UPDATED TO VERSION 2
VERSION 2 removes the textenhance node to rid the workflow of the terrible time consuming encoding. also provided a secondary workflow inside that gives the correct distilled model settings all in 1 workflow!
PRO-TIP: if you have at least 24gb of COMBINED memory (8gb vram + 16gb ram)
you can fit Kijais fp8 scaled model. i tested it and its WAY better than the gguf formats. worth a try!
I HIGHLY RECOMMEND UPDATING YOUR BAT FILE WITH THESE FLAGS:
--lowvram --disable-xformers --use-pytorch-cross-attention --reserve-vram 2 --disable-smart-memory
(these flags will help with the text encoder tricking the push back onto cpu and will burn your vram as priority first)
__________________________________________________________________________________________________
LOW VRAM workflow for the LTX-2.3 DEV gguf!
NOTE: the workflow is set to 20 steps by default with the distill lora. if you dont use the distill lora, youll need 40 to 50 steps.
You can remove the textenhance node if you dont want to wait for token generation!
workflow is set to run both vaes and text encoder on cpu. if this is too slow for you, swap to main device on vaes and default on text encoder and that will speed it up but you may encounter OOMs because of it.
WARNING! GENERATION TIMES WILL BE EXTREMELY LONG! The Gemma text encoder generates tokens and this process takes VERY long... like over an hour in some cases. You CAN run the text encoders on your gpu but if you have anything less than 10gb itll probably OOM. I recommend chaining the Gemma API node in and generating an API key so you can save VRAM and time by encoding through the cloud. (ITS FREE AND EASY TO SET UP)
I HIGHLY RECOMMEND UPDATING YOUR BAT FILE WITH THESE FLAGS:
--lowvram --disable-xformers --use-pytorch-cross-attention --reserve-vram 2 --disable-smart-memory
(these flags will help with the text encoder tricking the push back onto cpu and will burn your vram as priority first)
FILES:
OPTIONAL Kijais fp8 Scaled (requires load diffusion model node instead of unet loader node and replaces the gguf entirely. )
https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/diffusion_models
DEV gguf (distilled ggufs are in the repo as well)
https://huggingface.co/unsloth/LTX-2.3-GGUF/tree/main
(Unsloth is usually the best quants. They use a special quantization format which keep precision high! get the UD version)
Gemma 3_12B FP4 text encoder
Audio VAE
https://huggingface.co/Kijai/LTX2.3_comfy/blob/main/vae/LTX23_audio_vae_bf16.safetensors
Video VAE
https://huggingface.co/Kijai/LTX2.3_comfy/blob/main/vae/LTX23_video_vae_bf16.safetensors
Text Projection text encoder
https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/text_encoders
Distill Lora
https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-distilled-lora-384.safetensors
Spatial Upscaler (latent upscale models folder)
https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-spatial-upscaler-x1.5-1.0.safetensors
Description
more condensed workflow for DEV model. removes the textadherence node to speed up generation time. also incorporates a secondary workflow inside for the Distill model
FAQ
Comments (40)
Wow! I'll test now
Im still generating my examples so i genuinely dont even know if it works, the text encoder is TIME CONSUMING
@realrebelai Okay, I'll wait. After all, it's a newly released model, and creating a workflow is complicated at first.
Hey there. If I wanted to try this, can you help me, I need to know where these models go?
Are both Audio and Video VAEs going to the normal vae folder?
Is the "text projection" going to the normal text_encoders folder with Gemma?
Thank you 🙌
yes both vaes go in vae folder, both text encoders go in text encoder folder! text projection is a text encoder, itll make sense when you open up the workflow :)
@realrebelai Thank you 🙌🙌🙌
Doesn't the API key send all your prompts to the LTX servers though?
yeah, i believe so. but its free
I dont know?
Make sure the text encoder node is on set device = default (GPU) , not CPU. Running it on CPU will take an eternity.
true, even still its entirely too slow for whatever reason. im trying to figure out a work around by incorporation a gguf of the gemma encoder. going to take some time to figure out because the text projection encoder isnt a gguf format.
@realrebelai Honestly, you can skip the TextGenerateLTX2Prompt node; it’s not essential. My CLIP Text Encode only takes 5-7 seconds. By the way, did you try turning off dynamic VRAM? The new Comfy update turns it on automatically.
@realrebelai I used a dual clip loader gguf node with the gemma 12b Q4 and the regular "ltx-2-19b-embeddings_connector_distill_bf16" clip loaded as the second clip and it worked as far as I can tell. This was with LTX-2 and I saw some examples of others doing the same.
Regarding Gemma and TextEnhancer: You can disable the Prompt Enhancer.
Yes, but some peoppe suck at writing prompts believe it or not 😅
And i wish i wouldve known that before i made a whole youtube video om it as well but you learn something new everyday 😅🤷♂️🫶
Is there a way to fix that? to disable cencored version?
@Donkeyworld remove the textenhance node
For those getting the error: VAELoaderKJ
ERROR: VAE is invalid: None If the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.
--
Reinstall KJnodes pack and make sure you have installed the dependencies.
I shouldve mentioned the dependencies! Thank you
getting this error
VAELoaderKJ
ERROR: VAE is invalid: None If the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.
Update your KJ nodes, should correct it.
Are you using the Gemma api node?
@realrebelai I'm not using any external API's, Just the workflow that's here. If you're asking about the text encoder then I am using gemma_3_12B_fp4_Mixed. I think this one was installed when I started using LTX2
@Artmagnet so i think the issue stems from the dependencies in the kj nodes you may have clashing custom nodes blocking the kj nodes git pull to update the nodes. It happened to me as well you might want to remove all your unnecessary custom nodes and try to git pull again and maybe that will correct it and force the update to the nodes because thats why your getting the error, the nodes arent updated to house the new VAE architecture.
Also, DONT FORGET TO RUN REQUIREMENTS!
same error, solved updating KJ Nodes as people is saying.
@KeDu2o Man I shit you not I've done everything at this point. I can't figure out what's wrong
@Artmagnet your best bet is to just remove all your custom nodes and update brother. Only thing that works for me when mine wont. Also can force the master branch if it wont do it on its own
@Artmagnet No, no, sorry brother. My intention was to confirm that this could be the solution to some of the problems. But as with everything in these situations, there could be many reasons.
took me a while but in the end all i needed to do was a comfyui git pull locally. All works great after that.
Reinstall KJnodes pack and make sure you have installed the dependencies.
getting this error:
mat1 and mat2 shapes cannot be multiplied (1024x4096 and 32x4096)
Do like everyone here and ask ChatGPT ?
@GlowingGuardianGirl yeah I did already. it told me not to use a gguf text encoder which I’m not, and not to use old loras which I’m also not. thanks for your input.
@DankGabrillo Not only a GGUF encoder but any version not compatible with the model will give you mat errors. Give GPT and Grok the full console lines showing what occurs, they will help you, and crosscheck their answers
For those of you looking the fix:
I’d downloaded a different gguf thinking I could save on vram. using this model: ltx-2-3-22b-dev-Q4_K_M.gguf worked fine.
Answer provided by another user on a different workflow page.
@DankGabrillo i downloaded the Unsloth ggufs, they use a really high precision quanitzation process. Their quants always end up being the best in my opinion!
@realrebelai thank you, I’ll be checking them out. I’m not sure but I think the issue was using the Q4_K_S instead of the Q4_K_M model.
@DankGabrillo deff could be it
the distilled GGUF workflow gives phenomical results, but using the dev model in the dev workflow generates very fuzzy and bad results, both are uisng Q6 and their own VAEs respectively
Yea its just a precision issue with the dev ggufs. Fp8 works alot better if you have the vram and ram for it