Kijai

/

LTXV2_comfy

like 353

Do you have the distill lora for the upscale part?
You need the distilled lora for the 2nd part if using DEV model (without distill lora) for the main part (1st part)
(you can use LoraLoaderModelOnly node to load the lora)

So in other words if you are Dev model WITHOUT distilled lora for main part:

• 1st part of the workflow (1st pass sampler) : DEV model 20+ steps cfg 4
• 2nd part of the workflow (2nd pass upscale) : DEV model + distill lora for the model input (upscale has cfg 1 and low steps)

--

Alternatively you can use distill lora with dev model at BOTH first and second part of the workflow

• 1st part (1st pass sampler) : DEV model + distill lora => cfg 1, 8 steps
• 2nd part of the workflow (2nd pass upscale) : DEV model + distill lora (upscale has cfg 1 and low steps)

Hope that made sense ;-) a little hard to explain.
Its a bit tricky when its 2 different models, to make it correct at both sampler steps

Do you have the distill lora for the upscale part?
You need the distilled lora for the 2nd part if using DEV model (without distill lora) for the main part (1st part)
(you can use LoraLoaderModelOnly node to load the lora)

So in other words if you are Dev model WITHOUT distilled lora for main part:

• 1st part of the workflow (1st pass sampler) : DEV model 20+ steps cfg 4
• 2nd part of the workflow (2nd pass upscale) : DEV model + distill lora for the model input (upscale has cfg 1 and low steps)

--

Alternatively you can use distill lora with dev model at BOTH first and second part of the workflow

• 1st part (1st pass sampler) : DEV model + distill lora => cfg 1, 8 steps
• 2nd part of the workflow (2nd pass upscale) : DEV model + distill lora (upscale has cfg 1 and low steps)

Hope that made sense ;-) a little hard to explain.
Its a bit tricky when its 2 different models, to make it correct at both sampler steps

Ok recap.

If using the dev model I need to use the distilled Lora also for second stage ???

Ok recap.

If using the dev model I need to use the distilled Lora also for second stage ???

Yes ;-) The 2nd step runs in "distilled mode" so to speak (low steps, cfg1)

This screenshot is from ComfyUI default workflow for FULL DEV model (20 steps , cfg 4), and as you can see there is a lora loader for distilled lora for the upscale part

Ok recap.

If using the dev model I need to use the distilled Lora also for second stage ???

Yes ;-) The 2nd step runs in "distilled mode" so to speak (low steps, cfg1)

ok can i use this one from KIJAI also???

ltx-2-19b-distilled-lora-resized_dynamic_fro095_avg_rank_242_bf16.safetensors

ok can i use this one from KIJAI also???

ltx-2-19b-distilled-lora-resized_dynamic_fro095_avg_rank_242_bf16.safetensors

definitively. Even the lower rank ones should you want. They all seem to work great
Its the one i use as well ;-)

ok can i use this one from KIJAI also???

ltx-2-19b-distilled-lora-resized_dynamic_fro095_avg_rank_242_bf16.safetensors

definitively. Even the lower rank ones should you want. They all seem to work great

ok thanks, i will try again tonight

ok can i use this one from KIJAI also???

ltx-2-19b-distilled-lora-resized_dynamic_fro095_avg_rank_242_bf16.safetensors

definitively. Even the lower rank ones should you want. They all seem to work great

but YOU are using the OG of 7GB lol wow

ok can i use this one from KIJAI also???

ltx-2-19b-distilled-lora-resized_dynamic_fro095_avg_rank_242_bf16.safetensors

definitively. Even the lower rank ones should you want. They all seem to work great
Its the one i use as well ;-)

ok question, is there a difference in quality using the DEV model compared to the DISTILLED model?

and with the distilled model i DONT need the distilled lora on stage 2 right?

ok can i use this one from KIJAI also???

ltx-2-19b-distilled-lora-resized_dynamic_fro095_avg_rank_242_bf16.safetensors

definitively. Even the lower rank ones should you want. They all seem to work great
Its the one i use as well ;-)

if i am using the LORA detailer, should i add it on the second pass also? or at the beginning?

ok question, is there a difference in quality using the DEV model compared to the DISTILLED model?

Distilled model : can use low steps (8) and cfg 1 = very fast
Dev model : uses more steps (20), and cfg 4 = slower, but potentially better quality

And here comes the confusing part : you can run the Dev model in "distilled mode" (low steps,cfg 1) you add the distilled lora (this is true for both first part and 2nd part, and for 2nd part its needed)

and with the distilled model i DONT need the distilled lora on stage 2 right?

correct ;-) if you use distilled main model you don't need (and shouldnt use) the distilled lora neither for the first part or the 2nd part (upscale)

if i am using the LORA detailer, should i add it on the second pass also? or at the beginning?

Actually not sure about that one. I havent used that lora. You can always try, and remove it from 2nd part if the result is too much ("burned", too sharp etc).

I see the workflows with camera lora have them for both first and 2nd part, so might be also true for the detail lora

if i am using the LORA detailer, should i add it on the second pass also? or at the beginning?

Actually not sure about that one. I havent used that lora. You can always try, and remove it from 2nd part if the result is too much ("burned", too sharp etc).

I see the workflows with camera lora have them for both first and 2nd part, so might be also true for the detail lora

thanks it worked now!!!