CivArchive
    WAN2.2-14B-Rapid-AllInOne workflow with lora & color match - flf2v_v1.0
    NSFW

    <flf2v_ver1.0>

    This is a First-Last Frame to Video compatible workflow.

    <ver1.1>

    In this workflow, noise appears in the first frame of the video, so we used the “Select Image” node to cut the first frame.


    <ver1.0>
    Adding Lora nodes and color match nodes to the WAN2.2-14B-Rapid-AllInOne workflow.

    https://huggingface.co/Phr00t/WAN2.2-14B-Rapid-AllInOne

    Description

    This is a First-Last Frame to Video compatible workflow.

    If the first and last frames are not the same, activating the "Select Images" node and cutting out around the first 5 frames will stabilize the video.

    FAQ

    Comments (7)

    blobby99Aug 31, 2025· 6 reactions
    CivitAI

    WARNING- if you are using scaled FP8 for your clip (the LLM), change it to bf16 or a GGUF. I literally could not believe it when in A-B testing I discovered scaled FP8 was responsible for ignoring prompt details. Months wasted, because this model format is useless (on Blackwell at least- it is possible recent Comfy so-called Blackwell optimisations are to blame, but I'm not going to roll back Comfy to check).

    With Wan2.2 istv, GGUF Q8 is infinitely better than the scaled FP8 main model (and then as I said moving clip to GGUF gives more improvements again). This will apply to Wan2.2 i2v and t2v as well.

    Light6969Sep 1, 2025

    Interesting so you think it makes a big different with nromal 2.2 workflows too? What kind of difference are we talking about, like you tell the characters to walk right and it doesn't listen and now with the change it listens?

    panjoolSep 1, 2025

    how about the wan model itself, do u see better results from q8 compare to fp8 scaled?

    eshekieriSep 7, 2025

    Bro you just solved my headache with one comment

    Light6969Sep 7, 2025

    @eshekieri What model did this info help you with?

    kevinpawseyNov 24, 2025

    @Light2020462 probably all, because @blobby99 is not talking about the model, he is talking about the clip llm, so the "umt-xxl..." model. It does make sense what they are saying, I just noticed that a new workflow I am playing with is using fp8_scaled, so I am going to try switching. I might even try the GGUF Q8, as that is just about the original quantized model, but in memory-friendly layers.

    Thank you @blobby99 for this comment!

    Fil_Is_HereSep 15, 2025· 1 reaction
    CivitAI

    Funny thing, the first and last frame do not match the ones I set. It's so weird, I don't even understand the problem in theory. The result is a bit similar to my frames, but hmm... It seems like it shouldn't work like that.

    Workflows
    Wan Video 2.2 I2V-A14B

    Details

    Downloads
    2,993
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/31/2025
    Updated
    5/13/2026
    Deleted
    -

    Files

    wan2214BRapidAllinone_flf2vV10.zip

    Mirrors