CivArchive
    Breast Size Slider — MiniMax H3-Image - v2.0
    NSFW
    Preview 149690600
    Preview 149690601
    Preview 149690604
    Preview 149690615
    Preview 149690638
    Preview 149690650
    Preview 149690656
    Preview 149690665
    Preview 149690675
    Preview 149690678
    Preview 149690681
    Preview 149690689
    Preview 149690694

    A signed LoRA slider for changing apparent breast volume in MiniMax H3-Image. It works without breast-size or breast-shape adjectives in the prompt, so you can keep the scene description neutral and let the LoRA supply the change.

    What's new in v2.0

    Users reported that v1.0 at high weights changed the face. We measured it, and they were right: at ±1 and ±2 the face, and sometimes the whole figure, moved with the size. v2.0 is a full retrain with a different method, built to change the chest and leave the rest of the image alone.

    • Localized change. Pose, framing, hands, clothing folds and background stay put across the range. Compare the two same-seed ladders in the gallery.

    • Faces stay closer. We measured face similarity (ArcFace) against the weight-0 image on four test women, with the same seed in every frame. The mean was 0.73 at ±1 (v1.0: 0.60) and 0.57 at ±2 (v1.0: 0.53). For reference, adding the sentence "She has very large, full breasts." to the prompt, with no LoRA at all, moved the same faces to 0.37.

    • Full, even range. It runs from nearly flat at -2 to very large at +2, and the steps in between are even.

    How to use v2.0

    • File: Novarchi_H3Image_BreastSizeSlider_v2.safetensors. No trigger words.

    • Positive weights make the breasts larger, negative weights smaller. Start between -1.5 and +1.5. The tested range is -2 to +2, and the endpoints are deliberately strong.

    • Keep the prompt neutral about size and let the slider do it. For a clean comparison, keep the prompt, seed, sampler, resolution and every other LoRA fixed, and change only this LoRA's weight.

    v2.0 showcase

    The two v1.0 comparison scenes, re-rendered with v2.0: a fitted opaque ribbed top, and a topless studio profile with the arms behind the back. Same prompts, weights -2, -1, 0, +1, +2, one shared seed for every frame (not v1.0's seed, which is outside the current graph's seed range). The prompts do not name breast size or shape. The cover is the topless scene at +1.5. All images are 3:4, 3552 × 4736 (16 MP), MiniMax H3-Image FP8 scaled with LCM, 6 steps and CFG 1. v1.0's showcase used Euler, so compare the two versions by shape and stability, not skin texture. Full prompt, seed and LoRA metadata are embedded in every PNG.

    v2.0 training

    • Concept slider trained from the prompt difference "larger" vs "smaller", with a chest mask. The direction is learned only inside the chest region. Everywhere else, an extra loss holds the prediction to the base model at both signs, and the face region gets three times that weight.

    • 447 real photographs as training images, with chest masks and face masks. Rank 8, learning rate 7.5e-5, guidance 2, AdamW 8-bit, trained on the FP8 base. 1500 steps (about 3.4 epochs); this release is the final checkpoint.

    v2.0 limitations

    • At +2 the belly, hips and face can still get a little fuller, and the areolae grow larger and darker. At -2 the chest goes nearly flat and the ribcage can show. MiniMax H3-Image couples these with breast size. The same happens, more strongly, when you ask for size in the prompt. A fully identity-locked size change is not something we could get from this base model.

    • This slider controls volume, not a specific shape. Shape, softness and placement still come from the base model, the prompt, the pose and any other LoRAs.

    • Not validated: men, character LoRAs, other body or breast-shape LoRAs, image editing, and weights beyond -2 / +2.

    v1.0

    How to use v1.0

    • Positive weights make the breasts larger.

    • Negative weights make them smaller.

    • Start around -1 or +1, then adjust in small steps.

    • The included comparison packs cover the full tested range from -3 to +3. The endpoints are deliberately strong.

    • No trigger words are required.

    For a clean comparison, keep the prompt, seed, sampler, resolution and every other LoRA fixed. Change only this LoRA's weight.

    v1.0 showcase

    The published showcase contains two same-seed ladders with seven weights each:

    • a fitted opaque ribbed top;

    • a topless studio profile with the arms behind the back.

    All fourteen frames use the same seed and the same H3-Image settings: 6 steps, CFG 1, Euler/simple, 3:4 at 16 MP. The prompts do not name breast size or breast shape. The separate cover image uses weight +2 and a fresh seed.

    The release checkpoint is step 800 from the h3_breast_size_slider_mid_v4_3000 run. Step 800 was selected instead of the final checkpoint because later training added more body and pose leakage without giving cleaner control.

    v1.0 limitations

    The response is not perfectly linear. Positive weights are generally stronger than matching negative weights. At high strengths the LoRA can also influence pose, torso angle, framing or nearby clothing folds. In the topless showcase, later positive weights shift slightly from strict profile toward a three-quarter view.

    This slider controls volume, not a specific breast shape. Shape, placement and softness still depend on the base model, prompt, pose and any other LoRAs in the stack.

    The published package does not test men, character LoRAs, anatomy or breast-shape LoRAs, reference-image editing, or weights beyond -3 and +3. Those combinations may work, but they are not validated here.

    v1.0 recommended settings

    • Base model: MiniMax H3-Image

    • LoRA: Novarchi_H3Image_BreastSizeSlider_v1_step800.safetensors

    • Starting range: -1 to +1

    • Tested range: -3 to +3

    • Trigger words: none

    • Sampler used for the showcase: Euler / simple

    • Steps used for the showcase: 6

    • CFG used for the showcase: 1

    License

    This LoRA is a derivative of MiniMax-H3-Image and is distributed under the MiniMax-H3-Image Community License Agreement, with no additional or different terms. A full copy of the license and the NOTICE file are in Novarchi_H3Image_BreastSizeSlider_v1_License.zip, attached to v1.0.

    Each version carries its own copy of the license and NOTICE: v1.0 in Novarchi_H3Image_BreastSizeSlider_v1_License.zip, v2.0 as LICENSE.txt and NOTICE.txt.

    MiniMax-H3-Image is licensed under the MiniMax-H3-Image Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

    By downloading or using this LoRA you agree to that license, including its use restrictions (Section V) and its Acceptable Use Policy (Exhibit A). In particular: the license grants rights only outside the European Union, the United Kingdom, the Republic of Korea and the United States; neither the LoRA nor its outputs may be used to improve any other AI model; and images you publish must be clearly marked as AI-generated. All showcase images on this page are AI-generated with MiniMax-H3-Image.

    Listed under Civitai's "MiniMax H3 Image" base model. The license that governs this LoRA is the MiniMax-H3-Image one above, not the MiniMax H3 (video) license.

    Powered by MiniMax-H3-Image.

    Created by Novarchi.

    Description

    A retrained, localized slider. The change is confined to the chest: pose, framing, hands, clothing and background stay where they are across the range, and faces stay noticeably closer to the weight-0 image than with v1.0.

    Positive weights make the breasts larger, negative weights smaller. Start between -1.5 and +1.5. The tested range is -2 to +2. No trigger words are needed.

    This is the final step-1500 checkpoint. In our same-seed tests it held faces better at ±1 than the step-800 and step-1100 checkpoints, at the same strength of effect.

    FAQ

    LORA
    MiniMax H3 Image

    Details

    Downloads
    552
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/21/2026
    Updated
    11/22/2026
    Deleted
    -

    Files

    NOTICE.txt

    Mirrors

    CivitAI (1 mirrors)

    Novarchi_H3Image_BreastSizeSlider_v2.safetensors