CivArchive
    LTX-2.3 All-In-One workflow for RTX 3060 with 12 GB VRAM + 32 GB RAM - LTX-2.3 v1.0
    NSFW

    [edit:

    24.04.2026: Update version 4.3 (see version description).

    Minor update and bug fix.

    Thanks to all users for the many inputs over the last days and weeks 🙂

    Attention:

    If you struggle with node conflicts or you get errors while running the workflow, please have a look at my short Trouble Shooting Guide note in the wokflow first. Most importent is to update all components sucsessfully! ]

    Special thanks to:

    @ArcleinSK for investigation and solving the FLF issue, as well as forcing the First-Mid-Last Frame option and last but not least for charing fantastic knowlage.

    @boinobin730 for initialising, forcing and supporting this project in all kinds of matter, like providing links, running tests, sharing knowlage and inspiring diskussions.

    @Urabewe for publishing the original, perfectly running 12 GB VRAM LTX-2.3 workflows mainly used here in this workflow.

    Features:

    Simple to use all-In-One LTX-2 workflow with options for:

    • Text to Video

    • Image to Video

    • First/Last Frame to Video

    • Fisrt/Mid/Last Frame to Video

    • Video to Video

    • Text + Audio to Video

    • Image + Audio to Video

    • First/Last Frame + Audio to Video

    • First/Mid/Last Frame + Audio to Video

    • easy switching between all options,

    • all steps highly automated: no manual frame or width/hight calculations necessary,

    • easy to set inputs by predefined sliders and aspeckt ratio inputs (no risk to set wrong frame counts or wrong width/hight values),

    • completely automated resizing and cropping (if necessary) of your input images/videos.

    • brilliant audio generation (speech/sound) with LTX-2.3.

    LTX-2.3 specifications:

    Workflow version v4.3 consistently follows the LTX-2.3 specifications for 16:9/9:16 aspect ratios, including automatic width/hight calculations, as well as automatic input image/video resizing/cropping.

    In addition you can simply choose now any other aspect ratios according to your needs while still getting the right values calculated for width/hight and automatic image/video resize/crop.

    Requirements:

    • GPU with 12 GB VRAM (some users reported they got it running with 8 GB too),

    • 32 GB VRAM,

    • Swap file size: 64 - 128 GB.

    Speed and video length:

    Runs very fast: 5 second (1280 x 864) Video: < 10 minutes.

    Generation of long high quality videos in one run possible: 10 - 20 seconds without any issues,

    Testrun: 30 second video (1024 x 704) tooks around 40 minutes without any OOM errors. Longer videos might be possible, but not tested yet.

    Important:

    This workflow is intended for advanced comfyui users who know how to install and operate the system and are able to resolve basic system errors themselves, like as node conflicts, or general system issues.

    About this workflow:

    This workflow is mainly based on the fantastic LTX-2.3 workflows of @Urabewe.

    As far as I know, those were the first workflows running LTX-2 with 12 GB VRAM. All credits goes to the original creator.

    My job was only to combine and organise the different workflows in a simple to use all-in-one design.

    Description

    New LTX-2.3 model.

    • Version v1.0 is early "alpha" state, but ready to test. All options should work properly. Main downside: we have to use the old ltx-2 Loras actually.

    • Added Text + Audio to Video + Audio option.

    Please let me know, if you find any bugs or better Loras to use.

    FAQ

    Comments (121)

    artwex311Mar 8, 2026· 2 reactions
    CivitAI

    I'm encountering this error when using i2v: "Cannot execute because a node is missing the class_type property: Node ID '#209:183'". Could you please help me take a look? Thank you.

    arkinson
    Author
    Mar 8, 2026· 1 reaction

    @artwex311 Are you allready on the latest workflow version v1.1? Please see my updated model description for the right upscaler model too (the link in the workflow still points to the wrong old model). t2v is running? Your error message points to a missing/not installed node. Important: do the complete update routine as described at "Troubleshooting" and check for custom node conflicts.

    artwex311Mar 9, 2026· 1 reaction

    @arkinson It worked fine after I reinstalled it using comfyui-easy-install, thank you!

    arkinson
    Author
    Mar 9, 2026

    @artwex311 Thanks for your feedback. I`m glad you got it running 🙂

    mosagonemad107Mar 8, 2026· 4 reactions
    CivitAI

    Getting VEA error :(

    arkinson
    Author
    Mar 8, 2026· 1 reaction

    First, please look at my updated model desription for the right upscaler model (the link in the workflow still points to the old ltx-2 upscaler). Otherwise give a usefull and complete error desription.

    arkinson
    Author
    Mar 9, 2026· 1 reaction

    KJ nodes updated?

    antigen3Mar 9, 2026· 2 reactions

    I was getting this and it was a comfi update issue, I had to update comfi a couple of times before it worked (was getting errors when updating that it couldn't find some files IDK why)

    arkinson
    Author
    Mar 9, 2026

    @antigen3 Thank you for your feedback 👍

    mosagonemad107Mar 12, 2026

    Got it fixed after reinstalling ComfyUI , Now wokring perfectly in the RTX 5080 16gb GPU 9950X and 32Gb (16x2) 5600mts DDR5 RAM ( I have bottle necked RAM as i got the system recently :( )- Kj node update didnt work so I went all out reinstalling - . Thanks guys

    arkinson
    Author
    Mar 12, 2026· 1 reaction

    @mosagonemad107 Hi - thank you too. I`m glad you got it running.

    mosagonemad107Mar 13, 2026· 1 reaction

    @arkinson You are the best ...

    BbirdMar 8, 2026· 3 reactions
    CivitAI

    Thank you! This is the first LTX 2.3 workflow that I got to work without any errors or exotic missing nodes!
    Speed is great, on my 4060TI/16GB a 15 seconds video renders in about 6 minutes at 1280x*

    Just a basic question: Does the prompt influence the output on a I2V+audio.mp3 render?

    arkinson
    Author
    Mar 8, 2026

    @Bbird Thank you for your feedback 😋 Please have a look at my updated model description here for the right upscaler model (the link in the workflow is wrong and points still to the old ltx-2 upscaler). For more detailed discussions about quality and issues you might also look here.

    Sorry, I do not get your question. What do you mean with "prompt influence the output on a I2V+audio.mp3 render"? Do you want to use externel mp3 software?

    BbirdMar 8, 2026

    @arkinson No no, I just wondered if I need a text prompt at all for I2V+audio , or if my audio track does all the magic.

    arkinson
    Author
    Mar 8, 2026

    @Bbird Ok, so you talk about image+audio 2 video, not i2v.

    It mainly depends from your start image and audio input. And of cource, your prompt should describe as well as possible, what you want to get out. But keep in mind, ltx in generall seems to need very special prompting.

    BbirdMar 9, 2026· 1 reaction

    @arkinson Yeah thanks!
    Sorry, I was a LTX noob until yesterday.

    After some testing, the video follows the prompt "kind of", but enough to use some instructions.
    Thank you again providing the workflow. ;)

    arkinson
    Author
    Mar 9, 2026· 1 reaction

    @Bbird I`m glad you got it running 🙂 Good luck!

    MangledHeavenMar 8, 2026· 2 reactions
    CivitAI

    Great workflow! Worked on my first try with no tweaking. One thing though, the Text to Video workflow is generating a video with no audio, a video with audio, and a still image of the last frame. How do I get it to just output the video with audio? I disabled the Save Last Frame Image node but it still outputs the last frame image and the video with no audio. Thanks again for a wonderful workflow!

    arkinson
    Author
    Mar 8, 2026· 1 reaction

    @MangledHeaven The workflow allways generates 3 output files: 1 png file (this is the last frame image) and 2 mp4 files (the devil knows why the video combine node generates 2 mp4 files). Anyway: the file named ltx2_xxxxx-audio.mp4 is the final (video + audio) file. If you do not hear any audio, just try another/better prompt.

    MangledHeavenMar 9, 2026· 2 reactions

    I figured it out. Go to ComfyUI settings (Gear Icon) then go to the settings for VHS, then in the Output section at the top, turn off "Keep required intermediate files after successful execution" and "Save png of first frame for metadata". Once you do this, it will only output the video + audio file. This keeps your output folder much cleaner. Thanks again for the awesome workflow!

    arkinson
    Author
    Mar 9, 2026· 1 reaction

    @MangledHeaven Brilliant! I never saw the VHS options in the comfyui settings. Thank you for the hint.

    Rifler1Mar 8, 2026· 4 reactions
    CivitAI

    I tested 2.3 with this workflow. i2v works very well.

    Tips for avoiding errors:

    - Update Comfy and KJNodes to the latest version. This is important!

    - Download the new version ltx-2.3-spatial-upscaler-x2-1.0 (https://huggingface.co/Lightricks/LTX-2.3/tree/main)

    PopHorn1956Mar 8, 2026

    It does not work to me. Well, it creates slow mo vids. Like 2x slower movement

    Rifler1Mar 8, 2026

    @PopHorn1956 Which sampler and frame rate? Im using Euler Ancentral and 24 frames.
    Also, maybe you changed some other workflow settings? Maybe even accidentally. Try loading default workflow. Ofc, this is just a noobs assumption))

    PopHorn1956Mar 8, 2026

    @Rifler1 euler, lcm, euler a - does not matter. I double checked all. I used just the wf from zip - q4km etc

    PopHorn1956Mar 8, 2026

    Using Lore Image2Vid-Adapter from LTXV2 hepls. That is weird. Or image compression to 42 for trade of quality. Voodoo

    arkinson
    Author
    Mar 8, 2026

    @Rifler1 Thank you for the hints. Yes updating comfyui + all custom nodes is necessary. I added a hint in the workflow description for the right upscaler model (will fix this in the next update). For more discussions about quality and issues look here too please.

    arkinson
    Author
    Mar 8, 2026· 1 reaction

    @Rifler1 Oh - and thank you for buzzing 😋

    LeskiSTLMar 10, 2026

    @PopHorn1956 Update nodes and comfyui. It works for me.

    blhllMar 8, 2026· 3 reactions
    CivitAI

    12mins for 20 sec I2V, with 8Q instead of 4Q on 3090.

    blhllMar 8, 2026· 1 reaction

    I also get better prompt following with gemma abliterated (heretic). with standard gemma it loses the plot for longer videos. just what works for me.

    arkinson
    Author
    Mar 9, 2026

    @blhll Thank you for your feed back and hints 👍 Do you use the LLM model gemma abliterated (heretic) for the prompt generation inside comfyui or with external software or online? I tried LLM`s inside comfyui several times, but with 12 gb vram it mostly breaks the limits.

    blhllMar 9, 2026

    @arkinson no I generate prompts with LM studio outside of comfy, but I gen prompt with Gemma as well in LM, so when I use prompt then with also but then gemma abliterated, they kinda talks same language, if you know what I mean.

    arkinson
    Author
    Mar 9, 2026

    @blhll Ah, LM Studio. I tried it for myself, but never got well outputs. Do you use a special system prompt to force ltx style prompt outputs? I never dived in very deep yet, but will give it a next try...

    blhllMar 9, 2026· 2 reactions

    @arkinson I created this System Prompt: "Analyze the image and generate text-to-video prompt for a 20 seconds long video based on the given image;

    Strictly follow these guidelines:

    Video length is ~20 seconds. Given image is the first frame of the video, so prompt should continue from that point.

    Core Actions: Describe events and actions as they occur over (20 seconds) time;

    Visual Details: Describe all visual details and be creative with what you want to appear in the video;

    Audio: Describe sounds and dialogue needed for the scene."

    *** then I feed photo I will be making video with, and ask Gemma- here is the image for the prompt. Gemma then uses system prompt as its guidelines and outputs prompt for whatever length in seconds you added in system prompt. increase the temperature if you want prompt to go haliucinating from initial image, say above 0.5.

    (Little hack: If you dont tell Gemma that image is 1st frame, she just describes what is in the image, LTX then randomly creates completely different person during video gen, if video is long, or if in the scene person disapears and then reapears. )

    arkinson
    Author
    Mar 9, 2026

    @blhll Thank you so much. Very interesting. I will try this soon. I did not know, that it is possible to feed images in LM too.

    Btw.: boinobin730 provided a link for a node pack called Easy-Prompt. It generates ltx prompts inside comfyui (even with only 12 gb vram) and you can simply add it to my workflow - so it runs automatically in front of every generation. Installation and usage however, is a little bit tricky and buggy in several parts. If you like, have a look at our discussion here (starts about at the second commend).

    blhllMar 9, 2026

    @arkinson when you are selecting which LLM to use, you need to look for model with VISION capabilty, only those models can process images on top of documents. I will check your link to see if its any better for my workflow, thanks

    arkinson
    Author
    Mar 9, 2026

    @blhll Thank you for the hint with vision too. Your explanations are very helpfull. I did some first quick tests and it works pretty well (even with the "smalest" Q2 model). Very cool stuff.

    If you like to test the easy-prompt node be prepaired: the manual installation is tricky and at the last state i tested it, it was very buggy. I needed several hours to get the models running. T2V works slightly well. For I2V we did not get the vision node running as you can see in the discussion with binobin730. So this stuff is all very experimenting, but interesting of course.

    blhllMar 10, 2026· 2 reactions

    @arkinson yeah I did download his workflow but looks too messy atm, I prefer this wf, and will wait for someone implement those bits, I am too lazy for that :D

    arkinson
    Author
    Mar 10, 2026

    @blhll Yeah, that`s why I warned you. It is not for the lazy boy, but more for "enthusiasts" having fun to ruin their working comfyui system 😂🙂

    Crescend_FlowerMar 9, 2026· 2 reactions
    CivitAI

    I get 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." on the Audio VAE.

    But if I replace LTX23_audio_vae_bf16 with LTX2_audio_vae_bf16, everything works.

    ComfyUI has been updated.

    I generated the video using the promt, lore, and other settings from the video generated in LTX2. The result on LTX2.3 is much better in terms of physics, as if 2.3 understands objects and body parts, but for some reason the video quality is slightly worse. I probably need to experiment with the sampler or models. I used Q8 (Unsloth). I can't comment on the audio, as I used the LTX2's VAE.

    Special thanks to Arkinson for the work done, even though I switched from 3060 to 3090, I still use your workflow.

    Crescend_FlowerMar 9, 2026

    I replaced the "VAELoader Audio" node with the "LTXV Audio VAE Loader" node, and the process started with LTX23_audio_vae_bf16. I had to create a link to the VAE in the "...\models\checkpoints" folder.

    LazerCakeMar 9, 2026

    Update the KJ node and then ComfyUI

    arkinson
    Author
    Mar 9, 2026

    @Crescend_Flower Without any node conflicts, the workflow should work out of the box. As LazerCake allready said, update your KJ nodes and make sure, there are no other conflicts in your comfyui system.

    Crescend_FlowerMar 9, 2026

    @arkinson comfyui 0.16.4, KJNodes 1.3.3

    arkinson
    Author
    Mar 9, 2026

    @Crescend_Flower Yes, I just updated for myself and it works without any issues. I only had to manually enable the KJ nodes after the comfyu update.

    Crescend_FlowerMar 9, 2026

    @arkinson Well, "VAELoader Audio" still doesn't work for me, but it does work if I replace it with "LTXV Audio VAE Loader")

    arkinson
    Author
    Mar 9, 2026

    @Crescend_Flower Ok, I see you have a workaround. But something is wrong in your comfyui installation. Strange thing for me is: your VAELoader Audio don`t work, but VAELoader Video works??? Cause both are KJ nodes.

    Crescend_FlowerMar 11, 2026

    @arkinson "VAELoader Audio" works if I select the "LTX2_audio_vae_bf16" model, but it doesn't work with version 2.3. It might work if I uninstall and reinstall KJNodes 1.3.3, but I don't want to risk it until I have a workaround. =)

    arkinson
    Author
    Mar 11, 2026

    @Crescend_Flower Ah ok. You can test it risk free: Just move the KJ folder somewhere as a backup, restart comfyui and install KJ nodes. If it don`t work, just delete the new folder and move your old folder back. That is sayed without any garanty and warranty 🙂

    DNFYMar 15, 2026

    I was also having the same error. The fix was going into the Comfy Manager > Update all. After that everything worked without issue. ComfyUI v0.17.2 and KJNodes 1.3.4.

    darkdMar 9, 2026· 2 reactions
    CivitAI

    Hi, thanks for sharing your workflow. When generating the video, it seems that on the second pass (the 3-step one), the video becomes completely gray and blurry; I mean, you can't see anything, it's like blurry noise. I'm using img2vid. Do you know why this happens and how to fix it? I have the same graphics card as you, a 3060 12GB. Thanks a lot.

    arkinson
    Author
    Mar 9, 2026· 1 reaction

    @darkd Looks like a model mismatch. Do you use the right ltx-2.3 latent upscale model? The link in the workflow still points to the old model. Use the link published in the first lines here in the model description. Check all other models too.

    CybernixMar 9, 2026· 2 reactions
    CivitAI

    Many, many, many thanks! It works perfectly on RTX4060ti, although I had to update the nodes and comfyui itself

    arkinson
    Author
    Mar 9, 2026

    Thank you so much for your feed back and buzzing and have fun 🙂 I`m glad it works.

    elijahcole001154Mar 10, 2026· 1 reaction
    CivitAI

    How are ya'll getting such speeds? I'm at 5060ti (16gb), 32gb ram, 60-70gb swap and I've been generating the default t2v prompt (edited to be 3 seconds) for 30 minutes now

    blhllMar 10, 2026

    t2v should be even faster than i2v, I edited nothing in workflow, works fast out of the box, granted I'm with 24gb vram.

    arkinson
    Author
    Mar 10, 2026

    @elijahcole001154 There must be something wrong at your side. Please do a clean test: reboot your OS (Windows). Make sure, there are running no other heavy tasks on your machine. Run a 5 second t2v generation with "default" resolution and with only the necessary first Lora activated. Check vram and ram usage and wich process took too much time.

    elijahcole001154Mar 12, 2026· 1 reaction

    @arkinson @blhll Weirdly it finished at around 40 minutes (it was a bad generation though) but the next attempts were hitting the expected speeds. Thank you!

    arkinson
    Author
    Mar 10, 2026
    CivitAI

    @remulocosta453 Hi - thank you so much for buzzing 😋🙂

    NovellusMar 10, 2026· 5 reactions
    CivitAI

    I just said "oh my god" seeing the Model Links tab you provided in the workflow. Thank you soo much for this!!! Having the download links and the locations where to put them, this helps me soo much!! I wish all workflows included this

    arkinson
    Author
    Mar 10, 2026· 1 reaction

    @Novellus Hi - thank you for your feedback 🙂 And yes I really know what you mean from my own experiances. So this was one of the main reasons to publish my own workflows.

    Btw: please note my hint here in the model description for the right upscale model. The link in the workflow still points to the old model.

    Good luck and happy generating 🙂

    NovellusMar 10, 2026

    @arkinson I'm having an issue running it.

    When starting up ComfyUI, I see this:
    This workflow uses custom nodes you haven't installed yet.
    Installation Required
    Install RequiredFinalFrameSelector in subgraph 'New Subgraph'
    Install RequiredJWFloatToInteger in subgraph 'New Subgraph'
    Install RequiredJWIntegerToFloat in subgraph 'New Subgraph'

    When running img2vid, I see this:
    SaveImage 'NoneType' object is not subscriptable

    When I use Manager to install missing custom nodes, I see none to install.

    I see FinalFrameSelector on git DoctorDiffusion/ComfyUI-MediaMixer I Install that missing custom node into the custom_nodes folder, but this message still appears.

    arkinson
    Author
    Mar 10, 2026

    @Novellus Unfortunately, that's all "common" node conflicts at your system you have to solve.

    1. do all necessary updates (see my short Trouble Shooting guide at the workflow).

    2. JW nodes: If these nodes still not installable/working, see the FAQ at my Wan model description for a workaround.

    3. temporarly disable all nodes and reinstall only necessary nodes (see Trouble Shooting), etc.

    Or go the easy way (on windows) and set up a fresh installation just for video generation with Comfyui-Easy-Install (see my Wan workflow description for help). It takes you just a few mouse clicks and round about 30 minutes and you are up and running, including Triton + SageAttention.

    NovellusMar 11, 2026

    @arkinson Okay I got it to somewhat work. It begins generating a video, but then as soon as the upscaler begins, the whole video becomes distorted and early AI looking. I downloaded the recommended ltx-2.3-spatial-upscaler-x2-1.0.safetensors upscaler.

    Also, for a 10s img2vid video, it only takes me 3min to generate. Is this normal on a 4090 with the recommended resolution?

    Edit: Audio comes out great. It's just the video that's distorted or just black.

    arkinson
    Author
    Mar 11, 2026

    @Novellus Ok. I assume you have done all updates, all nodes working and you have no node conflicts anymore. Check twice, you have selected the right models in every loader node. Start easy: t2v, low resolution, 3 seconds, simple prompt and only the first Lora activated. With an error free comfyui system, this should work out of the box.

    InsistentMar 11, 2026· 1 reaction
    CivitAI

    with a 5070ti 16gb and 32gb ram, is it worth to try messing with Q5KM or Q5KS model for ltx 2.3?

    arkinson
    Author
    Mar 11, 2026

    Hi, sorry I have no experiances with higher models for myself. You might have a look here in the comments, but I am not sure, if anyone had tried it yet.

    grbear750611Mar 12, 2026

    I'm using Q6 and Q8 and im getting very fast results on a 4070Ti Super 16GB VRAM and 64GB RAM. 1280x1280 10secs generation in less than 6-7 minutes (image2video) and 20secs almost 9 minutes

    InsistentMar 23, 2026· 1 reaction

    @grbear750611 Thanks, I'll try Q6 or Q8, although I might hit a wall with my 32gb RAM with 100gb page files. :P

    gkirMar 11, 2026· 1 reaction
    CivitAI

    Привет ! спасибо за труды! у меня почему то 1, 2 шаг быстро делает , с 3 шага тормозит ! на ltx2 все летало .

    arkinson
    Author
    Mar 11, 2026

    Hi, sorry I don`got you. What do you mean with steps 1,2, and 3?

    gkirMar 12, 2026

    @arkinson @arkinson What I mean is that when generating steps 1 and 2 the process is fast, but steps 3 and beyond are generated much slower than in LTX 2.0.

    arkinson
    Author
    Mar 12, 2026

    @gkir My question was, what do you mean with steps? We have a two pass generation. 1. pass = 8 steps (first ksampler), 2. pass = 3 steps (second ksampler). Last operation is vae decode (latent to image) and video combine. So I really don`t get about what you are talking.

    gkirMar 13, 2026

    @arkinson , этот переводчик меня не правильно переводит), 8 шагов в первом проходе, так вот первый шаг и второй шаг быстро делаются, 345678 шаги с q4 очень долго делаются. перешел на полную модель , все норм, спасибо за рабочий процесс!

    arkinson
    Author
    Mar 13, 2026

    @gkir The переводчик works fine😉 It just was not understandable, that you talked about a slow down in the first path. Probably a vram issue and something wrong in your system/configuration. Strange thing is, that you say it works with a full model instead. But anyway, I`m glad you got it running. Happy generating 🙂

    gkirMar 13, 2026· 1 reaction

    @arkinson спасибо, у меня 5060 ти 16 гигов, 32 оперативной памяти ! классика

    CybernixMar 11, 2026· 1 reaction
    CivitAI

    In the guide, it's better to change the upscaler to 1.5, because it doesn't produce artifacts. :)

    https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-spatial-upscaler-x1.5-1.0.safetensors

    arkinson
    Author
    Mar 11, 2026

    @Cybernix Thank you for the hind. I will test this soon.

    sneedingonmyligma420Mar 11, 2026· 1 reaction
    CivitAI

    It just works! even better than LTX2 workflows i tried in the past.

    arkinson
    Author
    Mar 16, 2026· 1 reaction

    Uhh - I missed your comment. Thank you for your feedback 🙂

    8htifj2h3jf8239hzf892hnMar 11, 2026· 2 reactions
    CivitAI

    Edit: I'm an idiot and forgot I changed the Lora Name, which has been skipped in return.... IT WORKS! :D

    Hey, thanks for sharing. However I have an issue. I get blurred outputs no matter the prompts. Only thing I’ve changed was clip length to 5s for a test run. I downloaded the upscaler from the description but that didn’t help. 4080S (16GB VRam), 32GB ddr4, 60GB swap file (bazzite Linux). Everything else has been updated to the latest versions.

    arkinson
    Author
    Mar 11, 2026· 2 reactions

    @8htifj2h3jf8239hzf892hn No idiots here, just enthusiasts 😂🙂

    schschMar 11, 2026· 2 reactions
    CivitAI

    [TIPS and REVIEW] Kudos!! With a 3060 (12gb), even 1gb being used by other apps and my FULL HD monitor (yes, you guys know that higher resoltutins TAKES up some VRAM), it worked with a 890x928 picture (irregular size), 10 seconds. I used 32gb RAM, 60gb SWAP.

    LTXAV
    100%|████████████████████████████████████████████████████████████████████████████████████| 8/8 [03:27<00:00

    VIDEOVAE
    100%|████████████████████████████████████████████████████████████████████████████████████| 3/3 [04:11<00:00, 83.79s/it]

    Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
    (Here it takes up to 2 minutes)

    [Prompt executed in 00:12:35]

    BEFORE, I was getting a 'strange mosaic video' in the end (like a surrealistic painting) and it was the fact I was using \latent_upscale_models\ltx-2-spatial-upscaler-x2-1.0 , but the OP said it must be ltx-2.3-spatial-upscaler-x2-1.0 instead.

    You can really make people 'talk' with audio, even with foreign languages. But most of the time, it's unsynced with the mouth. But nothing that a post-work in AVIDEMUX can't do. Also, I dont know how to avoid background music. AI likes the put a music tune in the background.

    I am so far A-MA-ZED with the results! It puts WAN 2.2 'inside the pants pocket'.

    arkinson
    Author
    Mar 11, 2026· 1 reaction

    Thank you for your review 👍

    T2V: speach syncing works very well with most prompts.

    I2V: speach syncing depents heavily on your start image and prompting. Some ideas may work pretty well, others don`t work at all. So, often it is more easy to try something completely different, rather than try to force the AI in one direction.

    "Magic" background music: I never had issues with unwanted music yet. Maybe wrong prompting or a start image wich forces some music??

    Prompting: LTX seems to react very sensible on prompting. Generally try to use correct LTX style prompting.

    haymakerMar 11, 2026· 1 reaction
    CivitAI

    Is it just me or do generated I2V video always become less saturated within the first few frames? I can see the color quickly becoming a little washed out right after the beginning of every video.

    arkinson
    Author
    Mar 11, 2026

    It was a problem with Wan, but I never saw this with LTX.

    haymakerMar 13, 2026

    @arkinson But this does keep happening with my gens. Never happened with LTX-2.

    saurabhwe4u689Mar 17, 2026· 1 reaction

    is there any solution to it it keeps happening to my videos too

    kolompoiMar 11, 2026· 1 reaction
    CivitAI

    Hi,

    Thank you for sharing this amazing workflow.

    I have a problem I haven't been able to solve: in the last two seconds of the video, text and effects appear in the middle of the screen.

    I also hope the next update will include a new option for generating using the first frame and the last frame.

    arkinson
    Author
    Mar 12, 2026

    @kolompoi Hi - thank you. LTX seems often to generate some unwanted text. Just try another seed/prompt/start image.

    First to Last Frame: Yes, I have it allready in mind. Do you have a link to an existing workflow, wich runs with 12 gb vram?

    haymakerMar 13, 2026

    This has happened with my gens too. It shows some random letters in the last few frames.

    arkinson
    Author
    Mar 13, 2026

    @kolompoi Hi, thank you so much for the link 👍 I will try this soon.

    arkinson
    Author
    Mar 13, 2026

    @kolompoi Workflow version v2.0 is out now. Implementing the First/Last Frame part was quite easy and should work technically. I actually have not the capabillities to do larger/more serious test runs. Please try it and let me know if there are any issues.

    kolompoiMar 15, 2026

    I've tried the new version and it's almost complete; it's only missing one workflow, the one that uses ControlNet. As you know, they've released a new version that includes all three options in one file.

    You can find more details at the following link. https://www.youtube.com/watch?v=o7Qlf70XAi8&t=42s

    arkinson
    Author
    Mar 15, 2026

    @kolompoi Thank you for the link. Did you tried the workflow? Actually I´m getting only blurry outputs. Have to check the models.

    ViperimbaMar 12, 2026· 1 reaction
    CivitAI

    any hints how to solve this?
    Error(s) in loading state_dict for LTXAVModel: size mismatch for audio_embeddings_connector.learnable_registers: copying a param with shape torch.Size([128, 2048]) from checkpoint, the shape in current model is torch.Size([128, 3840]). size mismatch for audio_embeddings_connector.transformer_1d_blocks.0.attn1.q_norm.weight: copying a param with shape torch.Size([2048]) from checkpoint, the shape in current model is torch.Size([3840]). size mismatch for audio_embeddings_connector.transformer_1d_blocks.0.attn1.k_norm.weight: copying a param with shape torch.Size([2048]) from checkpoint, the shape in current model is torch.Size([3840]). size mismatch for audio_embeddings_connector.transformer_1d_blocks.1.attn1.q_norm.weight: copying a param with shape torch.Size([2048]) from checkpoint, the shape in current model is torch.Size([3840]). size mismatch for audio_embeddings_connector.transformer_1d_blocks.1.attn1.k_norm.weight: copying a param with shape torch.Size([2048]) from checkpoint, the shape in current model is torch.Size([3840]). size mismatch for video_embeddings_connector.learnable_registers: copying a param with shape torch.Size([128, 4096]) from checkpoint, the shape in current model is torch.Size([128, 3840]).

    arkinson
    Author
    Mar 12, 2026

    @Viperimba Seems like a model mismatch. Check every loader node twice, for selecting the right model. Look at my hint for the right upscaler model too. On the other hand - did you followed my short trouble shooting guide? Everything updated, no node conflicts???

    ViperimbaMar 12, 2026

    thanks! @arkinson, managed to get it done by forcefully updating GG UF node, but the video is tiled and greyish, what can cause it?

    arkinson
    Author
    Mar 12, 2026

    @Viperimba You managed the right upscaler???

    ahegoMar 14, 2026

    Thank you very much. I also encountered this problem. I originally thought it was a very troublesome issue. After reading your comment, I upgraded the gguf node and the problem was solved.

    grbear750611Mar 12, 2026· 3 reactions
    CivitAI

    Thank you very much for the share. Amazing workflow, and extremely fast results!!! 10seconds 4070Ti Super 16GB 1280x1280 image2video+audio in less than 7 minutes

    arkinson
    Author
    Mar 12, 2026· 1 reaction

    Thank you so much 🙂

    grbear750611Mar 12, 2026

    @arkinson i've pushed it to 20 seconds with my own audio. again EXTREMELY fast generation in 9 minutes! on the last second, it pushes some kind of graphic in front of the last few frames. have you encountered something like this ?

    arkinson
    Author
    Mar 12, 2026· 1 reaction

    @grbear750611 You guys with the capable hardware are crazy 😅

    For myself I sometimes notice some text or artefacts anywhere in the video. But that`s all random and not reproducible. Some others also reporting small issues in the first or last frames. Someone suggested to use 1.5x upscaler instead of 2.0x to reduce artefacts. I have not testid it yet.

    grbear750611Mar 12, 2026· 1 reaction

    @arkinson thank you very much :)

    ionelbolintineanu707Mar 12, 2026
    CivitAI

    Working like a charm on 5060ti / 16 Gb and 64 Gb RAM on Linux. Thank you for your work!
    It's only one thing bothering me: a short overlay at the end of the clip, slideshow like, when running V2V. How to solve that?

    arkinson
    Author
    Mar 12, 2026· 1 reaction

    @ionelbolintineanu707 Thank you. Try other seed, prompt, start video. Or use external tools to cut it off.

    Roxikon299Mar 12, 2026· 1 reaction
    CivitAI

    An error appears - CLIPTextEncode

    'Linear' object has no attribute 'weight'. Anyone have a solution?

    Roxikon299Mar 12, 2026· 2 reactions

    Got it sorted - Update KJNodes as per suggestion and had to update Comfyui. Great workflow btw. Thank you so very much!

    arkinson
    Author
    Mar 12, 2026

    @Roxikon299 Thank you, I`m glad you got it running.

    Roxikon299Mar 12, 2026· 1 reaction

    @arkinson Np. Loving the results so far. Thanks again.

    mosagonemad107Mar 12, 2026

    I had to reinstall the whole ComfyUI to make it work after getting this error ! may be I had come version conflicts.

    damoclesongMar 13, 2026· 2 reactions
    CivitAI

    I have an error for this Audio VAE.
    Runtime error : VAE is invalid,

    oh stupid me, just need to update the custom node, and it will works.

    arkinson
    Author
    Mar 13, 2026· 1 reaction

    @damoclesong No stupid one here, but you definately should update all components of your system 😉

    piehound0101723Mar 13, 2026
    CivitAI

    Thanks so much! I do have one question. Is there a way to change the frame rate and number of steps?

    arkinson
    Author
    Mar 13, 2026

    @piehound0101723 The number of frames is automatically calculated based on the clip length for 24 fps for each option.

    If you like to change the framerate you have to edit the values and the calculations in the subgraphs as well as in the video combine node by yourself.

    Generations steps you can change in the subgraphs too.

    Workflows
    LTXV2

    Details

    Downloads
    202
    Platform
    CivitAI
    Platform Status
    Available
    Created
    3/7/2026
    Updated
    4/30/2026
    Deleted
    -

    Files

    ltx23AllInOneWorkflowForRTX_ltx23V10.zip