CivArchive
    Stillroom - Krea/MOSS/H3 - H3 1.0
    NSFW
    Preview 148662934
    Preview 148663138
    Preview 148662937
    Preview 148663225
    Preview 148663229
    Preview 148663005
    Preview 148663006
    Preview 148663007

    Quiet rooms, cold light, a woman alone with the afternoon. Stillroom is Krea 2 Turbo with a style merged into its weights: washed, cold color, low natural light, worn European interiors, the texture of a film photo.

    Write the scene. Stillroom brings the rest.

    Use

    • Put the file in models/diffusion_models and load it with Load Diffusion Model, like the official Krea 2 Turbo. Same text encoder (qwen3vl_4b_fp8_scaled, type krea2) and VAE (qwen_image_vae).

    • 8 to 10 steps, CFG 1, euler, simple.

    • Tested at 1024x1280 (4:5).

    • No trigger word. The style lives in the weights.

    • No need to ask for film, grain, palette or mood. The prompts in the gallery describe the scene and nothing else.

    Files

    • fp8 (13.1 GB): runs on a 16 GB card. Same format as the official krea2_turbo_fp8_scaled from Comfy-Org.

    • bf16 (26.3 GB): full precision, for merging and training on top.

    What to expect

    • Wide city squares drift toward Paris.

    Made from

    Original style. Adults only, in the data and in the gallery.

    License

    Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.

    This model was modified: it is Krea 2 Turbo with a style merged into its weights. It is not an official Krea product and is not endorsed by Krea.

    Reuploading this model, hosting it on another site, or training another model on its outputs needs our permission first. For a commercial license, message us here.

    Description

    The same register on MiniMax H3, used as an image model: one frame, cold washed color, low natural light, rooms that have been lived in, merged into H3 FL2VA (pruned).

    • Load it with Load Diffusion Model, like the official minimax_h3_fl2va_pruned_int8_convrot. Text encoder qwen3vl_32b_minimax_h3 (type minimax), VAE minimax_h3_video_vae. The audio VAE is not needed.
    • One frame: in the MiniMax H3 Image To Video node set length to 5 (its minimum, a single latent frame), decode, and keep the first image with Image From Batch (index 0, length 1).
    • 20 steps, res_multistep / simple, BasicGuider (no CFG). Tested at 1024x1280.
    • Write the prompt in H3's own format, with the scene after [Shot 1] Live-action.:
      integrated_multimodal_description: [Shot 1] Live-action. A woman sits by a kitchen window holding a cup of tea.
      overall_soundscape: N/A
      non_diegetic_music: N/A
    • No trigger word. Write the scene; the style lives in the weights.
    • int8 convrot (21 GB): same layout as the official int8 file. About 90 seconds per image on a 16 GB card.
    • The cold wash shows best in daylight and window light. Night scenes under warm lamps stay warm.
    • The style was trained on stills and tested as an image model. Video with this file is untested.

    On .red

    It leans toward undress: "topless" came out fully nude in testing. Name the clothes that stay. Describe an adult body in plain words: a fit, toned body, a defined waist, full natural breasts. "Toned" can tip into gym abs; drop it if the waist gets hard.

    MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. This model was modified: it is MiniMax H3 FL2VA (pruned) with a style merged into its weights, quantized to int8. It is not an official MiniMax product and is not endorsed by MiniMax. The license text and the notice come with the files of this version. Every image in the gallery carries its prompt and settings.

    FAQ

    Checkpoint
    MiniMax H3

    Details

    Downloads
    234
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/11/2026
    Updated
    11/16/2026
    Deleted
    -

    Files

    stillroom-h3-license.zip

    Mirrors

    CivitAI (1 mirrors)

    stillroom_h3_image_v1_int8_convrot.safetensors