CivArchive
    LTX 2.5 - dev
    NSFW
    Preview 139435338

    LTX-2.5 is an open world model with open weights, built for local execution and fine-tuning. Its established use is generating synchronized, high-fidelity video and audio from text, image, and video inputs; applicability to emerging domains such as robotics and physical AI is developing.

    Full control and customization — self-host on your own infrastructure. No per-generation billing, no per-seat lock-in, no forced API dependency. Revenue is measured across the whole entity, including subsidiaries and affiliates under common control. The full, binding terms live in LICENSE.

    What's new in LTX-2.5

    • Native multishot generation — generate connected scenes in a single pass: multiple shots that hold character identity, environment, lighting, voice, and visual style across cuts (previous versions produced a single continuous shot).

    • Diffusion fidelity rendering — Instead of locking every scene to one compression rate, our model dynamically allocates compute by scene complexity and budget, rendering flawless detail where it matters, efficient everywhere else.

    • New diffusion video decoder — replaces the VAE reconstruction stage; sharper faces, textures, and on-screen text, better motion, and fewer artifacts in demanding scenes.

    • Custom Gemma 4 12B text encoder — holds complex prompts together (multiple characters, camera moves, lighting, actions) instead of dropping details across a longer sequence.

    • Prompt enhancer — expands a short prompt into richer cinematic instructions at minimal extra compute.

    • Duration predictor (optional) — an opt-in node predicts a clip's length from the prompt and sets the frame count for you, instead of relying on a fixed-duration parameter.

    • Substantially improved distilled model — retains much more of the full model's visual quality, prompt adherence, and motion consistency in a smaller, faster checkpoint.

    Description

    FAQ

    Comments (34)

    BzzzDarklordAug 11, 2026· 39 reactions
    CivitAI

    Who cares...

    JiBareinAug 12, 2026· 11 reactions

    Literally everyone who want to try new models

    BzzzDarklordAug 12, 2026· 8 reactions

    @JiBarein nobody cares what YOU want

    pezpok1929Aug 12, 2026

    i'll finish your sentence off for you ...what the price of fish is on a rainy Wednesday afternoon in Chicago?
    So you were asking a question. Well, the person/people who are looking to buy the fish on a rainy Wednesday afternoon in a fish shop in Chicago would care. Unsure what your question has to do with the model. But glad i could help clear up your question.

    elevendrAug 12, 2026

    What video model do you care about? Minmax H3 or Seedance 2.0

    fauxtographyproAug 12, 2026· 3 reactions

    What is the charge?! Eating a meal? Eating a succulent Chinese meal?! THIS IS DEMOCRACY MANIFEST!

    fierceysAug 12, 2026

    Sorry that your life is terrible

    Noone3r0f3890490Aug 11, 2026· 1 reaction
    CivitAI

    Very cool!!!

    watsont474Aug 11, 2026· 1 reaction
    CivitAI

    Thanks, I'll give this a try.

    aomyouza2543Aug 12, 2026· 3 reactions
    CivitAI

    it's a bug-fix patch and other corrections, huh? Truly lives up to being an "open world model AI" Lol 😂

    orzechowy3334318Aug 12, 2026· 6 reactions
    CivitAI

    Image 2 Video - super consistent. Thank you! Congratulations!

    civitaxianAug 12, 2026

    A little bit better than 2.3 but not that good in my experience.

    mcyhmen757567Aug 12, 2026· 1 reaction
    CivitAI

    is 2.3 Loras works in 2.5 ? if yes, for the better ? or the same ?

    orzechowy3334318Aug 12, 2026

    Works better. Think as if to your car they assembled improved engine.

    j_cubeAug 13, 2026

    2.3 furry+helper together work great

    JoeyDiazAug 12, 2026· 8 reactions
    CivitAI

    Now the real question, Does it beat Minimax H3? From what I've seen already absolutely not, But Ltx in has always been a very easy model to fine tune and improve, So we'll see what the community can turn it into

    WallyhojoAug 12, 2026

    LTX is faster, but it struggles to accurately render new layers or elements that aren't already present in the source image when doing I2V.

    JoeyDiazAug 13, 2026

    @Wallyhojo Yea after using it for a full day I can confidently say it's not touching minimax at any capacity but the speed is absolutely mind bugling to me

    orzechowy3334318Aug 12, 2026· 7 reactions
    CivitAI

    I spent a few hours of testing denoising parameters. The winner for distilled model IMHO is:
    Sampler: Euler_ancestral
    Scheduler: beta
    Steps: 8
    I tested more then 8 steps, but more steps can be an improvement if we could "reduce" distill weigths - I mean in "dev + lora" case.

    JunkieMonkey69Aug 12, 2026

    Im using the Dev+ distilled LoRA setup. How do i increase my step? can you please guide me here?
    i can only see Manualsigmas
    1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.725, 0.421875, 0.0

    putratroopsAug 12, 2026

    hi can i ask what is the diff beetwen the distilied or dev,, i using low vram btw, hope you can give answer?

    orzechowy3334318Aug 12, 2026

    @JunkieMonkey69 You have to replace "manual sigmas" node with "BasicScheduler" node.

    orzechowy3334318Aug 12, 2026

    @putratroops "Dev" model means full model designed for slow but precise denoising process. It requires lot of steps (takes time). Distilled model is "compressed" model requires few times less steps (faster) with a LITTLE worse result. Speaking about VRAM "distilled - dev" is less important. Important is what quantization the model has. Making the thing simple - bigger weigths file -> better results, more VRAM.

    putratroopsAug 12, 2026

    @orzechowy3334318 noted thank you smuch for the explanation

    orzechowy3334318Aug 12, 2026

    @JunkieMonkey69 You can also add numbers to the string from 1 to 0 be aware of format . ,

    RandoTheBarbarianAug 12, 2026
    CivitAI

    Curious to see how this model compares to MiniMax. 2.3 didn't really hold a candle to it.

    civitaxianAug 12, 2026· 6 reactions
    CivitAI

    MiniMax H3 is still much better. Much easier to prompt minimax, does what i tell it to do. Hopefully this will be fixed in LTX, because its much faster than minimax.

    virtyos07Aug 13, 2026· 4 reactions

    minimax is an useless shit

    chamo9009Aug 13, 2026

    @virtyos07 I support this guy, I've done everything I can to make it work and all my fucking generations are black

    BigSad11Aug 13, 2026

    @chamo9009 sounds like you have some built in safety filters or something because that's NOT normal.

    chamo9009Aug 13, 2026

    @BigSad11 You think? Use in8 version, use the right components, comfyui branches for Mini Max, cloned and updated nodes, cross Pytorch attention, vae in FP32, I throw everything and it doesn't work! I already despaired, because it doesn't work and I've been trying to understand why for 6 days. My error is a Runtime warning Invalid values found in cast or something like that

    bungelAug 13, 2026
    CivitAI

    The speed and image quality is impressive to say the least, but i think the 2.3 quirks are still there

    chamo9009Aug 13, 2026
    CivitAI

    For me, this is better than Mini max in the aspect of profitability.

    StrangerFogAug 13, 2026· 1 reaction
    CivitAI

    Yes, it is faster than Minimax, but the understanding of physics is certainly worse here, the understanding of Prompt is also worse, the speech has become better, but Minimax made a huge impression on me; I am still more willing to wait a long time for generation for a good result.

    Checkpoint
    LTXV 2.5
    by CivitaiOfficial
    Download (Beta) View on CivitAI
    base model

    Details

    Downloads
    532
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/12/2026
    Updated
    8/14/2026
    Deleted
    -

    Files

    ltx25_dev.safetensors

    Size:
    20.03 GB
    SHA256:
    2edbdb4465cd6c3b532cd67a31ddb38a63e97dcad20be3729675e2a4e8caf92b

    Mirrors

    HuggingFace (3 mirrors)
    CivitAI (2 mirrors)

    ltx25_dev.safetensors

    Size:
    249.81 MB
    SHA256:
    2bc3300f2b3c3c1834d72164fbf13a3b9fd73e5a741e8a2c3f4035f89a75c3fe

    Mirrors

    HuggingFace (133 mirrors)
    CivitAI (2 mirrors)
    ModelScope CN (1 mirrors)

    ltx25_dev.safetensors

    Size:
    39.13 GB
    SHA256:
    792a2bad501ca03262c0bc2ce7a2949e85b142ce18e30894aad5bc849c8e7584

    Mirrors

    HuggingFace (6 mirrors)
    CivitAI (1 mirrors)

    ltx25_dev_txt.safetensors

    Size:
    14.32 GB
    SHA256:
    09a89e084de1a149c3de60cfe9dfd3e5161967eb09eea39e806fcdeffdd568de

    Mirrors

    HuggingFace (5 mirrors)
    CivitAI (3 mirrors)

    ltx25_dev.safetensors

    Size:
    949.65 MB
    SHA256:
    eb5a71fe4068ee87ccdb1c3aa635e547ca76bd2d30ae20ae889f2c325c0677e8

    Mirrors

    HuggingFace (11 mirrors)
    CivitAI (2 mirrors)

    ltx25_dev.safetensors

    Size:
    8.29 GB
    SHA256:
    86370bbf79a9eb4edaa158907e2b48a5188fe4c5dc8ce30c7eb8f2f131a9bbf5

    Mirrors

    HuggingFace (9 mirrors)
    CivitAI (2 mirrors)