CivArchive
    Natural Breasts — MiniMax H3-Image - v1.0
    NSFW
    Preview 149052703
    Preview 149052748
    Preview 149052747
    Preview 149052758
    Preview 149052762
    Preview 149052782

    A LoRA for MiniMax H3-Image that gives breasts a natural, believable shape: soft weight, gentle sag, fuller lower curves, a smooth transition into the chest and slight asymmetry instead of high, round, implant-like geometry. It works under thin fitted clothing as well as on bare skin.

    What's new in v2.0

    v2.0 is a full retrain on a much larger, more varied set: 561 curated photographs instead of 119. They span small to very large sizes and many ethnicities, poses and lighting setups. Each image has a detailed caption describing breast size, shape, weight and real skin texture. Faces are masked out of the loss, so the LoRA learns the body and leaves faces to the base model.

    • More natural geometry with less prompting. Soft weight, fuller lower curves, slight asymmetry and real areola and nipple detail come from the LoRA itself. In our A/B tests the prompts named only the size, with no shape words, and v2.0 still moved the base away from high, round shapes.

    • Gravity reads better in forward bends, side-lying poses and under thin, braless tops.

    • Size follows the prompt. "small" stays small (the base tends to inflate everything), and "medium" stays medium.

    • No plastic skin. Skin keeps the base model's texture at every checkpoint we tested.

    How to use v2.0

    • File: Novarchi_H3Image_NaturalBreasts_v2_step1750.safetensors. Weight 1.0. No trigger words.

    • Write the size you want (small, medium, large, very large). v2.0 reads "large" a little more modestly than the base. If you want them big, say very large or lower the weight toward 0.7–0.8; the lower weights are untested.

    • For real skin, describe it plainly ("real, unretouched skin with visible pores") and avoid words such as "cover" or "magazine", which pull H3-Image toward airbrushed gloss.

    v2.0 showcase

    The six v1.0 showcase scenes, re-rendered with v2.0 at weight 1.0. The prompts are the v1.0 prompts, kept word for word except that "cover photograph" became "photograph", so they still describe the natural shape in words. Three images use the v1.0 seed and three use a new one. All are rendered at 3:4, 3552 × 4736 (16 MP), on MiniMax H3-Image FP8 scaled with LCM, 6 steps and CFG 1. v1.0 used Euler, and LCM renders micro-texture a little smoother, so compare shape between the two versions, not skin. Full prompt, seed and LoRA metadata is embedded in every PNG.

    v2.0 training

    • 561 curated photographs, buckets 1024 / 1280, face masks (masked areas weighted 0, inverted mask prior 0.5).

    • LoRA rank / alpha 16 / 16, learning rate 2.5e-5, AdamW 8-bit, EMA 0.995, frozen text encoder, trained on the FP8 base.

    • Run length 3000 steps. This release is the step-1750 checkpoint, about 3.1 epochs.

    v2.0 limitations

    • "large" comes out a bit smaller than with the base model, and figures trend slimmer, with more visible ribs and abs.

    • Lying on the back is still the weakest pose: the breasts can stay rounded and centred instead of spreading sideways.

    • H3-Image tends to finish half-done clothing states, so a blouse "pushed down to the elbows" may end up fully on or fully off. This happens with or without the LoRA.

    • Stacked with our Asian Beauty LoRA it works cleanly (tested at 1.0 + 1.0 in one scene), but both slim the figure, so breasts come out smaller. Other weights, men, character LoRAs, other body LoRAs and image editing were not validated.

    v1.0

    How to use v1.0

    • Recommended weight: 1.0. Every published image uses this weight.

    • No trigger words are required.

    • Describe the size you want in the prompt (for example large natural breasts). This LoRA shapes the result; it is not a size slider.

    • For the most natural result, keep the shape language in the prompt soft as well: soft weight, gentle sag, full lower poles, slight asymmetry. Avoid words such as perky, round or push-up, which pull the model back toward its default geometry.

    v1.0 showcase

    Six images, each a fresh seed at weight 1.0: a hotel bathroom, a sunlit bedroom, a seaside terrace in a thin ribbed tank top with no bra, a rainy night window, a sports cheerleader and a sunset terrace cover.

    All frames: MiniMax H3-Image FP8 scaled, 6 steps, CFG 1, Euler / simple, 3:4 at 3552 × 4736. Full generation metadata (prompt, seed, workflow, LoRA) is embedded in every PNG. The final frames went through a neutral DLSS5 neural-rendering pass and a light RCAS sharpen (strength 0.1) inside the same ComfyUI graph.

    These prompts do describe the intended natural shape in words. The showcase is therefore a LoRA-plus-prompt result: it shows what you get when you use the LoRA the way it is meant to be used, not a prompt-free transfer test.

    v1.0 training

    • Base: MiniMax H3-Image (trained on the INT8 ConvRot base; the showcase was generated on the FP8 scaled base).

    • 119 curated, captioned images at 1280 resolution, with per-image masks focusing training on the torso.

    • LoRA rank / alpha 16 / 16, learning rate 2.5e-5, AdamW 8-bit, frozen text encoder.

    • The run went to 1200 steps; this release is the step-900 checkpoint. It gave the best balance across front, profile, seated, fitted-knit and thin-tank test scenes. Later checkpoints drifted back toward higher, rounder, more projected shapes, and step 950 broke arm and torso pose in a leaning profile.

    v1.0 limitations

    • Lying on the back is the weakest pose. Breasts can stay too high and centred instead of spreading sideways with gravity.

    • It can slightly shift faces toward a different look than the base model would give for the same seed.

    • Only weight 1.0 has been tested. Lower and higher weights, men, character LoRAs, other body or size LoRAs and reference-image editing were not validated.

    v1.0 recommended settings

    • Base model: MiniMax H3-Image

    • LoRA: Novarchi_H3Image_NaturalBreasts_v1_step900.safetensors

    • Weight: 1.0

    • 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_NaturalBreasts_v1_License.zip, attached to v1.0.

    Each version carries its own copy of the license and NOTICE: v1.0 in Novarchi_H3Image_NaturalBreasts_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

    Selected step-900 checkpoint of the MiniMax H3-Image Natural Breasts LoRA.

    Gives breasts a natural shape: soft weight, gentle sag, fuller lower curves, smooth chest attachment and slight asymmetry. It works on bare skin and under thin fitted clothing. Use weight 1.0; no trigger words are required. Describe size in the prompt; the LoRA shapes it.

    Step 900 was chosen from a 1200-step run because it gave the best balance across front, profile, seated and fitted-clothing test scenes. Later checkpoints drifted toward rounder, more projected geometry. Known weak spot: lying-on-the-back poses can stay too high and centred.

    FAQ

    LORA
    MiniMax H3 Image

    Details

    Downloads
    1,344
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/15/2026
    Updated
    11/20/2026
    Deleted
    -

    Files

    Novarchi_H3Image_NaturalBreasts_v1_License.zip

    Novarchi_H3Image_NaturalBreasts_v1_step900.safetensors