only pass) and advise if the max and base shifts can be played with, and how, and if the terminal value matters or does anything. I'm playing with them but can't really tell, and AI grok/chatgpt don't really know. My testing is unclear, and I just want to allow more motion and expression from the loras and prompts... Lastly, I use 2 chunks at 4096 at the beginning after the dynamic_fro09 lora (which I believe allows lower steps if I'm not mistaken), and someone said to lower the value on this to 0.1 or 0.2 ranther than 0.6... My tests again are a bit inconclusive, and ai doesn't know. Any ideas?

@Delavestra can you share your settings for that node? the first pass uses ltxvscheduler by default which with max_shift 2.05 and base_shift 2.0 would be weighting the sigmas heavily towards the higher end.
for example, i'm using 12 steps, at 2.05/2.00 with stretch=true and terminal=0.1, the following sigmas are generated: 1.0000, 0.9834, 0.9640, 0.9409, 0.9130, 0.8784, 0.8348, 0.7778, 0.7001,0.5883, 0.4132, 0.1000, 0.0000 (i use Display Any from rgthree to view these, just connect the sigmas output to it)
the higher ranges (~0.85+) dominate this and i get some decent movement out of it.
try plugging a Display Any node into the Sigmas output and test some different values, get a sense of which ranges mainly control which aspects (e.g. higher ranges tend to be movement, mid ranges tend to be rough detail, low ranges tend to be fine detail - as an extremely rough guide)

@MrXin no worries. i found a nice node in the kjnodes pack called VisualizeSigmasKJ which makes a nice graph, i'm still trying to figure out more myself. putting together a new workflow that uses only manualsigmas for testing

@seaborgiumthemad Manual sigmas works! Fixes the blurriness, thank you. motion also seems to be decent

@seaborgiumthemad o cool thx, my current settings are: max shift 3.5, base shift 2.5, and terminal 0.03. (Grok suggested, this is with me not understanding what these values do). Do you recommend I change these values, move to the manual sigmas node, or something else? Lastly, I have a 5090, is there a better/different WF I should use for better results? Or is this one still solid? I'm seeing some gens, like the one on the orgasm lora, that are way more animated, higher quality, and really crisp compared to anything I've gotten out of this wf so far (the only one I've tried). Wondering if it's a model issue or NSFW loras issue... thx!

You can add this in yourself pretty easily. Just add in the node to set the w/h or length of a short or long side.

try nvidia rtx super resolution node at 1.25 t 1.5 scale at ultra quality. It's blazing fast.

@FantasticFever problem with teeth

@Axsee yep, can be an issue, teeth have to be almost perfect before 

@FantasticFever When i enable this the resulting video looks no different - im not getting any errors so not sure if its working or not?

Run this on the first layer of you ComfyUI folder:
.\python_embeded\python.exe -m pip install -U --no-build-isolation nvidia-vfx --index-url https://pypi.nvidia.com

@MrXin Yes ty, playing with 1 pass and it's starting to look great.

@MrXin Ah, ok. Then I will wait.

Thanks! I guess many people will use only the first pass flow until they read this!

I must be blind, I can't find the node your talking about to bump it up from 3 steps.

@sixchin You need to click the Processing Video node in the upper right to get into the inner bits, you'll find the Final Pass Sampler (wrote it backwards earlier xD) and the LTXVScheduler in there.

@MrXin cant wait to see your magic