CivArchive

    ALL QUESTIONS WILL BE ANSWERED IN THE COMMENT :)

    Cause there is a lot to say about how the model and workflows works...
    I will try to answer and help you if you get some problems :)


    (Upload Still in progress...)


    Upload In Progress....
    COMING SOON :

    • LTX+2 Q6_K.GGUF

    • LTX2 Audio VAE BF16

    • TEXT ENCODER GEMMA 3 12B

    • LTX-2-1B-Embeddings_connector_distill_BF16

    • Special GGUF EDITED WORFLOW IMG2IMG / TXT2IMG

    • LORA DISTILLED VERSION

    • SPATIAL UPSCALER

    • TEMPORAL UPSCALER

    • CAMERA CONTROL LORAS.

    • CONTROLNET AIO LTX2

    • Workflows I2V / V2V / T2V / VDETAILER.

    (Keep in mind i'm not the owner of this model)
    I made this upload from the official Public LTX Huggingface repo for Civitai.
    https://huggingface.co/Lightricks
    I just want to centralize the most useful models on this Civitai page,
    so everyone doesn’t get lost among the many models available on the original repo.
    I’m not asking for any “buzz” or anything like that, it’s just to help the community have a solid LTX base on Civitai (for people not familiar with Hugging Face).
    I won’t upload big models, I leave that to people who actually want to use them, even if the benefits of doing so will be really minimal.
    But I will also centralize here my own releases and base-model variations of unofficial LTX.


    ⚡ LTX-2 FP8 — Distilled (Fast & Lightweight)

    What is LTX-2 FP8 Distilled?

    The FP8 Distilled version is a compressed and accelerated variant of LTX-2, trained to replicate the behavior of the full model while being faster and lighter.

    Distillation reduces model complexity, making it more efficient — at the cost of some fine-grained detail.

    ✅ Key Characteristics

    • Faster generation speed

    • Lower VRAM requirements

    • Quicker prompt response

    • Slightly reduced fine detail compared to full FP8

    • Excellent quality-to-performance ratio

    🎯 Best Use Cases

    • Rapid iteration & testing

    • Prompt exploration

    • Draft videos and previews

    • Creators with limited hardware

    Recommended if:
    You want speed and accessibility, and are willing to trade a small amount of detail for faster results.


    🔹 LTX-2 FP8 — Standard (Full Quality)

    What is LTX-2 FP8 (Standard)?

    The FP8 Standard version is a full-quality LTX-2 model quantized to FP8 precision.
    It preserves the complete architecture and capabilities of the original model while reducing memory usage.

    This is NOT a simplified model.
    Only the numerical precision is reduced — the model’s intelligence, structure, and behavior remain intact.

    ✅ Key Characteristics

    • High visual fidelity and detail

    • Strong temporal consistency

    • Full audio-video synchronization

    • Lower VRAM usage than FP16

    • Stable and reliable for long generations

    🎯 Best Use Cases

    • Cinematic video generation

    • Final renders and high-quality outputs

    • Creators who want maximum quality with lower hardware requirements

    Recommended if:
    You want the best possible quality in FP8, with no compromise on features or flexibility.


    🧠 Which One Should You Choose?

    • 🎬 Go with FP8 Standard if quality and consistency matter most

    • Go with FP8 Distilled if speed and efficiency are your priority

    Both versions are fully compatible with ComfyUI workflows and part of the same LTX-2 creative ecosystem.


    📌 What is LTX-2?

    LTX-2 is a powerful multimodal AI model that transforms text prompts, images, or other media into fully synchronized audiovisual videos — with motion, dialogue, music, and ambient sound generated in one unified pass. It’s built on a hybrid Diffusion-Transformer (DiT) architecture designed specifically for efficient spatiotemporal generation and audio-video alignment. LTX-2+1

    This approach lets creators go from idea to cinematic result without stitching separate audio tracks manually — a major step beyond typical text-to-video systems. LTX-2


    Key Features & Capabilities

    🎥 Cinematic Quality Output

    • Native 4K resolution support with playback up to 50 FPS, delivering smooth, high-detail video clips ideal for cinematic, commercial, or creative use. LTX-2

    🎵 Unified Audio & Visual Generation

    • Generates synchronized audio — including dialogue, ambience and music — alongside the video in a single generation pass, removing the need for external audio sync tools. LTX-2

    🔄 Flexible Input & Output Modes

    • Works with text prompts, image references, multi-keyframe conditioning, and more to animate concepts or stills into motion. LTX-2

    ⚙️ Performance Modes

    • Multiple performance configurations (Fast, Pro, Ultra) allow creators to balance speed and quality according to project needs — from quick drafts to production-ready renders. LTX-2

    🧠 Efficient & Accessible

    • Highly optimized for consumer-grade GPUs — efficient enough to run on ~16 GB VRAM hardware with FP8/FP4 quantization options — making AI video production more accessible. Reddit

    🛠️ Open & Extensible

    • Fully open weights, codebase, and workflows, enabling fine-tuning, custom LoRAs, and integration into tools like ComfyUI. Hugging Face


    📈 Improvements Over Earlier Versions

    Compared to the original LTX family and other open video models, LTX-2 raises the bar in several key areas:

    Audio Integration Built-In
    Instead of generating silent videos and requiring post-processing, LTX-2 outputs audio and visual streams together with temporal coherence. LTX-2

    Higher Resolution & Frame Rates
    Supports native 4K at up to 50 frames per second, reaching cinema-grade quality unlike many earlier community models that cap at lower resolutions or fps. LTX-2

    Longer Clips
    Offers extended duration generation (up to ~20 s clips) with continuous quality and audio coherence — exceeding many alternatives. LTX-2+1

    Expanded Workflows
    Native support in ComfyUI plus custom workflows empowers users with text-to-video, image-to-video, multi-keyframe conditioning, and creative control nodes. comfyui.org+1


    🧠 Typical Use Cases

    🔹 Cinematic storyboarding & concept visuals
    🔹 Social media & marketing video content
    🔹 Animated storytelling & motion design
    🔹 Game cutscenes & immersive narratives
    🔹 Product visualizations & dynamic ads

    Whether for rapid prototyping or production output, LTX-2 empowers creators with professional-grade generative video. LTX-2


    🧩 Included Files & Variants

    Depending on the checkpoint uploaded, this collection may include:

    • Full Model Checkpoints (bf16 / fp8 / fp4) — maximum quality with quantization options

    • Distilled Variants — faster iteration with lighter compute cost

    • Spatial & Temporal Upscalers — improve resolution or frame rate via multiscale pipelines

    • LoRA & Fine-Tuning Packs — custom stylistic or control extension modules Hugging Face


    🔧 ComfyUI Integration & Workflows

    Included workflow templates help you use LTX-2 in ComfyUI with nodes for:

    📌 Text-to-Video — generate animated clips from prompts
    📌 Image-to-Video — animate still images with camera motion and style
    📌 Video Conditioning — extend clips forward/backward or refine motions
    📌 Keyframe Controls — precise guidance over scene transitions

    These workflows are designed for ease-of-use and creative flexibility while demonstrating best practices for prompt structure and smooth temporal motion. LTX Documentation


    🧠 Foundation Model Philosophy

    LTX-2 goes beyond a single task — it’s a foundation model for audiovisual creative AI. Open access to its weights, code, and tools encourages developers, artists, researchers, and hobbyists alike to customize, extend and innovate on a common platform. Hugging Face


    📌 Summary

    LTX-2 is not just another video model — it is a production-ready, synchronized audio-video foundation model that pushes the boundaries of what open discourse video generation can achieve. With cinematic output quality, flexible workflows, and a fully open ecosystem, LTX-2 stands as one of the most capable generative video tools available today. LTX-2

    Description

    This is the quantized version of LTX-2 DEV (can be used with 8steps distilled lora.)

    FAQ

    Comments (69)

    MrReclusive666Jan 24, 2026· 1 reaction
    CivitAI

    you should add the fp4 dev model. I kept reading don't try it unless you have a 5090, nah dude, runs great on 4090, no speed boost over fp8, but huge memory savings, don't have to quantize or distorch swap, running 1600x960 10 second clip right now, 2 minutes, was taking me 5 minutes on fp8 model because i had to memory swap the model, i can even push more, that was only 19gb usage.

    EKKIVOK
    Author
    Jan 24, 2026

    hmmm ....yeah i will upload an FP4 version

    spiralzJan 24, 2026
    CivitAI

    doesn anyone know which the most efficient gemma model to use? So many versions are usable now

    MrReclusive666Jan 24, 2026

    ive been using the 400mb unsloth version, is it the best? idk, but its 400mb lol

    EKKIVOK
    Author
    Jan 24, 2026· 1 reaction

    @spiralz if you got a lot of vram and ram i suggest abliaterated version.
    if not....use the one with less weight :) that will do the job or use GGUF (i will upload the best one)

    EKKIVOK
    Author
    Jan 24, 2026

    @MrReclusive666 basicaly yes it's text encoder :)

    MrReclusive666Jan 24, 2026

    @EKKIVOK i still can't get abliatarated working right, i did some videos yesterday, ended up using the 400mb unsloth gemma 3, but tried it with abliaterated, ended up with 2 dudes sifting though rocks? it completely ignored my prompt, but the 400mb gemma 3 gave me exactly what I asked for.

    EKKIVOK
    Author
    Jan 24, 2026

    @MrReclusive666 hmm did u use the connector embeddings text encoder with it ?

    MrReclusive666Jan 24, 2026

    @EKKIVOK yeah, i did, i think my issue is multigpu, certain things don't like to run if another model it wants to talk to is on another gpu, my assumption is that "connector" needs the model still, but that model is isolated on its own gpu so it can't find it.

    EKKIVOK
    Author
    Jan 24, 2026

    @MrReclusive666 i don't use it so i can't help you im really sorry about that but my release of tomorrow (special workflow) will maybe help you out) i will try to edit and include a special version for you :)

    MrReclusive666Jan 24, 2026

    @EKKIVOK ill take a look when you release it, i don't look at a lot of workflows on here, because the multigpu thing, don't need to try and cram vae,clip, and model all on one gpu when i have 3 gpu's

    EKKIVOK
    Author
    Jan 24, 2026

    @MrReclusive666 oh so the problem comes from there 3 gpu comfyui don't really like that especially since the last updates...

    MrReclusive666Jan 24, 2026

    @EKKIVOK its been running fine, just certain things don't like to talk to eachother if they are on separate gpu's, it hasn't been a big deal, just wish i could find a solution to gemma 3, don't like having half a 4090 dedicated to it.

    EKKIVOK
    Author
    Jan 25, 2026

    @MrReclusive666 in fact i never had multiple gpu i have only a 5090

    MrReclusive666Jan 25, 2026

    @EKKIVOK yeah, i got 2 4090's and a 4080. would like a 5090, but honestly, half tempted to start looking for used 3090's since they had nvlink, get 4 of those at a decent price, bam 96gb vram.

    EKKIVOK
    Author
    Jan 25, 2026

    @MrReclusive666 that's really nice and tho i didn't know comfyui can work with multiple card xD i guess i will buy a second 5090 then xD

    MrReclusive666Jan 25, 2026· 2 reactions

    @EKKIVOK yeah, unfortunately without nvlink, they work independently, so i have 2 24gb cards and a 16 gb card, not a 64gb card, so only comes in handy separating things like vae and clip from unet, but massive unets, still a pain, which is why fp4 working was a nice find, the unet is only 14gb at fp4

    Seii1Jan 24, 2026
    CivitAI

    is IMG TO video is doing a lot longer than t2v, t2v only like 4minutes 720p, but i2v like 20minutes wth am i doing it wrong,32gb ram/16gbvram

    EKKIVOK
    Author
    Jan 24, 2026

    @Seii1 wjhat is exactly your specs ?

    Seii1Jan 24, 2026

    @EKKIVOK 4060ti with 16gb vram + 32gb ram, my virtual i sett t0 100gb on nvme

    EKKIVOK
    Author
    Jan 24, 2026

    @Seii1 hmm seems a little bit weak for dev version....did you tried distilled version 8 steps ?
    if yes what res did u use for your videos ?

    Seii1Jan 24, 2026

    @EKKIVOK using q6 720

    EKKIVOK
    Author
    Jan 24, 2026

    @Seii1 with 20 or 8 steps ?

    Seii1Jan 24, 2026

    @EKKIVOK 8 steps

    EKKIVOK
    Author
    Jan 24, 2026

    @Seii1 so try to use it with the distilled lora it should accelerate the process :)

    Seii1Jan 24, 2026

    @EKKIVOK btw do use have Workflow for the GGUF version ?

    EKKIVOK
    Author
    Jan 24, 2026

    @Seii1 Yeah i made one it will be release tomorrow with some kind of ...surprises :D

    vAnN47Jan 29, 2026

    hi @Seii1 where did you find the i2v workflow?

    Seii1Jan 30, 2026· 3 reactions

    @vAnN47 from this post, just download the video and place it into ur comfy

    EKKIVOK
    Author
    Jan 31, 2026· 1 reaction

    @vAnN47 @Seii1 exactly <3 but the final workflows will be released soon :)
    The video was made with a none definitive one, expect better results with the new one !

    vAnN47Jan 31, 2026

    @EKKIVOK cant wait for the workflow :)

    lidianeporto9248Jan 24, 2026· 3 reactions
    CivitAI

    16gb-vram here. I gave up on LTX2 because the second sampler (that one that processes the upscaled latent) runs very slow. I truly believe that in a few weeks or months there will be better models for Us, the people that can not afford an 5090. And yes, the worflows works, I produced around 30 funny videos using it.

    EKKIVOK
    Author
    Jan 24, 2026

    si i'm happy that it worked for you too, yes, the spatial upscaler is a little bit slow BUT if you lower the res to 880x440 or 440x880 it will work better and way faster. The maximum for LTX2 is basicaly 720x1280 res. i have a 5090 and able to do 1080p but the result is not better with those classic res :)

    EKKIVOK
    Author
    Jan 24, 2026· 1 reaction

    @lidianeporto9248 every videos i made in that page are basicaly 720p and 440p :)
    so try to lower the resolution upscaler will do a better job and faster ;)

    lidianeporto9248Feb 5, 2026

    @EKKIVOK Thanks, friend. In fact now I'm very satisfied after I added this to command line "--cache-none --reserve-vram 3 --preview-method none". No OOM (out of memory) 24 fps, perfect lip sync, fast generations. I use gemma q2-gguf.

    EKKIVOK
    Author
    Feb 6, 2026

    @lidianeporto9248 You can use my personal Args too and see if it's fit for you :)

    --windows-standalone-build --use-sage-attention --fast fp16_accumulation --cuda-malloc --auto-launch

    AginoJan 24, 2026
    CivitAI

    How did you do the One Punch Man video? I have a hard time making anything with anime stuff. What prompted you to look like an anime scene?

    EKKIVOK
    Author
    Jan 24, 2026

    hmmmm download the video and drag it on your comfyui workspace :D
    Magic can happen .....Just sayin ..... <3 @Agino

    AginoJan 24, 2026

    @EKKIVOK I did that and didn't work, it can't find the workflow from the video

    EKKIVOK
    Author
    Jan 24, 2026

    @Agino so ....wait for it, it come tommorow :)

    AginoJan 24, 2026

    @EKKIVOK ok i will wait then

    EKKIVOK
    Author
    Jan 25, 2026· 1 reaction

    @Agino okay <3 trust me you will like those workflows :)

    AginoJan 25, 2026· 1 reaction

    @EKKIVOK ok i will be wait for the good new

    HoneyphoriaJan 25, 2026· 5 reactions
    CivitAI

    Hi bro! How do you think, is there any chance for NSFW LTX 2 like Wan 2.2 in the future?

    EKKIVOK
    Author
    Jan 25, 2026· 2 reactions

    @Honeyphoria yeah and you know bro ? Check this page this week and you will know why :D

    HoneyphoriaJan 25, 2026· 2 reactions

    @EKKIVOK OMG. I will support you with all money i can😁 It's gonna be new era of AI NSFW waifu creators!

    EKKIVOK
    Author
    Jan 25, 2026· 4 reactions

    @Honeyphoria 3 NSFW versions coming soon :)

    HoneyphoriaJan 25, 2026· 1 reaction

    @EKKIVOK ohhh boy, i'm so excited!! You're a legend!!

    xrc25toJan 26, 2026· 2 reactions
    CivitAI

    Hi! Did I miss something or is there a workflow with the model? :)

    EKKIVOK
    Author
    Jan 27, 2026· 4 reactions

    sure there is there is 3 :) search ekkivok workflow ltx. but this early stage, i'm working on a mucj much MICH more efficient and advanced workflows for those models :)

    wrOngplanetFeb 3, 2026
    CivitAI

    I've only got 16 GB VRAM (64 GB DDR4). But:


    How do you get these nodes?

    VHS_VideoCombine

    easy cleanGpuUsed

    CM_FloatToInt

    Only the 2 first is from your workflow and my comfy is missing those.

    Last one is not in your workflow, but in another LTX2 workflow (but then its model is 40 GB...), and btw it's found in "ComfyMath" nodes, which again IS found in Custom Nodes Manager.

    The others are not. Not directly at least.

    "VHS_VideoCombine" is mentioned in the description of "Video Output Bridge" node. Is that the one to install?

    "easy cleanGpuUsed" is not found in that manager. Do I just search it up and (if found) git pull it into custom_nodes?

    Also, the name in the "Load Checkpoint" node is different, and apparently I can't click on the name to find another one, it's fixed. I got your "all in one" models, in subfolers of "checkpoints". I also tried having them in "diffusion_models" and "unet". Maybe they should be in another folder?

    Also, since I suspect this is needed: How do I load GGUF's? (something about unet nodes but I have no idea which one that shows up in that Nodes Manager).

    Yeah I have no idea what I'm doing, I've barely gotten ComfyUI to work (again, this time via yet another install method, but it works the same afaik, I can do some Z-image gens just fine),

    Sorry but this is very frustrating.

    EKKIVOK
    Author
    Feb 4, 2026· 1 reaction

    pretty simple you need to use comfyui manager. Comfyui Manager is able to find and install missing nodes in a workflow :) you NEED that custom node in order to use any workflows properly :)

    wrOngplanetFeb 4, 2026

    @EKKIVOK Thanks for replying!

    Okay, so maybe there's a simple fix like that, I'll keep trying.

    Just FYI or something, but when I open a workflow with missing nodes, I get a pop-up listing missing nodes, Just a list, and I've only tried a simple copy-paste names in Custom Nodes Manager, to no avail ofc because apparently they are packaged with other (custom) nodes.

    Sometimes I get a pop up about missing models, but at least they have a download link. (As said, kind of new to ComfyUI).

    wrOngplanetFeb 5, 2026

    Sorry for this mess... this is not for your workflow anymore but for the "LTX-2 Text to Video" template (or I guess workflow) that I somehow got in ComfyUI's Templates sidebar button.

    I no longer get the list of missing nodes (Unless I use your workflow!), I think all I did was update comfyUI (which I had done earler also).

    But I can't select a model. The "Load Checkpoint" node's name field is grayed out. Same for the "LTXV Audio VAE Loader" node. Text encoder node worked (for selecting a model at least).

    According to this: https://comfyui.dev/docs/guides/nodes/load-checkpoint/

    The checkpoint should be located in the expected place (models/checkpoints). But ofc that doesn't work. (I also tried diffusion_models folder).

    Now I do use that extra_model_path.yaml and point to another drive, but not even when copying a checkpoint to ComfyUI's install folder and model/checkpoints/ helped (refreshed browser ofc). I only tried with one smallish checkpoint though. Maybe it's a wrong one that the loader don't recognize, or something else. I have no idea.

    I do use subfolders, like "LTX2" for those models, and "Z-Image" for those to try and keep things somewhat organized, as long as I can click and get that drop-down list it works just fine for Z-image, Qwen-TTS and ACE Step 1.5, which are the ones I've made to work in ComfyUI so far.

    EKKIVOK
    Author
    Feb 6, 2026· 1 reaction

    @wrOngplanet I personally don’t have this issue on my side, but I’m going to push updated workflows as fast as I can today on Civitai and make sure they work as perfectly as possible.

    PixivTdogFeb 7, 2026· 14 reactions
    CivitAI

    Can you like provide a workflow that DOESNT have renamed models? Wtf is all this shit? Why are there so many downloads. Your CHATGPT copy paste read me is fucking useless. Can you name things like a normal person?

    EKKIVOK
    Author
    Feb 7, 2026

    No. and if you're not ok with that there is other releases.
    Good luck. youy didn't made the job to unite all that different kind of models in one post.
    I did it, stop complaining like a baby. be thankfull for someone did that for you.
    Have a good night <3

    PixivTdogFeb 9, 2026· 3 reactions

    @EKKIVOK imagine telling your customer stop complaining when you can't even do a simple job LOL.

    EKKIVOK
    Author
    Feb 9, 2026· 5 reactions

    @PixivTdog 

    Do the d*mn job first, then we can talk.
    Stop whining about free stuff that people spent their own time uploading to a site.
    Dude, it’s FREE, so quit complaining or go make it yourself.

    If you don’t know how to do it, stop being rude and just ask nicely for once.
    Don’t come crying when the answer matches the vibe of your first message.
    I’m not here to deal with people like that.

    When customers pay me, I do the work, I update it however they want, done.
    Here? It’s free. All the useful links and direct download links are right there in front of your eyes and you’re still crying like a baby.
    Come on, man… move on.

    Oh and another thing it literally says “Upload in progress”.
    The post isn’t even 100% finished. I’ve got a family, a job, hobbies…

    I’m not sitting here 24/7 waiting for your personal satisfaction, especially not when someone’s bitching like that over free content.

    Not happy? The original Hugging Face repo is linked in the description too.
    Deal with it. But stop raging at free upload. Or DIY.

    Final message i'm done talking to ppl like that.

    TononMauricioFeb 9, 2026

    Get a job

    EKKIVOK
    Author
    Feb 10, 2026

    @TononMauricio "The post isn’t even 100% finished. I’ve got a family, a job, hobbies…" learn how to read bro.

    philcoreFeb 10, 2026· 5 reactions

    @EKKIVOK it was quite rude of you to release your workflow for free, I guess ... or whatever they are complaining about. just joking of course, simply wanted to send you some good vibes. don't let them get to you and keep your spirits high. best wishes. ;)

    EKKIVOK
    Author
    Feb 11, 2026· 4 reactions

    @philcore thanks bro, all my releases are free, it's for that reason when i receive that kind of comments it make me mad....xD it's the rules i guess....but not the good ones...

    PixivTdogFeb 21, 2026· 2 reactions

    @EKKIVOK I didn't read anything you said, I guess you enjoy using twitter.

    coachbateMar 9, 2026· 2 reactions

    @EKKIVOK there's always a troll or two - just ignore

    TononMauricioMar 24, 2026

    @EKKIVOK I was obviously talking to the comment creator. such a fragile ego gez

    ac728857Feb 18, 2026
    CivitAI

    Hello,

    I am trying to use the workflow that you posted "EKKIVOK - LTX2 IMG2IMG.json", however I can not seem to find "proj_linear.safetensors" anywhere, might you know how I can get it? Here is the error message:

    "LTXVGemmaCLIPModelLoader

    No such file or directory: C:\Users\User1\Downloads\ComfyUI\models\tts\Qwen3-TTS\Qwen3-TTS-12Hz-1.7B-VoiceDesign\proj_linear.safetensors"

    cg1Mar 11, 2026

    can you have a gguf workflow? i can't seem to redo this in gguf models

    obsidiancloudFeb 22, 2026· 2 reactions
    CivitAI

    your gemma 3 is corrupted, also you didnt provide any explanation of how to use this fp8 ver (I'm assuming)? The workflow u have says download entire gemma3 repo

    Checkpoint
    LTXV

    Details

    Downloads
    1,314
    Platform
    CivitAI
    Platform Status
    Available
    Created
    1/24/2026
    Updated
    6/11/2026
    Deleted
    -

    Available On (1 platform)

    Same model published on other platforms. May have additional downloads or version variants.