CivArchive
    akira style embedding - V5
    NSFW
    Preview 70065
    Preview 70064
    Preview 70063
    Preview 70062
    Preview 70061
    Preview 70060
    Preview 70059
    Preview 70058

    PLEASE comment the stuff you make with this, I really wanna know what you end up doing.

    mostly trained on sd1.5 pruned

    it likes to generate pretty, jumbled messes, but that's kinda the movie. very grimy outputs.

    but also I didn't try that hard for the example photos so maybe you can do better than me.

    also people generated are usually screaming, that's another issue with the movie, everyone is usually screaming.

    anyways I trained it on like 30 images from the movie that I thought looked cool, or filled in gaps that the others left. captioned the images by hand, idk if it's better than using the ai but whatever.

    trained on 424x424 images because I'm using a gaming laptop for the training lol,,, any bigger and it crashes

    6 vectors per token I think ? idk

    anyways I think this embedding is really cool and you should try it, works pretty well for img2img type stuff too

    Description

    trained more on images with faces,, got rid of images that sorta looked like they were cropped, didn't fix the cropping problem, but it's better now. I think I did like 300 steps on hassanblend but idk

    FAQ

    Comments (1)

    KlokinatorFeb 3, 2023· 1 reaction
    CivitAI

    I've tested your embeds and they don't seem very good on any of the models I've tried. Then I saw that you were using Hassanblend to train this embed.

    Akira is an anime, and you should be using an anime styled model if possible. Hassanblend is fine for creating realistic women, but no offense to the creator, it's a bit sloppy in how things are mixed and is not good for training embeds. One other problem is that it has a ton of floating point errors you probably overlooked and those will translate to worse embeddings.

    I recommend using Anything 3, NAI, or you can even try one of the models I've recently put out. I think you'll find that the Dutch models, which are custom mixed and trained via LORAs, are extremely embed friendly. I recommend using either the Macross Embed model, or you might try the Olympus model. Both are very versatile.

    https://civitai.com/models/6077

    This might seem like shilling, but I would LOVE to see better embeddings on Civtai and I'm just trying to help. Akira is also a god-tier anime, and if we can somehow replicate its visual style, I'd be ecstatic!

    TextualInversion
    SD 1.5

    Details

    Downloads
    574
    Platform
    CivitAI
    Platform Status
    Deleted
    Created
    2/3/2023
    Updated
    5/27/2026
    Deleted
    4/27/2026
    Trigger Words:
    akira5

    Files

    akira5.pt

    Mirrors

    HuggingFace (2 mirrors)
    CivitAI (1 mirrors)

    Available On (1 platform)

    Same model published on other platforms. May have additional downloads or version variants.