Works well
Not sure if its imagination or not, but seems to work a bit easier on the computer when split into more "regular" models.
And slightly more familiar workflow with that. Thanks for that ;-)
(i didnt render the video in full res, just made for fun ;-))
Not sure if its imagination or not, but seems to work a bit easier on the computer when split into more "regular" models.
And slightly more familiar workflow with that. Thanks for that ;-)(i didnt render the video in full res, just made for fun ;-))
Yeah many people on a discord server said it works better.
I took WF from the video and got an error on audio vae. What could be the error? Both vae files are located in the vae folder.
‘VAE’ object has no attribute 'latent_frequency_bins'
I took WF from the video and got an error on audio vae. What could be the error? Both vae files are located in the vae folder.
‘VAE’ object has no attribute 'latent_frequency_bins'
KJNodes probably not updated, I had to add some code to allow loading the audio VAE with it.
can someone share a working workflow please?
im also getting error -> "‘VAE’ object has no attribute 'latent_frequency_bins'". tried with load vae node also, but still getting different error
"raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
RuntimeError: ERROR: VAE is invalid: None"
can someone share a working workflow please?
im also getting error -> "‘VAE’ object has no attribute 'latent_frequency_bins'". tried with load vae node also, but still getting different error
"raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
RuntimeError: ERROR: VAE is invalid: None"
Currently for the audio VAE you can use the "LTXV Audio VAE loader" node, make sure to keep the audio vae in the checkpoints folder for now, soon after some days the normal vae loader will work
can someone share a working workflow please?
im also getting error -> "‘VAE’ object has no attribute 'latent_frequency_bins'". tried with load vae node also, but still getting different error
"raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
RuntimeError: ERROR: VAE is invalid: None"
Currently for the audio VAE you can use the "LTXV Audio VAE loader" node, make sure to keep the audio vae in the checkpoints folder for now, soon after some days the normal vae loader will work
Awesome. it works fine. Thanks
Also curious whats cooking at KJNodes ;-) Saw some LTX related stuff.
We'll wait and see ;-)
Update KJNodes even though there is no update. Solved the Audio VAE problem.
Works great!. Where's the non-distilled full version of the workflow?
Works great!. Where's the non-distilled full version of the workflow?
Hi, non one can share a simple working worflow ?
Thks in advance.
Hi, non one can share a simple working worflow ?
Thks in advance.
Download one of the videos and drag and drop the video into ComfyUI. The workflows are embedded in the videos. Enjoy!
Works great!. Where's the non-distilled full version of the workflow?
Awesome, works great too.
Works great!. Where's the non-distilled full version of the workflow?
with the DISTILLED version, i got them working MOST of the time, but with the NON distilled versions, it is NOT working, do i need by obligation to use the LORA in this case?
Works great!. Where's the non-distilled full version of the workflow?
with the DISTILLED version, i got them working MOST of the time, but with the NON distilled versions, it is NOT working, do i need by obligation to use the LORA in this case?
Just use the lora since you save on disk space as well, for distillations I would always recommend loras
Works great!. Where's the non-distilled full version of the workflow?
with the DISTILLED version, i got them working MOST of the time, but with the NON distilled versions, it is NOT working, do i need by obligation to use the LORA in this case?
Just use the lora since you save on disk space as well, for distillations I would always recommend loras
no no, my question is that IF using the NON distilled version the MAIN MODEL, if i nEED to use the DESITILLED LORA so the MAIN model will work? or not?
Works great!. Where's the non-distilled full version of the workflow?
with the DISTILLED version, i got them working MOST of the time, but with the NON distilled versions, it is NOT working, do i need by obligation to use the LORA in this case?
Just use the lora since you save on disk space as well, for distillations I would always recommend loras
no no, my question is that IF using the NON distilled version the MAIN MODEL, if i nEED to use the DESITILLED LORA so the MAIN model will work? or not?
Yes it will work, using the distill lora on the dev version will basically turn the dev model into distilled model during inference.
Works great!. Where's the non-distilled full version of the workflow?
with the DISTILLED version, i got them working MOST of the time, but with the NON distilled versions, it is NOT working, do i need by obligation to use the LORA in this case?
Just use the lora since you save on disk space as well, for distillations I would always recommend loras
no no, my question is that IF using the NON distilled version the MAIN MODEL, if i nEED to use the DESITILLED LORA so the MAIN model will work? or not?
Yes it will work, using the distill lora on the dev version will basically turn the dev model into distilled model during inference.
but WE have also the distilled main model. the point of using the NON distilled model is to make Better quality at 20 steps..... doesnt make sense, in that case i prefer to use the distilled model with NO LORA, im lost now.....
Works great!. Where's the non-distilled full version of the workflow?
with the DISTILLED version, i got them working MOST of the time, but with the NON distilled versions, it is NOT working, do i need by obligation to use the LORA in this case?
Just use the lora since you save on disk space as well, for distillations I would always recommend loras
no no, my question is that IF using the NON distilled version the MAIN MODEL, if i nEED to use the DESITILLED LORA so the MAIN model will work? or not?
Yes it will work, using the distill lora on the dev version will basically turn the dev model into distilled model during inference.
but WE have also the distilled main model. the point of using the NON distilled model is to make Better quality at 20 steps..... doesnt make sense, in that case i prefer to use the distilled model with NO LORA, im lost now.....
Just use the distilled lora on the dev model if you want to generate fast. If you want higher quality, remove the lora.
Works great!. Where's the non-distilled full version of the workflow?
with the DISTILLED version, i got them working MOST of the time, but with the NON distilled versions, it is NOT working, do i need by obligation to use the LORA in this case?
Just use the lora since you save on disk space as well, for distillations I would always recommend loras
no no, my question is that IF using the NON distilled version the MAIN MODEL, if i nEED to use the DESITILLED LORA so the MAIN model will work? or not?
Yes it will work, using the distill lora on the dev version will basically turn the dev model into distilled model during inference.
but WE have also the distilled main model. the point of using the NON distilled model is to make Better quality at 20 steps..... doesnt make sense, in that case i prefer to use the distilled model with NO LORA, im lost now.....
Just use the distilled lora on the dev model if you want to generate fast. If you want higher quality, remove the lora.
ah now makes sense, but is not an OBLIGATION to use it to get a result, it is just optional then..... ok thanks.
The reason i am asking if i need the LORA by obligation or not, is because after 20 steps of rendering, i am not getting the results like here - i am leaving EVERYTHING in default as it is when you drag and drop the video into COMFY..... i dont change ANYTHING, just like RUNE XX did, i dont change anything, and THESE ARE MY RESULTS........
very rare then,,,, using the same seed as RUNE XX
The reason i am asking if i need the LORA by obligation or not, is because after 20 steps of rendering, i am not getting the results like here - i am leaving EVERYTHING in default as it is when you drag and drop the video into COMFY..... i dont change ANYTHING, just like RUNE XX did, i dont change anything, and THESE ARE MY RESULTS........
very rare then,,,, using the same seed as RUNE XX
Something might be bugged in your WF https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_ltx2_i2v_distilled.json try this WF
The reason i am asking if i need the LORA by obligation or not, is because after 20 steps of rendering, i am not getting the results like here - i am leaving EVERYTHING in default as it is when you drag and drop the video into COMFY..... i dont change ANYTHING, just like RUNE XX did, i dont change anything, and THESE ARE MY RESULTS........
very rare then,,,, using the same seed as RUNE XX
Something might be bugged in your WF https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_ltx2_i2v_distilled.json try this WF
but i am using the same workflow are RUNE XX, the one that made the video.... does not makes sense..... here look again...... and also, you are sharing a DISTILLED workflow...... i want a NON distilled WF in which i can use KIJAI models.... do you have one NON distilled?
The reason i am asking if i need the LORA by obligation or not, is because after 20 steps of rendering, i am not getting the results like here - i am leaving EVERYTHING in default as it is when you drag and drop the video into COMFY..... i dont change ANYTHING, just like RUNE XX did, i dont change anything, and THESE ARE MY RESULTS........
very rare then,,,, using the same seed as RUNE XX
Something might be bugged in your WF https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_ltx2_i2v_distilled.json try this WF
but i am using the same workflow are RUNE XX, the one that made the video.... does not makes sense..... here look again...... and also, you are sharing a DISTILLED workflow...... i want a NON distilled WF in which i can use KIJAI models.... do you have one NON distilled?
https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_ltx2_i2v.json
The reason i am asking if i need the LORA by obligation or not, is because after 20 steps of rendering, i am not getting the results like here - i am leaving EVERYTHING in default as it is when you drag and drop the video into COMFY..... i dont change ANYTHING, just like RUNE XX did, i dont change anything, and THESE ARE MY RESULTS........
very rare then,,,, using the same seed as RUNE XX
Something might be bugged in your WF https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_ltx2_i2v_distilled.json try this WF
but i am using the same workflow are RUNE XX, the one that made the video.... does not makes sense..... here look again...... and also, you are sharing a DISTILLED workflow...... i want a NON distilled WF in which i can use KIJAI models.... do you have one NON distilled?
https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_ltx2_i2v.json
yes, those are the One from COMFY TEMPLATES, then i need to change MANY nodes to the KIJAI NODES, thats fine.... thanks for replying..... those are the WF i am using since i mentioned the errors, no harm, have a great day.....
You could also try these workflows, and simply swap out model loading with KJ model loader nodes exactly same way as you have in your already workflows.
https://github.com/Lightricks/ComfyUI-LTXVideo/tree/master/example_workflows
I had quite nice results with these, might work better. But the native also works for me.
(if you dont already have ComfyUI-LTXVideo nodes, you need to add that first if you want to try those workflows)
A little GGUF test run, works great ;-) might help those that get OOM on lower ram systems
NB! this support is not out yet, but will be soon https://github.com/city96/ComfyUI-GGUF/pull/399
Sorry guys, but the Kijai version is very bad.
Back to the template fp8 version with the same prompt and seed:Sorry guys, but the Kijai version is very bad.
Back to the template fp8 version with the same prompt and seed:
but you are using KJ nodes also..... and combine with the NATIVE nodes? WTH? lol but nice result.......
Sorry guys, but the Kijai version is very bad.
Back to the template fp8 version with the same prompt and seed:but you are using KJ nodes also..... and combine with the NATIVE nodes? WTH? lol but nice result.......
First one is with distilled model meant to be used with CFG 1 and low steps like 6-8, also the scheduler is different.
The reason i am asking if i need the LORA by obligation or not, is because after 20 steps of rendering, i am not getting the results like here - i am leaving EVERYTHING in default as it is when you drag and drop the video into COMFY..... i dont change ANYTHING, just like RUNE XX did, i dont change anything, and THESE ARE MY RESULTS........
very rare then,,,, using the same seed as RUNE XX
I tried you prompt as T2V without image and it only gives a still image with zooming in effect. I had to mention she is talking and says ... to make it work.
Sorry guys, but the Kijai version is very bad.
Because you are using distilled model with CFG there.....
Sorry guys, but the Kijai version is very bad.
Because you are using distilled model with CFG there.....
I need to officially say apologies! Kijai version works and indeed good result. CFG 1.0 and Step 4 is working fine. Bloody fast. 23s 121 Frames on 1280x704. The only thing I changed is that I bypassed the 0.5 Upscale node as well as the post-upscale group nodes. Additionally added a Basic Scheduler with beta57. Different, but still good result.
For the static zoom image "issue", i think it all comes down to the prompt.
Apparently you should not prompt what you already see, and not like image generation AI
Instead focus on sequence of actions to happen.
basically something like this :
- Style first (cinematic, anime etc)
- Set the scene (what kind of atmosphere, color tones, lighting, etc)
- ACTION (important) - describe the sequence of things to happen - including dialog
- Camera (optional) : how should the camera act/move
- Audio (optional): background sounds, music etc.
LTX-2 Prompting Tips (from the creators of LTX)
Core Actions: Describe events and actions as they occur over time
Audio: Describe sounds and dialogue needed for the scene
Reference Image: Do not repeat details already present
Consistency: Avoid instructions that do not match the reference image, as this will degrade results
I collected some of the LTX-2 prompt guides and instructions (if you use prompt enhancer) here: https://github.com/kijai/ComfyUI-KJNodes/issues/489#issuecomment-3730593217
Hopefully that helps, its just a theory, but i have not had any static image happen ever since ;-) (it could of course also be due to updated models, but prompting the "correct way" wont hurt)