CivArchive
    [Minimax H3] Anime2Real Experimental (Anime to Realistic) - v2.0
    NSFW
    Preview 139287787
    Preview 139287788
    Preview 139289160

    Just for fun, maybe will have a official image editing model in the future.

    This is an example of Minimax-H3 turing object from style A to B, and you can try any other prompts to convert X style to Y.

    Download single frame VAE: https://huggingface.co/Mamad8/MiniMax-H3-Image-VAE to avoid grid artifact on the image.

    v2.0

    v1.0

    If you like converting anime to realistic photos, please check my other models as well.

    Here is the general purpose prompt for the generation, the more frames (<=24) you generate the better conversion result it will be:

    For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
    
    integrated_multimodal_description: [Shot 1] A completely static shot begins with <object> rendered in pure anime style. Over the full 1-second duration, <object> undergoes a smooth, continuous transformation into a photorealistic live-action human cosplay. Skin texture directly turned to realistic and get realistic pores, subtle imperfections, and natural subsurface scattering. Body proportions kept the same. Stylized anime lighting transitions into natural, physically accurate lighting with soft shadows and realistic highlights. Facial expression and body pose remain completely unchanged throughout. Simultaneously, the entire background environment transforms from anime style into a matching photorealistic setting under the same realistic lighting. The camera remains perfectly locked with zero movement of any kind.
    
    overall_soundscape: N/A
    
    non_diegetic_music: N/A

    V2.0 prompt:

    For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
    
    integrated_multimodal_description: [Shot 1] A completely static shot begins with <object> rendered in pure anime style. Over the full 1-second duration, <object> undergoes a smooth, continuous transformation into a photorealistic live-action human cosplay. Skin texture directly turns to realistic with visible pores, subtle imperfections, and natural subsurface scattering. Moderately retain the original anime’s exaggerated body proportions and stylized curves (slightly enhanced bust, waist, and hip ratios) while converting them into believable real-world anatomy. Stylized anime lighting transitions into natural, physically accurate lighting with soft shadows and realistic highlights. Facial expression and body pose remain completely unchanged throughout. Simultaneously, the entire background environment transforms in perfect visual continuity from anime style into a matching photorealistic setting under the same realistic lighting, ensuring no sudden shifts, discontinuities, or mismatched elements between frames. The camera remains perfectly locked with zero movement of any kind.
    
    overall_soundscape: N/A
    non_diegetic_music: N/A
    
    character traits: .
    
    environment: .

    Description

    reduce inference steps to 5 cut half of the gen time

    FAQ

    Comments (4)

    FantasyAlAug 10, 2026
    CivitAI

    Awesome thanks for sharing this workflow and your prompts, I have tried out a few of your other anime2real workflows and enjoy them. I was wondering what model you think gives you the best results overall?

    Also two things released today that will probably interest you if you haven't seen them already:

    https://www.reddit.com/r/StableDiffusion/comments/1vk2qd3/minimaxh3imagevae_experimental_minimax_h3/

    https://www.reddit.com/r/StableDiffusion/comments/1vjo4qh/kroma_v02_released_chroma_krea2/

    xmsthc
    Author
    Aug 10, 2026· 1 reaction

    Thanks for your comment!

    - I was wondering what model you think gives you the best results overall?

    I think Pornmaster version Krea2 can get the best result for now, except for it randomly add cum on girls.

    - https://www.reddit.com/r/StableDiffusion/comments/1vk2qd3/minimaxh3imagevae_experimental_minimax_h3/

    This one is just the VAE that I'm using in this workflow haha.

    - https://www.reddit.com/r/StableDiffusion/comments/1vjo4qh/kroma_v02_released_chroma_krea2/

    For this one I think the idea is great, but I havent used Chroma before, I will try it after they can show some comparison examples.

    AzzyboiAug 10, 2026· 1 reaction
    CivitAI

    Huh interesting.... youre essentially using Minimax H3 as Flux 2 Klein 9b image editor lol

    Does this work on ref2va?

    I wonder what the inference time is. Do you pass only one frame to AI? Thus you can make 2mp images easily?

    What about concurrent image generation, where you give it one image, and itll give you 24 frames of alternatives, like a story style stuff... man the possibilities are endless xD

    xmsthc
    Author
    Aug 10, 2026· 1 reaction

    - Does this work on ref2va?

    Havent tried ref2va but I assume it will get better result...

    - I wonder what the inference time is.

    5s/iter 1536px 20steps on my 4070s

    - Do you pass only one frame to AI? Thus you can make 2mp images easily?

    Yeah, only one frame, but it can't gen clear images like a specialized image mode does.

    I don't think it will get better at 2mp though I dont have the hardware that can run that resolution lol, the displayed images are upscaled from 0.7mp.

    - What about concurrent image generation

    The frame-level generation may need a LoRA to make it possible.

    Since MMH3 is a multimodal model, I do believe all these features can really be implemented.

    Workflows
    MiniMax H3

    Details

    Downloads
    282
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/10/2026
    Updated
    8/12/2026
    Deleted
    -

    Files

    MinimaxH3Anime2real_v20.json

    Mirrors

    CivitAI (1 mirrors)