RuneXX

/

LTX-2.3-Workflows

like 74

Yes, been meaning to upload a DEV model workflow.
It's a lot the same, but as you hinted at the main changes are the steps at first pass.
So it can not use the pre-set sigma values from LTX, but rather a LTX Scheduler node where you enter the steps you want.

If you look closely, that node is already there in the workflow but i have collapsed it, and made it smaller. Its right above the sigma node in use.
Just connect that one instead to the sampler in first pass. Set steps to 20-30 and cfg to 4-7.

And use the distilled lora at the model loader (it used to keep the steps a bit less at step 1 (without it, its easily 40-60 steps), and used as well at step 2 upscale)

And if you get stuck, I'll upload a DEV workflow asap ;)

But its really just 2 changes to make it work (+ enable the distilled lora at the model loaders)

Yes, been meaning to upload a DEV model workflow.
It's a lot the same, but as you hinted at the main changes are the steps at first pass.
So it can not use the pre-set sigma values from LTX, but rather a sigma node where you enter the steps you want.

If you look closely, that node is already there in the workflow but i have collapsed it, and made it smaller. Its right above the sigma node in use.
Just connect that one instead to the sampler in first pass. Set steps to 20-30 and cfg to 4-7.

And use the distilled lora at the model loader (it used to keep the steps a bit less at step 1 (without it, its easily 40-60 steps), and used as well at step 2 upscale)

And if you get stuck, I'll upload a DEV workflow asap ;)

But its really just 2 changes to make it work (+ enable the distilled lora at the model loaders)

ok but regarding the LORA, i just activate it at the beginnig, ENABLE IT, and it will be applied also to the phase 2, i dont need to add it also on phase 2...... right.?

Yes, been meaning to upload a DEV model workflow.
It's a lot the same, but as you hinted at the main changes are the steps at first pass.
So it can not use the pre-set sigma values from LTX, but rather a LTX Scheduler node where you enter the steps you want.

If you look closely, that node is already there in the workflow but i have collapsed it, and made it smaller. Its right above the sigma node in use.
Just connect that one instead to the sampler in first pass. Set steps to 20-30 and cfg to 4-7.

And use the distilled lora at the model loader (it used to keep the steps a bit less at step 1 (without it, its easily 40-60 steps), and used as well at step 2 upscale)

And if you get stuck, I'll upload a DEV workflow asap ;)

But its really just 2 changes to make it work (+ enable the distilled lora at the model loaders)

one more thing, does the DIT accelerator cache, works in 2.3?

There is more that is perhaps wanted if one runs the DEV model, so I'll also upload a workflow for that.
A workflow that is slower, for higher quality generations

These are the things I am probably gonna add to a Dev model workflow:

• Dev model (with distilled lora) : 20-30 steps, cfg 4 for first pass
• Perhaps add multimodal guider from LTX
• Perhaps add temporal upscaler (double frame rate)
• Perhaps set the upscaler to 1.5x instead of 2x to make 1st pass even higher quality
• Add a few things from KJNodes such a sage attention, enhance-a-video etc etc

Basically a bit more advanced workflow for those with strong GPU, and tons of patience to wait on each run ;-)

one more thing, does the DIT accelerator cache, works in 2.3?

I haven't tried, but would be surprised if it did.
The Easy Cache node from ComfyUI is an option for sure.

one more thing, does the DIT accelerator cache, works in 2.3?

I haven't tried, but would be surprised if it did.
The Easy Cache node from ComfyUI is an option for sure.

ok wait, im lost a little, you mean that FOR DEV, the LORA is only so it can be used with CFG 1? so if i want to use CFG at 4, so i can activate negative prompts, i DONT need the DEV distilled lora then?

No its still CFG 4 or higher.
But, with the lora it will be able run with 20-30 steps.
Without the lora, you are easily looking at 40-60 steps .. or even more

But it is entirely possible to only use the lora at the upscaler 2nd pass though.

(LTX team uses the lora with the dev model themselves too, so i dont think its much gain to run without)

No its still CFG 4 or higher.
But, with the lora it will be able run with 20-30 steps.
Without the lora, you are easily looking at 40-60 steps .. or even more

But it is entirely possible to only use the lora at the upscaler 2nd pass though.

ahh ok thanks, i will try.....

so i can activate negative prompts

Btw, the negative prompts are activated with the distilled model and CFG 1.
Thats what the NAG node is for. It allows the negative prompt to be used even if CFG is just 1

https://chendaryen.github.io/NAG.github.io/

No its still CFG 4 or higher.
But, with the lora it will be able run with 20-30 steps.
Without the lora, you are easily looking at 40-60 steps .. or even more

But it is entirely possible to only use the lora at the upscaler 2nd pass though.

(LTX team uses the lora with the dev model themselves too, so i dont think its much gain to run without)

ok so, just change the model to DEV, and enable the LORA, and tahts it? no need to change the text encoder extension like before also to DEV? since i dont see anything in KIJAI to change the text encoder to DEV.... right?

Thats it yes ;-) only change the model + enable the distilled lora right below the main model node.

(+ set the steps higher and cfg higher at first pass)