CivArchive
    godago - v1.0
    NSFW

    v30

    The v10 and v20 felt a bit too plain, so I created the v30 with a more intriguing touch. I’m happy with it. Enjoy.

    ====================

    v20

    I refined v20 from v10 to make it more flexible when used with LoRAs.

    In the comparison image, the left is the base krea2 and the right is v20.

    This time, I have uploaded fp8, int8, and nvfp4 versions.

    ====================

    v10

    This doesn't seem particularly special

    Feel free to use the sampler and scheduler

    The list includes the NVFP4 file and the INT8_CONVROT file

    Description

    FAQ

    Comments (20)

    ferrrett33Sep 19, 2026· 7 reactions
    CivitAI

    "This doesn't seem particularly special"

    Then why post it? Use Hugging Face for free personal storage of AI models.

    SH_zeroSep 19, 2026

    Thanks for the advice, my boy. I’ll give Hugging Face a shot as well.

    hushichoSep 19, 2026· 3 reactions

    I'm glad this was posted. If it was yet another boob-centric anime girl model, maybe I wouldn't care, but this has a distinct style and concept. OP, please don't be driven to post less by the person who can't spell ferret properly.

    PartisanoSep 19, 2026

    33imbecccile "ferrrett33"

    rommix0Sep 24, 2026

    Why the heck do you care? Maybe other people like it, and it's a fact the world doesn't revolve around you. Think before you say stupid crap like that next time. People will like you more. Take care.

    ferrrett33Sep 26, 2026· 3 reactions

    @rommix0 Nah, you're wrong. Everyone can do whatever they want with AI, create their own loras, merge checkpoints. If you're going to post something in public, that's the next level. It has to be worth using for hundreds or thousands of people otherwise you're wasting hundreds or thousands of people's time and compute resources. It's not about whose brave enough to post something, if you want I'm brave enough to post some total garbage checkpoint for you and then waste your time.

    SH_zeroSep 26, 2026

    @ferrrett33 Wise words, my boy. Now stop wasting your time and go on with your life or keep hiding It's honestly better that way.

    ferrrett33Sep 29, 2026· 1 reaction

    @SH_zero I know bro, you're Chinese or some other nationality who doesn't speak English very well so you can't write a description in English, but we can help you. Just tell me in your own words what this model does and I'll help you with the English. Thanks

    SH_zeroSep 29, 2026

    @ferrrett33 omg.. You sound like a loser.

    ferrrett33Sep 29, 2026

    @SH_zero Nah you're a korean who doesn't understand how this site works and joined a week ago. I'm just trying to help you with your English. lol

    UnicomSep 19, 2026
    CivitAI

    BF16 please

    UnicomSep 19, 2026

    or at least FP8, because not everyone has Nvidia graphics cards.

    Mr_FeiSep 20, 2026
    CivitAI

    What style is this? Comic or anime style or realistic style?

    2TheMaxSep 20, 2026· 1 reaction
    CivitAI

    The style is the incredible!!! Masterpiece!!

    everylightSep 20, 2026
    CivitAI

    This is a great model. Could you please share FP16 or BF16?

    fjallSep 20, 2026
    CivitAI

    can you provide samples without lora?

    MirabilisSep 21, 2026
    CivitAI

    Hi, just a heads-up that the NVFP4 version appears to be broken.

    I tested it in Forge Neo and found that 76 of the 224 .comfy_quant tensors are completely zero-filled rather than containing the expected NVFP4 JSON metadata.

    The packed weights plus weight_scale and weight_scale_2 tensors are present. I reconstructed all 76 missing NVFP4 markers and got the model to load successfully in Forge Neo, but generations still produce pure noise.

    So it looks like the problem goes deeper than the missing metadata and the NVFP4 checkpoint probably needs to be re-exported/re-quantised from the original high-precision model.

    Hope that helps track it down.

    SH_zeroSep 21, 2026· 1 reaction

    Thanks for the heads-up and detailed report!

    I re-downloaded the uploaded checkpoint directly from the host and thoroughly inspected the tensor structure with a Safetensors inspection script:

    Total tensors: 1,102

    Quantized layers: 224

    Valid .comfy_quant metadata: 224 / 224 (0 zero-filled)

    Corrupted scales: 0

    The checkpoint loads and generates images cleanly in official ComfyUI with native NVFP4 ops without any pure noise issues.

    Given this, it seems likely that one of two things happened:

    Download corruption: The file might have been corrupted or truncated during download. Could you verify the SHA256 hash of your local file?

    SHA256: 37c86800f4

    Forge Neo compatibility: If the hash matches, the issue might lie in how Forge Neo's current loader/dequantization kernel parses ComfyUI-format NVFP4 metadata and scales.

    Could you check your file hash or try loading it in ComfyUI to see if the issue persists? Thanks again for taking the time to test!

    MirabilisSep 21, 2026· 1 reaction

    @SH_zero No worries. I think it must have gotten corrupted during download by the looks of it (Civit seems to be increasingly flaky I find) as I didn't get a matching SHA. I'll try downloading it again and see if I get better luck. Thanks,.

    Grief_Sep 21, 2026· 3 reactions
    CivitAI

    did your last account get blasted?? lol

    Checkpoint
    Krea 2

    Details

    Downloads
    208
    Platform
    CivitAI
    Platform Status
    Available
    Created
    9/19/2026
    Updated
    11/20/2026
    Deleted
    -

    Files

    int8_convrot.safetensors

    Mirrors

    CivitAI (1 mirrors)

    nvfp4_convrot.safetensors

    Mirrors

    CivitAI (1 mirrors)