CivArchive
    Wan 2.2 Lightning LoRAs - i2v-seko-v1-low-r64
    NSFW

    Extreme speed-up of Wan 2.2 T2V and I2V with just 4 steps

    https://huggingface.co/lightx2v

    Update: 25/10/30

    Update: 25/10/23

    • Added new release 1022 of Lighx2V LoRAs

    • Quality seems improved

    • Possible no ghosting anymore

    • Compatible to other LoRAs

    Update: 25/10/15

    • MoE-Distill is the latest speed-up LoRA

    • Patched version from KJ (HIGH)

    • Restores some motions compared to other speed-up's

    • Quality seems high

    Needs more testing. Some situations are still slow motion.

    Use: Euler+simple, Sigma-Shift 5, CFG 1

    ⚠️ Attention

    • Some configurations seems to blur/ghost

    • It seems incompatible with some LoRAs (e.g. NSFW)

    • It seems incompatible some checkpoints

    Update: 25/08/21

    • More Sampler+Scheduler info

    • ATM: The LoRAs cause a bit slowdown compared to native sampling without

    Update: 25/08/08

    • The I2V Lora-Upload seem to be broken, re-uploaded it again

    • Changed some Usage descriptions

    • Mentioned scheduler/sampler combo

    Update: 25/08/07

    • Added I2V LoRAs

    Update: 25/08/05

    • Added more Infos and LoRAs fix1

    Usage

    SwarmUI

    1. CFG: 1

    2. Steps: 4

    • For T2V Use with this at the end of your prompt: <base> <lora:Wan2.2-Lightning_T2V-A14B-4steps-lora_HIGH_fp16> <refiner> <lora:Wan2.2-Lightning_T2V-A14B-4steps-lora_LOW_fp16> (adapt the lora filenames to whatever specific filenames you have locally)

    • For I2V, use <video> <lora:...i2v-high> <videoswap> <lora:...i2v-low> and of course use the i2v loras

    • the I2V Lightning LoRA appears to target 16 fps

    General:

    • 1 strength on both

    • 2+2 steps (total 4)

    • switch at 2 (50%)

    • Sigma Shift: 5

    • Sampler + Scheduler combinations:

      • Euler + Beta

      • LCM + LTX

      • Euler + Simple

    • You need to be very descriptive, short prompts leading to much more random things.

    • work best at 81x480x832, 81x832x480, 81x720x1280, 81x1280x720

    usage tip first loras (outdated):

    • 0.125 strength on both

    • 4 steps

    • switch at 2

    Did not figured out optimal settings for comfyui or swarmui ... still testing. Let me know! :)

    Generation times <3 minutes

    YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.

    Description

    FAQ

    Comments (15)

    Lora_AddictAug 7, 2025· 1 reaction
    CivitAI

    This is getting so confusing ngl.. what is the difference between your Loras and these ones?
    https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Wan22-Lightning
    In your description you even use these ones in the SwamUI prompt?
    I don't get it, sorry!

    blobby99Aug 7, 2025· 11 reactions
    CivitAI

    WARNING- the discussions on Huggingface show the new LoRAs are broken again. Just use the 'old' 2.1 LoRAs- they work and work well. A lot of well-meaning AI people are stumbling in the dark, and all that matters is good reliable results.

    Using something 'new' for the sake of 'new' is plain silly- the massive community will let you know when a new solution has proven worth using in place of an old.

    BTW, Wan2.2 is a literal miracle, and destroys 2.1 in every way that matters. If you are not seeing this in your use case, you need to to research better workflows and settings. There is a lot of poor advice out there.

    sekaiwlc07860Aug 7, 2025· 1 reaction
    CivitAI

    I HAVE TRY THE NEW I2V...IT CAN`T WORK

    whatsthisaithingAug 7, 2025· 2 reactions
    CivitAI

    Looks like you're using the refiner instead of the video swap, I'm guessing so we can apply the separate loras for high pass and low pass. My head is spinning on the steps settings though.

    It looks like I'm setting 4 steps on the image to video section and NOT selecting the video swap model. That means it will actually run 4 steps on the high pass. Then in the refiner section we're setting control to .5 which seems like it'll only apply 2 steps to the low pass.

    Should we instead be setting 2 steps on the image to video settings, then control of 1 (not .5) on the refiner pass to run 2 steps on low, as well?

    Maybe I'm just missing something.

    delta45424155Aug 7, 2025· 4 reactions
    CivitAI

    How does this compare to using a 5B model instead? Is the difference in quality noticeable?

    nsfwVariantAug 8, 2025· 4 reactions
    CivitAI

    Edit: the I2V files are now fixed.
    I think you uploaded the wrong I2V loras. I couldn't get them to work (blurry output), but then I went and grabbed them from the huggingface source and those work fine.

    FealowAug 8, 2025· 5 reactions
    CivitAI

    4 steps my ass, the new I2V lora just produces a blurry smeared mess on anything below 8 steps. Even then 8 steps still smears and blurs the hands and facial features when moving.

    4598756Aug 8, 2025· 4 reactions
    CivitAI

    I'm looking how to avoida noise seed on the second pass, on usual worflow in the second pass it may be off, but i don"t find this option in swarmui

    Diffusion_4_The_PeopleAug 27, 2025· 3 reactions
    CivitAI
    I'm using the WAN 2.1 lighting version, and in terms of processing time and image quality, it's far better than this Lora. Honestly, I don't know if I'm doing something wrong, but it takes twice as long and the videos come out blurry.

    Ponder_StibbonsSep 5, 2025· 2 reactions
    CivitAI

    A hyperlink takes up less space than copying safetensors. I fail to see the point. Unless the point is to skim porn tokens.

    st3rb3nSep 13, 2025· 2 reactions
    CivitAI

    I do have one question though: I’m running the wan 2.2 GUFF 480p model with a Lightning LoRA (4 steps) to speed things up — a 6-second video takes about 60–70 seconds on an Nvidia H100. The main problem I’m facing is that facial features, especially the eyes, tend to drift or “melt” during generation.

    Do you think there’s a good way to stabilize this? I was even thinking about training a dedicated Low LoRA focused on faces/eyes to keep them sharper.

    smoke07Sep 26, 2025· 2 reactions
    CivitAI

    cool

    frenzzzzy155Sep 28, 2025· 2 reactions
    CivitAI

    Hello! Can anybody share a stable workflow with that lora?

    Sexybeast10Oct 9, 2025· 1 reaction
    CivitAI

    Maybe dumb question but how do I add LTX to scheduler?

    libenlubenschlagenOct 12, 2025
    CivitAI

    This workflow actually works
    https://huggingface.co/lightx2v/Wan2.2-Lightning/raw/main/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1-NativeComfy.json

    LORA
    Wan Video 2.2 I2V-A14B
    by darksidewalker
    Download (Beta)
    tool

    Details

    Downloads
    5,607
    Platform
    CivitAI
    Platform Status
    Deleted
    Created
    8/7/2025
    Updated
    4/27/2026
    Deleted
    4/27/2026

    Files

    Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1_low_noise_model.safetensors

    Size:
    1.14 GB
    SHA256:
    024f21de095bc8fad9809ded3e9e49a2e170dcf27075da8145ba7d60d8aab7f9

    Mirrors

    HuggingFace (132 mirrors)
    CivitAI (2 mirrors)
    ModelScope CN (2 mirrors)
    TensorArt (1 mirrors)
    TensorHub (1 mirrors)

    Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1_low_noise_model.safetensors

    Size:
    1.14 GB
    SHA256:
    8f42b7e1cc2f91cfb7cae7e972e18e051e2125d8070d8b33d715e37242d529d8

    Mirrors

    CivitAI (1 mirrors)