CivArchive
    LTX-2.3 22B GGUF WORKFLOWS 12GB VRAM - v1.0

    Small update. If you already downloaded before 3/7 there is a new distill LoRA being used. Download it here:

    https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/loras

    If you are coming here and downloading for the first time it's already available.

    LTX 2.3 just dropped and figured it was a good time to update a bit.

    Models all updated for LTX 2.3, basically the same workflow as LTX 2

    We are still using the old LTX2 distill lora as it is working just fine for now. The lora is HUGE at 7+gb total. Previously someone released a lower rank more manageable one which we use right now. Staying with this one for now until I can find a lower rank or make one myself.

    Removed the audio nodes as those were creating a lot of problems and hopefully LTX has fixed the audio which to me it seems much better.

    Removed the tvae previews, no longer working as 2.3 has a new VAE anyway. You just get blurred static but video comes out fine.

    LTX claims to have fixed the portrait video so those should work now. Create those tikkie tokkies to your hearts content

    https://huggingface.co/Lightricks/LTX-2.3 if you're looking for more info on the updates.

    No examples yet but I'll drop a vid when I get the chance.

    With this right now I'm only uploading the all in one zip which will contain the t2v, i2v, ta2v, ia2v, v2v as usual. I WANT ALL THE THINGS! will be the only one available for now.

    I'm doing some testing to see if these settings are going to stay the same, change, if we still need lower strength on distill lora etc. I'll update all workflows as needed. For now all tested and working by me.

    Description

    FAQ

    Comments (28)

    ViperimbaMar 12, 2026· 1 reaction
    CivitAI

    any ideas why output is blurry like this? https://civitai.com/posts/27197809
    I use latest latent upscaler in proper folder

    nktmanyik69Mar 12, 2026

    you are probably using incompatible lora, try disabling or reducing strength

    Miyagi2300Mar 13, 2026
    CivitAI

    Per Claude's analysis: ● Now I can see the exact bug. The Decoder.__init__ pre-computes output_channel (the starting/largest channel count) by

    iterating blocks in reverse — but it only accounts for res_x_y and compress_all multipliers. The LTX 2.3 decoder uses

    compress_space and compress_time blocks which also change channel counts, but those are not counted in the

    pre-computation.

    The math:

    - LTX 2.3 config has compress_space {multiplier:2} and compress_time {multiplier:2} in the decoder

    - These DepthToSpaceUpsample blocks change channels by their stride factors (4x for space, 2x for time)

    - Current code ignores them → computes output_channel = 256

    - Checkpoint was built with those accounted for → actual starting channel = 1024

    This is a bug in ComfyUI's causal_video_autoencoder.py that hasn't been patched yet.

    Urabewe
    Author
    Mar 14, 2026

    what bug exactly?

    benwah1111740Mar 14, 2026
    CivitAI

    tl;dr - In the V2V workflow, is there a way to have only the new segment of video go through upscale and output? I just want the new video output with the source video cut out, and I'll join them later.

    Basically what I'm trying to do is extend a video again and again indefinitely, while maintaining faces and voices. It OOMs on upscale due to running the source video that I already have as well as the new segment. I could just cut both durations in half, but even then, some portion of every video segment is redundant, and that's a new mess, as well as generation time being increased on video I don't need. I've tried some latent trimming nodes in different places, but as you can probably tell, I don't really know what I'm doing, and it barfs when I do that. Do you know of a way I can cut out the source video just before the upscaling stage, but keep all the context from it?

    Urabewe
    Author
    Mar 14, 2026

    there would be ways yes but i haven't put it in nor would i plan to. This would be done by putting nodes in before the upscale that would separate the two then upscale just the new segment.

    Possible yes, but I haven't ever looked into it so you'd have to figure that one out

    beepperson947Mar 14, 2026· 4 reactions
    CivitAI

    So I've been using the Comfy template of LTX 2.3 successfully, and just plugged in your workflow (thank you btw). I think I have everything in its right place, no missing nodes, no errors when I run a generation, but when I click "Run", it cycles through the workflow in like 2 seconds and says "Job Completed", but there's no output. Any idea what this could mean?

    *edit: I figured it out. If anyone else runs into this just make sure you READ THE INSTRUCTIONS IN THE WORKFLOW AND PUT YOUR TEXT ENCODERS IN THE FOLDER HE TELLS YOU TO. lol

    Urabewe
    Author
    Mar 14, 2026· 1 reaction

    Solid advice!! Hey, at least you got it working! Have fun

    beepperson947Mar 14, 2026

    @Urabewe Thank you so much! It's an incredible workflow. Especially coming from WAN 2.2

    TorvaldMar 14, 2026· 6 reactions
    CivitAI

    Anyone having an error with VAE, UPDATE ComfyUI-KJNodes

    gruevyMar 15, 2026
    CivitAI

    These workflows are great. Having a good time figuring out how to make a music videos.

    One request: Being able to start a clip at 0:30.5 instead of 0:30 or 0:31

    Natsu24Mar 19, 2026· 3 reactions
    CivitAI

    The custom node ComfyUI-Logic won't import :/

    I'm using the Easy-Install fork, and everything is up to date

    zexeorMar 20, 2026
    CivitAI

    I can't seem to make the lipsync work. I even changed the audio to stereo, prompted the exact words the character should say, changed the video length to match the audio length, etc. Any help, please?

    Urabewe
    Author
    Mar 20, 2026

    Are you using the crop audio node? That needs to be connected to the original audio node. I will have that fixed up with a few more fixes very soon.

    I'm just about finished with some tweaks and it seems everything has settled down a bit. New upsampler, previews, better connections with crop guides and other intermediate nodes

    zexeorMar 20, 2026

    @Urabewe Hey, yes, everything was connected. Managed to solve it. Elevenlabs audio comes too clean, for whatever reason, adding some background noise makes lip sync work.

    nikoeMar 24, 2026

    @zexeor Did you change anything else than the background noise to get it working?

    zexeorMar 25, 2026

    @nikoe No, that was enough to make it work

    bigmoneypat363Mar 20, 2026
    CivitAI

    OMG this workflow is great. Thanks. 1st time using LTX and it looks great (image to video). I am coming from Wan2.2! So much FASTER!!! I have an RTX5060ti and 32gb of memRAM. Just had to "Update All" in the manager and it worked great

    BobsbigtoeMar 23, 2026
    CivitAI

    is this supposed to merge the original video with the new extended video? im using the V2V and its making a great extended video but its output is just the new video.

    beepperson947Mar 23, 2026
    CivitAI

    Been having a blast with this one. really well made. question, is there a way to avoid the garbled text/logos/graphics that almost always appear at the end of my outputs? I've tried all the negative prompting I can think of but it doesn't seem to help. Thanks!

    Kagi13Mar 23, 2026
    CivitAI

    It doesn't work. First there were problems with the BAE, then with the sampler, and then finally just a black square. Nothing makes sense. There's no sound either.

    Urabewe
    Author
    Mar 23, 2026

    So there seems to be a lot of confusion about why something doesn't work in ComfyUI. Your problem is not the workflow it is your install of Comfy or it is the models you downloaded.

    If you post a few error codes or at least a description of what's going on I might be able to help.

    bigmoneypat363Mar 24, 2026

    Did you install all the links that were provided in the workflow?

    bigmoneypat363Mar 24, 2026
    CivitAI

    I'm getting good results with 11 second videos. Img to video. Is doing "Last Frame Loop" sime on this? My chatbot mentioned doing this for longer videos and changing scenes? I guess I'll look on youtube what that's all about.

    roby_shiMar 28, 2026
    CivitAI

    Hey gays, according to a thread on runninghub, I changed the first sigmas to 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.881203, 0.863321, 0.841251, 0.820089, 0.655, 0.381875, 0.0 . It's worth a try.

    sd_animation_artApr 1, 2026
    CivitAI

    NO SLIDER Clip duration (in seconds)

    michaelscoatthegoat461Apr 3, 2026
    CivitAI

    can this be used on wan2gp?

    velantegApr 11, 2026
    CivitAI

    Can it work without swap? Dont want to kill SSD.

    Workflows
    LTXV2

    Details

    Downloads
    10,364
    Platform
    CivitAI
    Platform Status
    Available
    Created
    3/11/2026
    Updated
    6/10/2026
    Deleted
    -

    Files

    ltx2322BGGUFWORKFLOWS_v10.zip

    Mirrors