LTX-2 19B — COMPLETE COMMUNITY TOOLKIT
Everything you need to start creating with LTX-2 19B in ComfyUI, gathered in one community-maintained Civitai hub: base models, GGUF and FP8 variants, text encoders, VAEs, LoRAs, upscalers, ControlNet resources and ready-to-use workflows.
IMPORTANT
LTX-2 was created by Lightricks. I am not the original creator of the base model. This page mirrors and organizes useful public files from the official Lightricks ecosystem while also hosting my own workflows and community resources.
Official source: https://huggingface.co/Lightricks
START HERE — CHOOSE ONE MAIN MODEL
You do not need to download every file on this page.
• GGUF Q4_M — Smallest GGUF option for constrained hardware
• GGUF Q6_K — Good balance between size and fidelity
• GGUF Q8_0 — Highest-precision GGUF option, with the largest footprint
• FP8 Distilled — Faster generation and rapid iteration
• FP8 Dev — Full model behavior and maximum workflow flexibility
The distilled workflow is the easiest starting point when speed matters. Choose the full FP8 or higher-precision GGUF variants when final quality and flexibility matter more.
INCLUDED RESOURCES
• LTX-2 19B GGUF models: Q4_M, Q6_K and Q8_0
• Standard and distilled FP8 models
• Gemma text encoders
• Video and audio VAE components
• Distilled, detail-enhancement and camera-control LoRAs
• Spatial and temporal upscalers
• Canny ControlNet
• EKKIVOK text-to-video and image-to-video workflows
COMFYUI QUICK SETUP
1. Update ComfyUI before loading an LTX-2 workflow.
2. For native workflows, open Template Library → Video → LTX-2.
3. For GGUF models, install ComfyUI-GGUF and use Unet Loader (GGUF). Place .gguf files in ComfyUI/models/unet/; recent versions can also expose ComfyUI/models/diffusion_models/.
4. Place LoRAs in ComfyUI/models/loras/ and text encoders in ComfyUI/models/text_encoders/.
5. Load the included workflow, install missing custom nodes through ComfyUI Manager, then restart ComfyUI.
6. Select the downloaded filenames in the corresponding loader nodes. A downloaded filename may differ from the name saved inside an older workflow, so relink it once if ComfyUI shows a missing model.
TROUBLESHOOTING
• GGUF model missing: verify that you are using Unet Loader (GGUF), refresh the model list and restart ComfyUI.
• Red or missing nodes: use Install Missing Custom Nodes in ComfyUI Manager, then restart.
• Workflow asks for a differently named file: select the equivalent file manually in the loader node; you do not need to download the entire repository.
• Out of memory: begin with a smaller GGUF quantization or the distilled workflow, reduce resolution and frame count, and avoid loading unused models simultaneously.
• Need help: include your GPU, VRAM, ComfyUI version, workflow name and exact error message in your comment.
CREDITS AND DOCUMENTATION
• Original LTX project: https://huggingface.co/Lightricks
• Official LTX documentation: https://docs.ltx.video/open-source-model/integration-tools/comfy-ui
• Official ComfyUI LTX-2 guide: https://docs.comfy.org/tutorials/video/ltx/ltx-2
• ComfyUI-GGUF: https://github.com/city96/ComfyUI-GGUF
If this hub or one of my workflows helps you, sharing a result and leaving a review is the best way to support continued updates.
Description
Original Distilled Version.
FAQ
Comments (50)
thx for our new home base :D
This is the best Video & audio model ever made (better than wan) !
"Some images have been hidden based on moderation preferences set by the creator" not my home if this censorship exists...
what even is this model. isn't https://civitai.com/models/2291192/lightricks-ltxv-2?modelVersionId=2578325 the original?
@IrakTheOracle yes the @ltxvideo files seems to be the originals, but @EKKIVOK was a day faster for us early creators, and he put a lot of effort in putting all files up. unfortunately with a little censorship settings. don't know if ltxvideo is censoring their uploads.
text encoder?
In the workflow you will get it it's from hugging face repo.
It's not GGUF or Safetensors Files. You need to manually download it into text encoders path.
I made a tuto in my workflows for that :)
will this be posible to run on 16gb vram 32gb ram ?
yes, slightly faster than wan 2.2
Doesn't work for me with the ComfyUI workflow and this checkpoint instead of the official checkpoint.
It crashes when trying to load lora.
yes that's why i uploaded the fp8 version. cause the dev version is up to 40go......
@jvachez594 just try my workflows :)
@ttaetherai wayyyy faster than wan ^^
@EKKIVOK actually, I tested it properly today. I did a 5s 1920x1080 in 18 minutes on my 4060ti.
couldnt go over 480x832 with wan 2.2, which did 4s videos in 15-18 minutes.
so yes. WAYYYY faster
@ttaetherai I have a rtx 5090 and 128go of ram at 6000Mhz.....
So maybe the difference of wpeed coming from that too ^^
@ttaetherai 50 steps wan 2.2 ?
After searching and reserver-vram 4 or 10 it start to work without oom on official workflow. @EKKIVOK those workflows go oom for me . 16gb vram and 32gb ram. THe only ones work for me is official ones distiled fp8 and dev fp8
The output is what concerns mostly i2v, you need to try hard to force to use image as refference. promting vs wan2.2 has to be on another level of details...
@poisas69220 To avoid OOM you need to use distilled version of the model (that model include the lightnbing lora = 8 steps) if you have a more powerfull gpu as me (5090) you can use the fp8 dev version. but you need to know something original model fp16 is 40Gb ....yeah that why i didn't uploaded it, it's a waste of time no one will be able to load that. 2nd things even with a rtx5090 and 128go of ram i experienced some problems. but not anymore with the distilled models, that why i uploaded workflow for that model in particular :)
@poisas69220 Oh and last thing you can bypass the "enchancer" node. it will keep some vram for generation usage. (Enhancer seems to take a lot of memory for barely nothing, just delete that node or just bypass it)
@EKKIVOK t's a waste of time no one will be able to load that? I'm running full model on 4090 and 192GB RAM. and still can play games while the workflow runs.
@janssenbutton I was talking about the 40Gb Models (those ones i didn't upload)
But even with that you will not be able to make video or train wan models and playing games lol. (and i know it, i have 128go 6400mhz and a RTX 5090 Astral.
Even Fp8 models (Wan and LTX2 and i don't talk about the Vae encode and decode + the audio Vae and loras....) take full of your vram AND your great part of your ram xD and you need vram for games...So....
Stop Fake Bragging. especially with a 4090 ...wich is good but not better than a 5090.
if you play Tetris whil generating with high-end models maybe you're correct.
But in the real world with triple A games and even lower you're dreaming.
@EKKIVOK no clue what you are on. I'm using the 40gb distilled model. vram is at 16GB during process while RAM gets filled to like 90GB before its at 180GB during interpolation. and there are games which don't need tons of hardware reservered. I know that I can't play battlefield 6 during that process. I never said that.
@EKKIVOK 128GB is not enough. so you wouldn't know. who is the one fake bragging? you even tell your MHz and videocard model unasked. just lol
@janssenbutton bro, i have an i9 14900KF + Rtx 5090 Astral OC + 128Go DDR5 6400Mhz + ATX power 1275Watt from asus strix :D, i know what i say and i didnt made that config for gaming it was for making models and workflows xD Si i know what i say here ;) (i'm not a gamer)
@janssenbuttonyou're talking about offloading here, when there is no enough vram into your gpu the rest of the models is offloaded inbto your ram xD yeah great but there is a price for that ....the execution time cause you know there is no way your ram could be faster than your Vram xD Offloading can be a solution for lower models but not for 40gb models....
there is always a price to pay. and i say it again if the models is offloading it's because you don't have enough vram....so NO WAY you can play a game that demand Vram to work.
It's just logic and simple. You're whining stop spamming my post or i block and report.
@EKKIVOK I simply said I was able to play games. I never said I play games that needs tons of vram. I'm totally aware that it's slower if using slower RAM instead of VRAM. but thanks for the reminder. I prefer quality over quantity. block and report what you want. I was clearly just answering to your silly comments claiming noone was able to run the 40gb model files. period.
@EKKIVOK I also have a 14900 CPU. just not the cheap KF version like you. mine got built-in Intel UHD Graphics 770. have a good one.
@janssenbutton cheap ? do you know i use KF and do you know what KF just is ? that's juste a version without any IGPU, cause why an IGPU when you have a 5090 ? again....bro stop...thanks. too many bs for this morning for me xD
@EKKIVOK by saying "mine got built-in Intel UHD Graphics 770" I already covered that you are lacking IGPU. less raging. you might get the point then if reading carefully.
@janssenbutton I'm raging now? Lol… omg…
By buying that version of that CPU, it is based on a choice, right?
Why i should rage then ? omg… I made that on purpose, bro, for the simple logic that I will never use any iGPU :D so that's a no-sense dialogue.
Ok, so if you want to play that game, right:
You have a great computer, bro. Nice! But you have a 4090 and I have a 5090 Astral.
You lose, I won. End of the versus of who's gonna pe* the farthest :)
@EKKIVOK I just said that your 'rage' is making you blind.. or ignorant. since I have to repeat what I already said. to the IGPU : the day will come our 4090/5090 need to be replaced. I could still use my PC then without videocard. theres is also other benefits if you don't have to use the 4090 for every single task. I didn't know it was a contest.. and if you can't run the 40GB model with your 'legendary' hardware, I actually won.
@janssenbutton bro, you don't know me right now i'm very calm cause...why i should raging against random ppl i don't know ? i didn't say that i can't run a 40gb model that will be a lie cause i can run higher than that but as i said everything has a price....
here the price is time consuming and weight on your hard drive :/
i have already 4To of loras ans 8To of models .....(i need to clean that mess btw xD)
so why i should use a 40go models that a 20go distilled model can do ?
That's why i didn't upload that an d put the link of LTX HF repo for ppl who don't know that repo exist and get the model over there ^^
See ?
Cause not a lot of people on civitai will download a 40gb models if there is a distilled version exist or even and FP8 one with distilled lora :D
Do you understand that point of view ?
@EKKIVOK I don't think anyone 'understands your point of view'. I literally just corrected you when you said "it's a waste of time no one will be able to load that". which is bullshit. I'm able to load and run it. that was pretty much the whole point. and thanks for reporting me. civitAI shadowbanned your profile for me. (i don't see it anymore when logged in, but I can see it in other browser if not logged in). talking to you is a pain in the arse anyway since I have to repeat over and over before YOU get the point. I was thinking you would be american for sure. but your word count is actually too high. unless you use bugged AI for those bullshit replies. don't give up. there will be a time, education won't be illegal in your country anymore. BYE!
@janssenbutton lol....O Key, bye bro, i hope you will find a good therapist. A really good one xD
@EKKIVOK found a school? :) can't ask in DM since you blocked me :(
@janssenbutton i never blocked you bro :/
@EKKIVOK 404
The page you are looking for doesn't exist
(when clicking your profile)
so yes, you did!
@janssenbutton But you can still post on this page right ? so....maybe a bug ?
@EKKIVOK no bug. I simply use the notification to scroll back to old comments. unblock now!
@janssenbutton
I checked again and indeed, I had blocked you (I honestly didn’t remember anymore).
That happened because of our recent exchanges.
The thing is, like I said before, I really dislike when people criticize work they don’t do themselves. That said, I don’t have anything against you personally.
That actually makes me wonder though ...
btw why ask to be unblocked to see releases that, as you said yourself earlier, didn’t interest you? That seems a bit contradictory.
@EKKIVOK there was no block option for me since you blocked first. seems like there is now a fix for that. I can block you now too. isn't that just great? enjoy blocking people who disagree with your opinions. since all I said was that I'm able to run full model while you said, noone was able to run it. nothing else. bye for good, kid :)
I need a workflow that works for this. Keep getting tokenizer errors with other workflow I found
just search EKKIVOK LTX2 WORKFLOW :)
Tokenizer errors can mean you didn't load the full directory of the text encoder if you chose a different gemma version.
@Elmer588 Gemma version is on my workflow on the tutorial note please download the one with 4 - 5 Go of size :)
your workflow gives tokenizers errors, while official Comfy workflow works.
@blhll 1 it's the same workflow 2 there is no "official workflows" wf are not made by LTX team.
3 You didn't use the right text encoder. ( It's Gemma or Connector text embeddings.. I will upload it, it's mentioned in the description ).
@EKKIVOK I think clip/text encoder loader is not the same as in comfyUI templates, thats where its causing tokenizer error. while ComfyUI template using only 1text encoder, yours Wf require 2 different things.
@blhll it's not 2 different things one is the text encoder and the other it's imply the main model itself ^^ there is no 2 text encoders ^^