This is a free LoRA, please don't use it for commercial stuff or other profit-making schemes.<h1 id="invisiblelivingclothes-minimax-h3-lora-full-bf16-+-pruned-int8">Invisible/Livingclothes - MiniMax H3 LoRA - Full BF16 + Pruned INT8</h1><h2 id="model-purpose">Model Purpose</h2>Invisible/Livingclothes is a Rank-32 BF16 concept LoRA trained for MiniMax H3 FL2VA. It is designed to improve the depiction of clothing supported and animated by a completely invisible wearer: garments retain believable human volume, posture, folds, inertia, and continuous motion while the wearer remains invisible.Please note that this is a pretrained preview release. Some parts of the dataset have not yet been fully captioned, so this version is intended for prerelease testing only. A formal version will be published after the annotation progress and status have been documented in detail.The current release candidates contain two independently trained branches:<ul><li>MiniMax H3 official Full non-pruned BF16</li><li>MiniMax H3 Pruned INT8 ConvRot</li></ul>Both branches use the same data split and training objective, but they use different base models. The LoRAs cannot be loaded interchangeably. Do not load the Full LoRA into the Pruned base model, and do not load the Pruned LoRA into the Full base model. For controlled comparisons, load only one version at a time.This is an FL2VA concept LoRA supporting T2V and I2V. It is not a paired Ref2VA R2V person-removal model.<h2 id="which-version-should-i-download">Which Version Should I Download?</h2><h3 id="bf16-or-int8-non-pruned">BF16 or INT8 non-pruned</h3>Diffusion model: <code>minimax_h3_fl2va_bf16/int8.safetensors</code> Initial release file: <code>Livingclothes-H3-Full-BF16-Epoch09-Best-HeldOut.safetensors</code><h3 id="pruned-int8">Pruned INT8</h3>Diffusion model: <code>minimax_h3_fl2va_pruned_int8_convrot.safetensors</code> Initial release file: <code>Livingclothes-H3-Pruned-INT8-Epoch08-Best-HeldOut.safetensors</code><h2 id="minimax-h3-strength-settings">MiniMax H3 Strength Settings</h2><ul><li>LoRA strength: 0.6-0.9. We recommend starting with 0.7, then comparing 0.6 / 0.8.</li></ul><h2 id="minimax-h3-prompt-example">MiniMax H3 Prompt Example</h2>Default trigger words: living cloth, living clothes, magic clothes, magic cloth, livingcloth, livingclothes.A fixed trigger word is not mandatory. The training set contains 695 explicit empty prompts, but during inference it is still recommended to clearly describe the garment, the invisible wearer, the motion rhythm, temporal changes, and camera constraints.<h3 id="t2v-example">T2V Example</h3>integrated_multimodal_description: High-quality video, 4K resolution.invisible, living cloth, living clothes, magic clothes, magic cloth, livingcloth, livingclothes. Outdoors beneath a vivid blue sky with leafy trees and distant buildings, a light blue living jacket fills out around an invisible body that poses for a selfie. The garment clings to the phantom torso and limbs with natural tension across the zipped front and raised sleeve, while the hood drapes over the unseen head and frames a beige sun mask and perched wire-rimmed glasses that suggest a direct gaze toward the viewer. Fabric folds pull and stretch in response to the invisible figure's subtle movements, the right sleeve drooping open as if an invisible hand has just slipped free, creating an intimate, playful tension between the animated garment and the hidden presence inside. The jacket's nylon shell rustles with phantom movement, the hood arching slightly as the invisible owner holds a confident, theatrical stance, blending everyday urban surroundings with this uncanny display of the living clothes guiding the posture against the breeze.overall_soundscape: fabric moving, background soundnon_diegetic_music: N/AThe current workflow does not use a separate negative prompt. Write exclusions directly in the main prompt:<pre><code>An outfit supported by a completely invisible wearer, behaving as living clothes,
preserving realistic garment volume, folds and continuous motion in the scene.
No abrupt cuts. No flicker. No extra subjects.
No text, logo, or watermark.
</code></pre><h2 id="known-limitations">Known Limitations</h2><ul><li>This LoRA is a test version and training is not yet complete. It generally performs well in typical SFW scenes, but its performance in NSFW scenes is limited.</li><li>Under certain conditions, especially with clothing or fabrics that have complex textures, image quality may degrade noticeably when LoRA strength is greater than 0.7.</li><li>It is incompatible with some existing well-known NSFW LoRAs. Combining them may reverse the intended Invisible/Livingclothes effect.</li></ul><h2 id="training-data">Training Data</h2>Both branches use the same frozen split:<ul><li>Valid unique media: 5,578</li><li>Training media: 5,528</li><li>Held-out: 50</li><li>Training images: 5,208</li><li>Silent training videos: 320</li><li>Intentional empty prompts: 695</li></ul><h2 id="shared-training-configuration">Shared Training Configuration</h2><ul><li>LoRA rank: 32</li><li>LoRA save precision: BF16</li><li>Optimizer: AdamW Optimi</li><li>Learning rate: <code>1e-4</code></li><li>Steps per epoch: 967</li><li>Image / video micro-batch: 8 / 1</li><li>Activation checkpointing: Unsloth</li><li>MiniMax H3 shift: image 1, video 12</li><li>The LoRA is saved after every complete epoch, followed by separate held-out evaluations for images and silent videos.</li></ul>

[Invisible person / Living clothes] MiniMax H3 T2V/I2V/FL2V

LoRA: Preview Pruned INT8 version v0.5 Name: Livingclothes-H3-Pruned-INT8-Epoch08-Best-HeldOut(QQQQ4413)

Files with SHA256: fd9bfb746378ae0d1ec0bcc8bc2d86593195a9ac6e5f57b9d602b717ccdab15e