RuneXX

/

LTX-2.3-Workflows

like 313

@runedo you have DEV workflows already set?

If you're seeing jagged pixel-like edges, try changing the Upscale Method from 'Nearest-Exact' to 'Lanczos', this made my renders much better and edges look smoother. If you're seeing bright artifacts or titles at the end of the videos, that's a separate issue and not related to the workflows, You can check the discussions here https://huggingface.co/Lightricks/LTX-2.3/discussions/13

DO i need to change the WF settings on PHASE 2 if i am using the DEV model? like MORE STEPS for upscale? CFG? image strength?

PHASE 2 is always same, both for distilled and dev model workflows. This part only does the upscale part

If you use DEV model you need to enable the Distilled lora right below where you load the Dev model.
This makes it possible to run Dev model with low steps. The workflows are all made to run fast with low steps.

Alternatively if you want to run with more steps and cfg higher, you would need to change more but in PHASE 1:

• at the first sampler pass group (phase 1), connect the LTX Sampler to the sigma input at the sampler
• set steps to 20++ at this LTX Sampler at first pass group
• Set CFG to 3-5 in the first pass group
• Enable the distiller lora

The last part, the distilled lora is perhaps the most tricky part. It probably runs fine with the 0.6 strength set as default for both first and second pass (but it could in theory be far less for first pass, and 0.6 in last pass)

But if you wait a little bit I will upload a workflow soon, that runs Dev model for a more slow render with full steps, that can give better results

If you're seeing jagged pixel-like edges, try changing the Upscale Method from 'Nearest-Exact' to 'Lanczos',

Nice tip, will try that here. See if it improves things ;-)

DO i need to change the WF settings on PHASE 2 if i am using the DEV model? like MORE STEPS for upscale? CFG? image strength?

PHASE 2 is always same, both for distilled and dev model workflows. This part only does the upscale part

If you use DEV model you need to enable the Distilled lora right below where you load the Dev model.
This makes it possible to run Dev model with low steps. The workflows are all made to run fast with low steps.

Alternatively if you want to run with more steps and cfg higher, you would need to change more but in PHASE 1:

• at the first sampler pass group (phase 1), connect the LTX Sampler to the sigma input at the sampler
• set steps to 20++ at this LTX Sampler at first pass group
• Set CFG to 3-5 in the first pass group
• Enable the distiller lora

The last part, the distilled lora is perhaps the most tricky part. It probably runs fine with the 0.6 strength set as default for both first and second pass (but it could in theory be far less for first pass, and 0.6 in last pass)

But if you wait a little bit I will upload a workflow soon, that runs Dev model for a more slow render with full steps, that can give better results

But if i use the DEV model with NO lora, can i do 50 steps and is the same result as 20 steps WITH lora? i only had a good resullt with distilled LORA at 1..... 20 steps......, also, one pass WF with DEVm gives me bad result....... maybe because of the DIT cache i used.....

please let me kn ow ehen you get the DEV woekflow so i can compare what i am doing wrong...

If you're seeing jagged pixel-like edges, try changing the Upscale Method from 'Nearest-Exact' to 'Lanczos',

Nice tip, will try that here. See if it improves things ;-)

BOOM! ONE PASS workflow worked well, i removed the stupid DIT CACHE, and did 30 steps and lora at .6 and use deforum LTX lora and look!!!!! 20 seconds DEV in just 14 minutes!!!!
what do you think? 768x768

Yes that looks great ;-) Nice nice ;-)
Reminds me of the Deforum days.. I gotta try that lora ;-) got nostalgic hehe

And in a 2 pass workflow you can absolutely remove the lora from the 1st pass, and do 50 ish steps with dev model ONLY (and leave the dev + lora for 2nd pass upscale)
Its a little bit "complicated" since you bascially have "two models" in the spaghetti mess (And i have "hidden" a lot of spaghetti with get set nodes).
But if you know how to re-range it a bit, its not too hard. Just move the lora loader for the distilled lora away and put it into 2nd pass group. And connect the model "Get" node to that etc

That being said, i plan to make a Dev Model workflow though. For those who dont want to mess around with the nodes and noodles ;-)

Yes that looks great ;-) Nice nice ;-)
Reminds me of the Deforum days.. I gotta try that lora ;-) got nostalgic hehe

And in a 2 pass workflow you can absolutely remove the lora from the 1st pass, and do 50 ish steps with dev model ONLY (and leave the dev + lora for 2nd pass upscale)
Its a little bit "complicated" since you bascially have "two models" in the spaghetti mess (And i have "hidden" a lot of spaghetti with get set nodes).
But if you know how to re-range it a bit, its not too hard. Just move the lora loader for the distilled lora away and put it into 2nd pass group. And connect the model "Get" node to that etc

That being said, i plan to make a Dev Model workflow though. For those who dont want to mess around with the nodes and noodles ;-)

Hi there, my lora's seems not working with the DEV workflow, do you know what's can be the reason ?

Hi there, my lora's seems not working with the DEV workflow, do you know what's can be the reason ?

Any specific lora? perhaps made for LTX-2.0 ? if so you can try set higher strength, even up to say 2.0

Hi there, my lora's seems not working with the DEV workflow, do you know what's can be the reason ?

Any specific lora? perhaps made for LTX-2.0 ? if so you can try set higher strength, even up to say 2.0

Hi Rune, hmm not gonna lie I mostly use nsfw lora's and they seems to not work with the WF ( very strong blurry effect motion, prompt not respected and sometimes it's change completely my subject in I2V )

For example : https://civitai.com/models/2471327/rogue-ai-missionary-cowgirl?modelVersionId=2778591

Im trying to do this kind of transitions scene : https://civitai.com/images/124677141, if by any chance you know how to do it let me know (can delete my post if nsfw is not allowed)

very strong blurry effect motion

You'd probably get better result with a distilled workflow, or increase the steps a lot, maybe all the way up to 40 ish.
But when you increase the steps the lora often gets "weaker", since its probably trained to work on distilled model. Not sure.
And if the lora is "old" (LTX-2.0), sometimes you need to increase the strength quite a lot even for low steps. Sometimes up to 2.0
(although this specific one seems trained on LTX-2.3 and should be fine)

At the link you gave it also looks like its a LOT of lora's in combination

sometimes it's change completely my subject in I2V

It might be trained on text to video only. In your link it as an extremely long prompt with timestamp etc. Probably it was done with T2V

prompt not respected

Also these loras might often have trigger words, or very specific prompt structure.
But first of all perhaps try the distilled workflow.

Also, as a sidenote, some of these loras are often quite simple trained, and does not have audio part of the lora. So they might produce very distorted metallic sounding audio.
For that try use KJNodes LTX Advanced Lora Loader and set all the bottom audio settings to zero.

Lastly, the prompt might get refused by Gemma, so you might have to use non censored version of Gemma ;-) Lots of them at Huggingface, search for abliterated or heretic or just uncensored

Also its 46 seconds long. So it might have been heavily edited ;-) Multiple generations cut together as one video in video editor
(LTX video clips are usually around 5-15 seconds or so)

thanks you very much! Indeed it's edited, will apply your recommendations

very strong blurry effect motion

Most of all think you might have wrong models loaded though.
Such as using the Distilled model, in the Dev workflow.
LTX has 2 models.. Distilled model for low steps and cfg 1 (fast render), and a Dev model that requires far more steps and higher cfg (slow render, but can give better results).
And to confuse even more, you can run Dev model as if it was Distilled model (aka low steps, cfg 1) IF you also use the distilled lora with the Dev model (and dont use this lora with the model already distilled)

Although, the dev workflow might been a bit too "optimistic" with 20 steps if you are aiming for high motions.
Perhaps try the res_2s in sampler settings if so, or bump up the steps a lot more in the 30 ish range or more ...

Also double check if you have the LTX 2.3 vae set, and not LTX 2.0 vae (if you used LTX 2.0 earlier).

(distilled model and distilled workflow are 90% as good (according to LTX), and runs super fast though... and might be the better choice for most use)