This is a free LoRA, please don't use it for commercial stuff or other profit-making schemes.
All-in-One Invisible Person & LivingClothes (Ref2VA V2V) - MiniMax H3 LoRA [Alpha Preview]
Model Purpose & Alpha Preview Notice
This is an early Alpha concept-validation preview release (Alpha-v2 / Step 2500) of an All-in-One person-removal and hollow-garment editing LoRA trained for MiniMax H3 Ref2VA (Reference-to-Video / V2V). It is not the final version.
Concept & Pipeline Origin: Building directly on the paired-edit concept explored and tested by gangstuhh (Special thanks!). I generated synthetic ground-truth pairs in Virt-A-Mate (VAM) alongside paired image-editing outputs (2D & 3D photorealistic control-to-target image pairs) and trained this LoRA using an IMAGE-ONLY paired dataset (zero real video clips in training, bridged into MiniMax H3's 3D temporal blocks via 5-frame pseudo-video <Video 1> conditioning).
Despite being trained exclusively on static image pairs, the learned removal and garment-hollowing operators transfer zero-shot to full 5+ second (124+ frame) V2V (Ref2VA) video editing across four core modes in a single unified checkpoint:
Remove Man (
REMOVE-MAN): Erases the male subject (body, head, arms, hands, fingers) from the input video while keeping the female subject, physical contact deformations/indentations, and background intact.Remove Woman (
REMOVE-WOMAN): Erases the female subject completely from the input video while keeping the male subject, his uncovered anatomy, and the background intact.Remove Man + Keep Living Clothes (
REMOVE-MAN-KEEP-CLOTHES+livingclothes): Turns the man's body, skin, head, and limbs invisible while preserving his outfit floating in 3D volume as animated living clothes.Remove Woman + Keep Living Clothes (
REMOVE-WOMAN-KEEP-CLOTHES+livingclothes): Turns the woman's body, skin, head, hair, and limbs invisible while keeping her garments floating and deforming naturally around her unseen body curves (also works in two-person scenes while keeping the man visible).
Current Status & Future Roadmap:
Currently this Alpha build ONLY supports V2V via
Ref2VA(MiniMaxH3ReferenceToVideowith<Video 1>). It does not support standalone T2V or FL2VA I2V base models.Because zero real video pairs were used in this initial image-only validation run, temporal motion consistency and camera framing lock still show some deviation on difficult clips (for example, occasional rhythmic motion lag or slight Frame-0 zoom shifts), though overall editing quality is already solid.
Next Steps: I am continuing research to add true paired video training sets for tighter motion/camera locking, and ultimately plan to merge all of my previously released T2V / I2V (
FL2VA) Invisible, Grope, and LivingClothes datasets together withRef2VAV2V into one unified All-in-One MiniMax H3 LoRA.
Base Model & File Information
This Alpha preview provides a single LoRA file trained for the Ref2VA Pruned base model family (compatible with both BF16 and INT8 ConvRot pruned weights):
Release File:
h3_invisible_allinone_ref2va_alpha_perview_r128_step02500(QQQQ4413).safetensors(Rank 128 BF16,1.18 GB, AutoV2 Hash:cc6ca34d)Required Diffusion Model (UNET):
minimax_h3_ref2va_pruned_bf16.safetensorsorminimax_h3_ref2va_pruned_int8_convrot.safetensors(Ref2VA Pruned models only; do not load into non-pruned Full, FL2VA, or T2VA models)Text Encoder (CLIP):
qwen3vl_32b_minimax_h3_int8_convrot.safetensors(type: minimax_h3)Video / Audio VAE:
minimax_h3_video_vae_fp16.safetensors/minimax_h3_audio_vae_fp32.safetensors
Important: Do not load this LoRA into minimax_h3_fl2va_* or T2VA models. It is trained exclusively for the minimax_h3_ref2va architecture.
How to Load & Recommended ComfyUI Settings (Ref2VA V2V)
Conditioning Node: Use
MiniMaxH3ReferenceToVideo. Connect your source video toref_videos(slot 0, referenced in the prompt as<Video 1>) and connectvideo_vae. Setref_image_sizetomatch, and match outputwidth,height(multiples of 32), andlength(native17n + 5grid, e.g.124frames at24 fps) to your source clip.LoRA Strength:
0.85 - 1.0. Start at1.0for clean full-body and limb removal; try0.85 - 0.90if a wide shot shifts slightly.Fast 8-Step Turbo Workflow (Recommended):
Chain
UNETLoader (minimax_h3_ref2va_pruned_bf16)→LoraLoaderModelOnly (This Invisible LoRA @ 1.0)→ [Optional Companion NSFW LoRA below] →LoraLoaderModelOnly (minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors @ 1.0)→MiniMaxH3SigmaShift (shift = 12.0).Sampler:
euler,steps = 8,CFG = 1.0. (Using 8 steps instead of 4-6 steps noticeably reduces camera drift and residual finger/skin ghosts.)
Native Non-Turbo Workflow: Skip the 4-step Turbo LoRA and run
24 - 28 stepswith standard flowmatch scheduling atCFG = 1.0 - 2.0.
Recommended Companion NSFW LoRA (Anatomy Repair Boost)
When removing an occluding partner in explicit NSFW scenes (for example, erasing the woman during oral/intercourse in REMOVE-WOMAN, or close-up REMOVE-MAN scenes), the model must reconstruct previously occluded genitalia cleanly. Stacking MysticXXX provides a strong anatomy fidelity boost without breaking invisibility:
Recommended Companion LoRA:
MysticXXX_MMH3-V5-ref2va.safetensorsRecommended Strength:
0.45 - 0.60(load it right after this Invisible LoRA at1.0, and before the Turbo 4-step LoRA at1.0).When to Enable: Strongly recommended whenever uncovered glans/shaft structure looks soft or truncated, or when pubic hair / vulva / foreskin detail needs cleaner anatomical reconstruction after partner removal.
Trigger Keywords & Prompt Examples (4 Core Modes)
Always reference the input clip as <Video 1> and select one primary mode prefix:
1. Remove Man:
Primary Trigger Prefix:
REMOVE-MAN, REMOVE-MALE, all men are invisible.Optional Interaction Sub-Tags:
invisiblesex,invisiblegrab,ghostgrope,gaping_vagina,hollow_insertion
2. Remove Woman:
Primary Trigger Prefix:
REMOVE-WOMAN, REMOVE-FEMALE, all women are invisible.Optional Interaction Sub-Tags:
invisiblesex(avoidinvisiblegrabon Remove-Woman!)
3. Remove Man + Keep Clothes:
Primary Trigger Prefix:
REMOVE-MAN-KEEP-CLOTHES, livingclothes, invisible_clothes, hollow_garment, invisible man wearing hollow clothes.Optional Interaction Sub-Tags:
invisiblesex
4. Remove Woman + Keep Clothes:
Primary Trigger Prefix:
REMOVE-WOMAN-KEEP-CLOTHES, livingclothes, invisible_clothes, hollow_garment, invisible woman wearing hollow clothes.Optional Interaction Sub-Tags:
invisiblesex,cowgirl,doggystyle,missionary
1. Remove Man Example (Preserving Contact Indentations & Woman's Own Hands)
REMOVE-MAN, REMOVE-MALE, invisiblesex, invisiblegrab, all men are invisible. Remove ONLY the man's entire body, skin, head, hair, arms, hands, and fingers completely from <Video 1> while keeping the woman and both of her own arms and hands 100% intact, preserving her natural body motion and hip contact indentations, and reconstructing the background seamlessly with the exact camera framing, scale, and perspective of <Video 1> locked across all frames.2. Remove Woman Example (Zero Phantom-Hand Rule Applied)
REMOVE-WOMAN, REMOVE-FEMALE, invisiblesex, all women are invisible. Remove the woman's entire body, skin, head, hair, and limbs completely from <Video 1> while preserving the man and his bare anatomy in their exact position and scale, reconstructing the bedding and background seamlessly with the exact camera framing and perspective of <Video 1> locked.3. Remove Man + Keep Living Clothes Example
REMOVE-MAN-KEEP-CLOTHES, livingclothes, invisible_clothes, hollow_garment, invisible man wearing hollow clothes. Remove the man's entire body, skin, head, hair, and limbs from <Video 1> while keeping his shirt, pants, and outfit floating as 3D hollow livingclothes clinging naturally to his invisible body posture, keeping the exact camera framing and perspective of <Video 1> locked.4. Remove Woman + Keep Living Clothes Example
REMOVE-WOMAN-KEEP-CLOTHES, livingclothes, invisible_clothes, hollow_garment, invisible woman wearing hollow clothes. Remove the woman's entire body, skin, head, hair, and limbs from <Video 1> while keeping her outfit floating as 3D hollow livingclothes animated by her unseen body curves (and keeping the man completely visible and intact if present), locking the exact camera framing and perspective of <Video 1>.Critical Prompting Tips (Important at CFG = 1.0)
Zero "Hand / Finger" Words in
REMOVE-WOMANPrompts: At distilledCFG = 1.0there is no negative prompt channel. Mentioning the literal tokenshand,hands,finger,fingers, orinvisiblegrabanywhere inside aREMOVE-WOMANprompt (even inside "remove her hands and fingers"!) activates hand cross-attention and can hallucinate disembodied hands on the man's body. Always writelimbsinstead ("Remove the woman's entire body, skin, head, hair, and limbs").Match Prompt Scale to Shot Framing: On wide or medium room shots, do not add close-up macro keywords such as
invisibleoral,close-up scale, or macro skin prose — doing so can pull Frame 0 into a zoomed-in close-up crop. Instead, explicitly name 2-3 static background landmarks (lamp, headboard, wall art, window) and keep the camera framing lock clause.No Camera Movement Adjectives: Avoid writing
pan,zoom, ortiltin the prompt; let<Video 1>drive the camera.
Known Limitations (Alpha Preview)
Ref2VA V2V Only: Does not work on T2V or FL2VA I2V workflows.
Motion & Alignment Variance: Trained 100% on static image pairs without real video supervision yet. Fast motion, heavy camera shake, or shots where the erased subject covers >70% of the frame can show mild rhythmic pauses, background hallucination, or slight scale shifts from the original video.
Heavy Occlusion Reconstruction: When an erased body completely blocks the kept person's head or genitals in the source clip, single-LoRA reconstruction may occasionally truncate or soften that body part — stack
MysticXXX_MMH3-V5-ref2va @ 0.45 - 0.60to mitigate this.
Training Data
Trained strictly on curated static control → target image pairs (0 video files in the dataset), organized into three core categories:
REMOVE-MAN:798unique image pairsREMOVE-WOMAN:158unique image pairsKEEP-CLOTHES / LivingClothes:224unique image pairsTotal (Image-Only, 0 Videos):
1,180unique paired images (class-balanced to2,064samples per epoch)Native Aspect-Ratio Buckets (
max_res = 1280, multiples of 32, zero aspect stretch):832×1216(523),1280×800(297),1024×768(151),1024×1024(129),1344×768(43),896×1152(23),704×1216(7), and remaining native buckets (7).
Training Configuration & Convergence Metrics
Trainer:
ostris/ai-toolkitBase Architecture & Official Components:
minimax_h3_ref2va(minimax_h3_ref2va_pruned_bf16.safetensors+qwen3vl_32b_minimax_h3_int8_convrot.safetensors+minimax_h3_ref2va_training_adapter_v2.safetensors+minimax_h3_video_vae_fp16.safetensors)Ref2VA Conditioning Settings:
image_refs_as_video: true,image_ref_video_frames: 5LoRA Network: Rank
128/ Alpha128, saved inBF16(dropout = 0.0)Optimizer & Schedule:
AdamW 8-bit, learning rate1e-4,constant_with_warmup(150warmup steps),max_grad_norm = 1.0Batching & Noise Scheduler: Batch size
1× Gradient Accumulation4(effective batch size4→516steps per epoch),flowmatchscheduler withshifttimestep samplingRelease Checkpoint: Step 2,500 (
Epoch 4.84)Loss Trajectory: Flow-Matching
EMA Lossconverged smoothly from>0.14(warmup) →0.089(Step 2,000) →
Description
h3_invisible_allinone_ref2va_alpha_preview_r128_step02500(QQQQ4413).safetensors