Test3lora / app.py
observantdistressed's picture
Remove concurrency_limit β€” no request queue, memory cleanup retained
995d272 verified
Raw History Blame Contribute Delete
118 kB
"""MiniMax-H3 `t2va` / `fl2va`, split deployment β€” the denoising half.""" from __future__ import annotations import glob import json import os import re import shutil import struct import tempfile import threading import time import traceback from functools import cache # Before anything that could initialize CUDA: `import spaces` patches `torch.cuda` so the 72 GiB load can happen at # startup rather than on GPU time. import spaces import gradio as gr MODEL_REPO = os.environ.get("H3_MODEL_REPO", "MiniMaxAI/MiniMax-H3") CONDITIONER_SPACE = os.environ.get("H3_CONDITIONER", "multimodalart/qwen3vl-conditioner") # `pack` places the transformer at startup, `lazy` moves everything on the first GPU call, `offload` hands placement to # `ComponentsManager.enable_auto_cpu_offload`. PLACEMENT = os.environ.get("H3_PLACEMENT", "pack").lower() # cuDNN's fused attention is 10-20% faster than the SDPA default on this pool and needs nothing installed. ATTENTION = os.environ.get("H3_ATTENTION", "_native_cudnn").lower() GPU_SIZE = os.environ.get("H3_GPU_SIZE", "xlarge") # larryvrh distillation LoRA (diffusers conversion): A/B factors stored on CPU, merged on-demand inside @spaces.GPU. TURBO_LORA_REPO = os.environ.get("H3_TURBO_LORA", "InstantX/MiniMax-H3-Turbo-Lora-Diffusers") TURBO_LORA_FILE = "minimax_h3_turbo_4step_ckpt500_diffusers.safetensors" TURBO_LORA_SCALE = 1.0 # alpha == rank # NSFW / anatomy LoRAs (ComfyUI format, converted at load). Lazy: downloaded + converted on CPU in generate(), # merged on GPU inside @spaces.GPU only when selected. Standard LoRA only β€” LoKR (.lokr_w1/.lokr_w2) is unsupported. NSFW_LORA_CONFIGS = { "nsfw_innie": { "repo": "Hearmeman/minimax-h3-loras", "file": "HMInnie_v1_e50.safetensors", "scale": 1.0, "label": "NSFW Innie Pussy (HMInnie V1)", "trigger": "inniepussy", "desc": "Renders the vulva as a neat 'innie' β€” labia tucked in with no protruding inner lips. Use for close-up genital shots where you want a clean, compact vulva look. Not for scenes needing prominent outer labia or detailed pussy anatomy β€” pair with HMPussy for that.", }, "nsfw_penis": { "repo": "observantdistressed/plora-h3", # CivitAI "Penis Lora - H3" v1.1, mirrored to HF "file": "PLORA_H3_V2-step00006300.safetensors", "scale": 0.7, "label": "NSFW Penis (Penis Lora - H3 v1.1)", "trigger": "PENISLORA", "format": "kohya", "desc": "Adds a detailed, erect penis to male subjects β€” the core anatomy LoRA for any scene with a visible penis. Use whenever a penis is in frame. Not for solo female scenes, and don't stack it with HMPenis v2.0 (same purpose, conflicting weights).", }, "nsfw_dildo": { "repo": "EllaPriest45/MinimaxH3_Actions", # CivitAI "Dildoing" model 2903074, backup repo "file": "Dildoing - MinimaxH3 - pumping a dildo insider her vagina.safetensors", "scale": 0.7, "label": "NSFW Dildoing (Dildo Insertion)", "trigger": "pumping a dildo insider her vagina", "desc": "Makes dildo insertion read naturally β€” the toy slides into the vagina with proper depth and motion instead of clipping or floating. Use for any dildo penetration scene. Not for scenes without a dildo, and don't combine it with other insertion LoRAs at full strength.", }, # Motion/action adapter β€” auto-loaded (not user-selectable) whenever any anatomy LoRA is on, so the # anatomy renders sharply AND the act moves naturally. The Hearmeman README says to stack anatomy # adapters underneath this action adapter. "nsfw_aio": { "repo": "Hearmeman/minimax-h3-loras", "file": "HMNSFW-AIO-V2.5.safetensors", "scale": 0.7, "label": "HMNSFW AIO V2.5 (motion)", "trigger": "hmmotion", "desc": "Motion/action adapter that makes anatomy LoRAs move naturally during sex acts β€” auto-loaded whenever any NSFW LoRA is selected. Use stacked underneath anatomy LoRAs for sharp rendering plus natural motion. Not for standalone use; it's a motion enhancer, not an anatomy fix.", }, # ── EllaPriest45/MinimaxH3_Actions catalog (t2va-compatible, standard LoRA only) ── # 93 LoRAs, classified by header inspection: 80 ComfyUI + 13 kohya. LoKR, I2V, REF2V, R2V, FL2VA excluded. # Triggers/scales extracted from the CivitAI filenames (best effort β€” refine per-LoRA as needed). 'nsfw_ahegao_face': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Ahegao Face - MinimaxH3 - She is making the ahegao face - eyes rolled back and tongue hanging out.safetensors', "scale": 0.7, "label": 'Ahegao Face', "trigger": 'ahegao face', "desc": "Makes the subject's face contort into the ahegao expression β€” eyes rolled back, tongue hanging out β€” for intense reaction shots. Use during orgasm or overstimulation scenes. Not for normal expressions or dialogue scenes; it overrides the face.", }, 'nsfw_all_tied_up': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'All Tied Up - MinimaxH3 - 0.8str.safetensors', "scale": 0.8, "label": 'All Tied Up', "trigger": 'all tied up', "desc": "Renders the subject bound and restrained, with rope or ties visible and functional. Use for bondage and restraint scenes. Not for scenes where the subject needs to move freely; it locks them into restraint poses.", }, 'nsfw_astro_nsfw': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Astro NSFW - MinimaxH3 - improves human movement and anatomy.safetensors', "scale": 0.7, "label": 'Astro NSFW', "trigger": 'astro nsfw', "format": "kohya", "desc": "General NSFW quality enhancer that improves human movement and anatomy in explicit scenes. Use as a base layer on top of specific anatomy or act LoRAs. Not a specific act LoRA β€” don't expect it to produce a particular act on its own.", }, 'nsfw_auto_cunnilingus': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Auto-Cunnilingus - MinimaxH3 - She performs oral sex on herself.She licks her vulva.safetensors', "scale": 0.7, "label": 'Auto-Cunnilingus', "trigger": 'She performs oral sex on herself.She licks her vulva', "desc": "Makes the subject perform oral sex on herself β€” licking her own vulva. Use for self-cunnilingus scenes. Not for partnered cunnilingus; that's a different act with its own LoRAs.", }, 'nsfw_beanflk_masturbation_finger_insertion_au': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'BEANFLK - Masturbation & Finger Insertion + Audio - MinimaxH3 - BEANFLK,0.4str.safetensors', "scale": 0.4, "label": 'BEANFLK - Masturbation & Finger Insertion + Audio', "trigger": 'BEANFLK', "desc": "Masturbation with finger insertion, plus audio cues for the act. Use for solo fingering scenes. Not for scenes without finger penetration; the audio may not fit non-masturbation content.", }, 'nsfw_bunny_general_motion_continuity_repair': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'BUNNY General Motion Continuity Repair - MinimaxH3 - for dynamic scenes,bunny_crisp_motion,resmultistep+simple,euler+beta,0.9str,stage1,0.6str,stage2,0.25str.safetensors', "scale": 0.9, "label": 'BUNNY General Motion Continuity Repair', "trigger": 'bunny_crisp_motion', "desc": "Repairs motion continuity in dynamic scenes β€” reduces warping and keeps movement coherent across frames. Use for fast-action or high-motion shots. Not for slow, static scenes where it adds little.", }, 'nsfw_better_motion': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Better Motion - MinimaxH3 - 720x1280,15-30steps,0.2-0.4str.safetensors', "scale": 0.3, "label": 'Better Motion', "trigger": 'better motion', "desc": "General motion-quality LoRA that makes movement smoother and more natural. Use as a base enhancer for any scene. Not for specific acts; it's a motion polish, not an action generator.", }, 'nsfw_better_titfuck_v0_95': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Better Titfuck v0.95 - MinimaxH3 - res_multistep+simple,titjob,titfuck.safetensors', "scale": 0.7, "label": 'Better Titfuck v0.95', "trigger": 'titjob,titfuck', "desc": "Makes titfuck scenes read properly β€” breasts wrap around the penis with realistic motion. Use for titjob/titfuck scenes. Not for other sex acts; it specializes in breast-on-penis action.", }, 'nsfw_blowjob_v3_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Blowjob v3.0 - MinimaxH3 - bl0w_j0b.safetensors', "scale": 0.7, "label": 'Blowjob v3.0', "trigger": 'bl0w_j0b', "desc": "Core blowjob LoRA β€” makes oral sex on a penis read naturally with proper head motion. Use for blowjob scenes. Not for scenes without a visible penis; it won't do anything useful.", }, 'nsfw_blowjob_wearing_thong': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Blowjob wearing Thong - MinimaxH3 - blowjob_wearing_thong.safetensors', "scale": 0.7, "label": 'Blowjob wearing Thong', "trigger": 'blowjob_wearing_thong', "desc": "Blowjob variant where the performer keeps a thong on. Use for thong-clad blowjob scenes. Not for nude blowjobs β€” the thong is part of the concept.", }, 'nsfw_bluebella_nova_lingerie': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Bluebella Nova Lingerie - MinimaxH3 - She is wearing COLOR Bluebella Nova Lingerie.safetensors', "scale": 0.7, "label": 'Bluebella Nova Lingerie', "trigger": 'She is wearing COLOR Bluebella Nova Lingerie', "format": "kohya", "desc": "Renders the subject wearing the Bluebella Nova lingerie set β€” a specific lace-and-straps design. Use for scenes featuring this exact lingerie. Not for other outfits; it's tied to this one garment.", }, 'nsfw_boob_physics': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Boob Physics - MinimaxH3 - jiggle,grab,massage,bounce.safetensors', "scale": 0.7, "label": 'Boob Physics', "trigger": 'jiggle,grab,massage,bounce', "desc": "Adds natural breast motion β€” jiggle, bounce, and sway during movement. Use for any scene with breast movement. Not for static shots; it only shows when there's motion.", }, 'nsfw_bustywomen': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'BustyWomen - MinimaxH3 - 0.75-1str.safetensors', "scale": 0.88, "label": 'BustyWomen', "trigger": 'busty women', "desc": "Makes female subjects busty β€” fuller, larger breasts. Use when you want a well-endowed figure. Not for flat-chested or petite figures; it conflicts with tiny-breast LoRAs.", }, 'nsfw_button_pop': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Button pop - MinimaxH3 - 0.2-0.5str,butt0n_p0p.safetensors', "scale": 0.35, "label": 'Button pop', "trigger": 'butt0n_p0p', "desc": "Makes buttons pop off clothing under strain β€” a clothing-failure effect. Use for scenes where a tight top bursts open. Not for general scenes; it's a specific gag/effect.", }, 'nsfw_cameltoe_v2_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Cameltoe v2.0 - MinimaxH3 - cameltoe.safetensors', "scale": 0.7, "label": 'Cameltoe v2.0', "trigger": 'cameltoe', "desc": "Renders a visible cameltoe through tight clothing. Use for scenes with tight bottoms. Not for nude scenes; it needs clothing to show through.", }, 'nsfw_can_t_stop_shaking': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": "Can't Stop Shaking! - MinimaxH3 - endface1,endbody1.safetensors", "scale": 0.7, "label": "Can't Stop Shaking!", "trigger": 'endface1,endbody1', "desc": "Makes the subject shake uncontrollably β€” body and face tremors, often during intense stimulation. Use for overstimulation and orgasm scenes. Not for calm scenes; it adds constant shaking.", }, 'nsfw_chasing_blowjob_v1_2': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Chasing blowjob v1.2 - MinimaxH3 - The girl performs a blowjob.safetensors', "scale": 0.7, "label": 'Chasing blowjob v1.2', "trigger": 'The girl performs a blowjob', "desc": "Blowjob where the performer chases the penis β€” active, eager head movement. Use for enthusiastic blowjob scenes. Not for slow or gentle oral; it's an energetic variant.", }, 'nsfw_combat_base_v2_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Combat BASE v2.0 - MinimaxH3 - prfight2,prfin1,res_multistep+simple,euler+beta.safetensors', "scale": 0.7, "label": 'Combat BASE v2.0', "trigger": 'prfight2,prfin1', "desc": "Base LoRA for fight and combat scenes β€” wrestling, grappling, and physical struggle. Use for combat or struggle scenes. Not for sex acts; it's a fight-motion base.", }, 'nsfw_cum_swallow': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Cum Swallow - MinimaxH3 - TADAA,0.4str.safetensors', "scale": 0.4, "label": 'Cum Swallow', "trigger": 'TADAA', "desc": "Makes the subject swallow cum after a facial or oral scene β€” the swallowing action reads clearly. Use for cum-swallow endings. Not for scenes without a cumshot; it needs the setup.", }, 'nsfw_cumshot': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Cumshot - MinimaxH3 - 20steps,euler+beta,0.85str.safetensors', "scale": 0.85, "label": 'Cumshot', "trigger": 'cumshot', "desc": "Renders a visible cumshot β€” ejaculation with realistic volume and motion. Use for cumshot scenes. Not for non-ejaculation scenes; it forces the act.", }, 'nsfw_deeper_hold_it': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Deeper... Hold It. - MinimaxH3 - oralmove1,oraldetail1,0.8str.safetensors', "scale": 0.8, "label": 'Deeper... Hold It.', "trigger": 'oralmove1,oraldetail1', "desc": "Oral-sex detail LoRA β€” deep-throat motion and oral detail. Use for deep oral scenes. Not for shallow kissing or foreplay; it's oral-specific.", }, 'nsfw_dildo_masturbation_v2_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Dildo Masturbation v2.0 - MinimaxH3 - She slide a dildo in and out of her pussy.safetensors', "scale": 0.7, "label": 'Dildo Masturbation v2.0', "trigger": 'She slide a dildo in and out of her pussy', "desc": "Makes solo dildo masturbation read naturally β€” the subject slides a dildo in and out of her pussy. Use for solo dildo scenes. Not for partnered penetration; it's solo-focused.", }, 'nsfw_disney_style_renaissance_era': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Disney Style - Renaissance Era - MinimaxH3 - In the style of Disney animation.safetensors', "scale": 0.7, "label": 'Disney Style - Renaissance Era', "trigger": 'In the style of Disney animation', "format": "kohya", "desc": "Applies a Disney Renaissance-era animation style β€” the classic 90s hand-drawn look. Use for stylized or animated scenes. Not for photorealistic content; it fights realism.", }, 'nsfw_drone_shot': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Drone Shot - MinimaxH3 - dr0nesh0t.safetensors', "scale": 0.7, "label": 'Drone Shot', "trigger": 'dr0nesh0t', "desc": "Adds a drone-style aerial camera shot β€” high-angle and sweeping. Use for establishing or overhead shots. Not for close-up intimate scenes; it pulls the camera away.", }, 'nsfw_facial_cumshot_v0_1': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Facial CumShot v0.1 - MinimaxH3 - cmst.safetensors', "scale": 0.7, "label": 'Facial CumShot v0.1', "trigger": 'cmst', "desc": "Renders a facial cumshot β€” cum landing on the face. Use for facial scenes. Not for internal or swallow endings; it's face-specific.", }, 'nsfw_faster_harder_shake_harder_v0_2': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Faster! Harder! Shake Harder! v0.2 - MinimaxH3 - dynv2.safetensors', "scale": 0.7, "label": 'Faster! Harder! Shake Harder! v0.2', "trigger": 'dynv2', "desc": "Makes the subject move faster and harder β€” intense, vigorous motion. Use for rough or fast scenes. Not for gentle or slow scenes; it escalates intensity.", }, 'nsfw_fetish_submission_chain_teddy': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Fetish Submission Chain Teddy - MinimaxH3 - She is wearing a red latex fetish submission chain teddy.Her hands are bound in front of her.safetensors', "scale": 0.7, "label": 'Fetish Submission Chain Teddy', "trigger": 'She is wearing a red latex fetish submission chain teddy', "desc": "Renders the subject in a red latex fetish submission chain teddy, with her hands bound in front. Use for latex or bondage outfit scenes. Not for casual outfits; it's a specific fetish garment.", }, 'nsfw_fingering': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Fingering - MinimaxH3 - 0.8-1str.safetensors', "scale": 0.9, "label": 'Fingering', "trigger": 'fingering', "desc": "Makes fingering scenes read naturally β€” fingers inside the vagina with realistic motion. Use for manual stimulation scenes. Not for scenes without hand-on-genital contact.", }, 'nsfw_framerush': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'FrameRush - MinimaxH3 - lights.safetensors', "scale": 0.7, "label": 'FrameRush', "trigger": 'lights', "desc": "Camera and lighting effect LoRA β€” adds a rushed, dynamic frame feel. Use for energetic camera work. Not for stable, slow shots.", }, 'nsfw_furry_enhancer_video_v1_5': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Furry Enhancer Video v1.5 - MinimaxH3 - 0.6-0.7str.safetensors', "scale": 0.65, "label": 'Furry Enhancer Video v1.5', "trigger": 'furry enhancer', "desc": "Enhances furry and anthro characters in video β€” fur texture and anthro anatomy. Use for furry content. Not for human subjects; it will fur-ify them.", }, 'nsfw_futanari_transformation_futanarization_v': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Futanari transformation (Futanarization) v5.1 - MinimaxH3 - futanaei transformation,futanaei transformation.safetensors', "scale": 0.7, "label": 'Futanari transformation (Futanarization) v5.1', "trigger": 'futanaei transformation', "desc": "Renders futanari transformation β€” a female subject gaining a penis. Use for futanari or transformation scenes. Not for standard male-female scenes; it changes anatomy.", }, 'nsfw_grindtime': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Grindtime - MinimaxH3 - grindtime,a woman performing continuous rhythmic gyrating hip motion while having sex,front-rear-side view.safetensors', "scale": 0.7, "label": 'Grindtime', "trigger": 'grindtime', "desc": "Makes the subject perform continuous rhythmic hip grinding during sex β€” front, rear, and side views. Use for grinding or riding scenes. Not for static poses; it's a motion LoRA.", }, 'nsfw_h3_footjob_v0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": "H3 Footjob v0 - MinimaxH3 - a photoealistic video depicting a woman sitting in front of the viewer.the camera angle is a male pov.the viewer's penis is exposed.the woman is performing a footjob.h3fj.safetensors", "scale": 0.7, "label": 'H3 Footjob v0', "trigger": 'h3fj', "desc": "POV footjob LoRA β€” the subject strokes a penis with her feet from a male POV. Use for footjob scenes. Not for scenes without a visible penis and feet; it's a specific POV setup.", }, 'nsfw_hmbreasts_v2_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'HMBreasts v2.0 - MinimaxH3 - tits,seeds2+ddim_uniform,8stepsturbolora.safetensors', "scale": 0.7, "label": 'HMBreasts v2.0', "trigger": 'tits', "desc": "Hearmeman breast LoRA β€” renders attractive, natural breasts. Use for breast-focused scenes. Not for flat-chested figures; it conflicts with tiny-breast LoRAs.", }, 'nsfw_hmmasturbation_squirting_and_female_mast': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'HMMasturbation - Squirting and Female Masturbation - MinimaxH3 - hmmasturbation,euler+simple,12steps,shift6,turbo lora 0.5str,use hmpussy at 0.4-0.8str.safetensors', "scale": 0.6, "label": 'HMMasturbation - Squirting and Female Masturbation', "trigger": 'hmmasturbation', "desc": "Hearmeman masturbation LoRA β€” solo masturbation with squirting. Use for solo scenes with squirting. Not for partnered sex; it's solo-focused.", }, 'nsfw_hmpenis_v2_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'HMPenis v2.0 - MinimaxH3 - penis,0.9str.safetensors', "scale": 0.9, "label": 'HMPenis v2.0', "trigger": 'penis', "desc": "Hearmeman penis LoRA β€” detailed erect penis anatomy. Use for any scene with a visible penis. Not for solo female scenes, and don't stack it with the dedicated penis LoRA (same purpose).", }, 'nsfw_hmpussy_v1_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'HMPussy v1.0 - MinimaxH3 - pussy,0.9str.safetensors', "scale": 0.9, "label": 'HMPussy v1.0', "trigger": 'pussy', "desc": "Hearmeman pussy LoRA β€” detailed vulva anatomy. Use for close-up genital shots. Not for scenes where the vulva isn't visible; it only affects what's in frame.", }, 'nsfw_hmsquirt_squirting_v0_1': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'HMSquirt - Squirting v0.1 - MinimaxH3 - euler+simple,hmsquirt,12steps,6shift,turbolora 0.5str.safetensors', "scale": 0.5, "label": 'HMSquirt - Squirting v0.1', "trigger": 'hmsquirt', "desc": "Hearmeman squirting LoRA β€” makes squirting read clearly during stimulation. Use for squirting scenes. Not for dry scenes; it forces the effect.", }, 'nsfw_hand_in_panties_masturbation': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Hand in Panties Masturbation - MinimaxH3 - 0.7str,She has her hand in her panties,She slides her hand in her panties.She rubs her vagina.safetensors', "scale": 0.7, "label": 'Hand in Panties Masturbation', "trigger": 'She has her hand in her panties,She slides her hand in her panties.She rubs her vagina', "desc": "Makes the subject slide her hand into her panties and rub herself. Use for clothed masturbation scenes. Not for nude masturbation; the panties are part of the concept.", }, 'nsfw_hip_gyration': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Hip Gyration - MinimaxH3 - gyrating her hips,0.8str.safetensors', "scale": 0.8, "label": 'Hip Gyration', "trigger": 'gyrating her hips', "desc": "Makes the subject gyrate her hips rhythmically β€” circular hip motion. Use for grinding or dancing scenes. Not for static poses; it's a motion LoRA.", }, 'nsfw_hip_sway_dance': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Hip Sway Dance - MinimaxH3 - dancing with arms in front.safetensors', "scale": 0.7, "label": 'Hip Sway Dance', "trigger": 'dancing with arms in front', "desc": "Dance LoRA β€” the subject dances with arms in front, swaying her hips. Use for dance scenes. Not for sex acts; it's a dance-motion LoRA.", }, 'nsfw_hitachi_magic_wand_vibrator': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Hitachi Magic Wand Vibrator - MinimaxH3 - using a Hitachi Magic Wand vibrator.safetensors', "scale": 0.7, "label": 'Hitachi Magic Wand Vibrator', "trigger": 'using a Hitachi Magic Wand vibrator', "desc": "Renders the subject using a Hitachi Magic Wand vibrator β€” the toy and its use read clearly. Use for vibrator scenes. Not for scenes without the toy; it won't add anything.", }, 'nsfw_holi_festival_of_colors': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Holi Festival of Colors - MinimaxH3 - Festival_of_Colors,multicolored smoke,multicolored paint,multicolored powder.safetensors', "scale": 0.7, "label": 'Holi Festival of Colors', "trigger": 'Festival_of_Colors', "desc": "Holi festival effect β€” multicolored smoke, paint, and powder. Use for colorful festival scenes. Not for NSFW acts; it's a visual-effects LoRA.", }, 'nsfw_huge_breasts_large_breasts': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Huge Breasts - Large Breasts - MinimaxH3 - bigsloppytits,huge-massive-enormous saggy breasts-bust,nip_protruding-flat-soft,areola_large-wide-medium-small.safetensors', "scale": 0.7, "label": 'Huge Breasts - Large Breasts', "trigger": 'bigsloppytits', "desc": "Renders huge, saggy breasts β€” massive with large areolas. Use for extreme-breast scenes. Not for natural or petite figures; it goes big.", }, 'nsfw_icy_ass_smacking': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Icy Ass Smacking - MinimaxH3 - icyslap.safetensors', "scale": 0.7, "label": 'Icy Ass Smacking', "trigger": 'icyslap', "desc": "Makes ass-smacking read naturally β€” the slap and the resulting jiggle. Use for spanking scenes. Not for non-spanking scenes; it forces the action.", }, 'nsfw_icy_twerk_pro_max_v2_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Icy Twerk Pro MAX v2.0 - MinimaxH3 - icytwerk.safetensors', "scale": 0.7, "label": 'Icy Twerk Pro MAX v2.0', "trigger": 'icytwerk', "desc": "Twerk LoRA β€” vigorous twerking with strong ass motion. Use for twerk scenes. Not for slow or gentle movement; it's high-energy.", }, 'nsfw_innie_pussy': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Innie Pussy - MinimaxH3.safetensors', "scale": 0.7, "label": 'Innie Pussy', "trigger": 'inniepussy', "format": "kohya", "desc": "Innie pussy variant (kohya format) β€” a clean, tucked vulva look. Use for close-up genital shots wanting a neat innie. Not for scenes needing prominent labia; same caveat as the dedicated innie LoRA.", }, 'nsfw_jav_moans_audio_only': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'JAV moans audio only - MinimaxH3 - jpnMoans.safetensors', "scale": 0.7, "label": 'JAV moans audio only', "trigger": 'jpnMoans', "format": "kohya", "desc": "Audio LoRA β€” adds JAV-style moans to the soundtrack. Use for scenes where you want Japanese AV-style vocalizations. Not for scenes with dialogue; it's audio-only.", }, 'nsfw_jiggle_walk': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'JIGGLE WALK - MinimaxH3 - JGLWLK.safetensors', "scale": 0.7, "label": 'JIGGLE WALK', "trigger": 'JGLWLK', "desc": "Makes the subject walk with a pronounced jiggle β€” breast and body bounce while walking. Use for walking scenes. Not for static or lying-down scenes; it needs locomotion.", }, 'nsfw_kaku_table_humping': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'KAKU - Table Humping - MinimaxH3 - 0.4-0.5str,KAKU,euler+simple.safetensors', "scale": 0.45, "label": 'KAKU - Table Humping', "trigger": 'KAKU', "desc": "Makes the subject hump a table or surface β€” grinding against furniture. Use for table-humping scenes. Not for partnered sex; it's a solo furniture act.", }, 'nsfw_lesbian_cunnilingus_v2_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Lesbian Cunnilingus v2.0 - MinimaxH3.safetensors', "scale": 0.7, "label": 'Lesbian Cunnilingus v2.0', "trigger": 'lesbian cunnilingus', "desc": "Lesbian cunnilingus β€” one woman performs oral sex on another. Use for lesbian oral scenes. Not for solo or hetero scenes; it's a partnered female act.", }, 'nsfw_licking_penis_and_testicles_nsfw_v1_2': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Licking Penis and Testicles (NSFW) v1.2 - MinimaxH3 - licking penis,licking testicles,licking penis tip,Only works if penis and testicles are in view.safetensors', "scale": 0.7, "label": 'Licking Penis and Testicles (NSFW) v1.2', "trigger": 'licking penis,licking testicles,licking penis tip', "desc": "Makes the subject lick a penis and testicles β€” detailed oral attention. Use for oral scenes focused on the penis and testicles. Only works if the penis and testicles are in view.", }, 'nsfw_mini_dick_fix_penis_helper': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Mini Dick Fix - Penis Helper - MinimaxH3 - 3r3kt (but not really needed).safetensors', "scale": 0.7, "label": 'Mini Dick Fix - Penis Helper', "trigger": '3r3kt', "desc": "Fixes the 'mini dick' problem β€” makes the penis render at a realistic size instead of shrinking. Use alongside any penis LoRA. Not a standalone anatomy LoRA; it's a helper.", }, 'nsfw_missionary_sex_pov': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Missionary Sex - POV - MinimaxH3 - having missionary sex.safetensors', "scale": 0.7, "label": 'Missionary Sex - POV', "trigger": 'having missionary sex', "desc": "Missionary sex from the male POV β€” the act reads naturally from above. Use for missionary scenes. Not for other positions; it's position-specific.", }, 'nsfw_moawxx_female_pleasure_moans_body_writhi': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Moawxx - Female Pleasure Moans + Body Writhing v0.1 2000e - MinimaxH3 - moawxx,0.6-0.85str.safetensors', "scale": 0.72, "label": 'Moawxx - Female Pleasure Moans + Body Writhing v0.1 2000e', "trigger": 'moawxx', "desc": "Moawxx pleasure LoRA β€” female moans plus body writhing during stimulation. Use for pleasure and orgasm scenes. Not for calm scenes; it adds audible and visible arousal.", }, 'nsfw_moawxx_female_pleasure_moans_body_writhi_2': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Moawxx - Female Pleasure Moans + Body Writhing v0.1 2500e - MinimaxH3 - moawxx,0.6-0.85str.safetensors', "scale": 0.72, "label": 'Moawxx - Female Pleasure Moans + Body Writhing v0.1 2500e', "trigger": 'moawxx', "desc": "Moawxx pleasure LoRA β€” female moans plus body writhing during stimulation (2500e training variant). Use for pleasure and orgasm scenes. Not for calm scenes; it adds audible and visible arousal.", }, 'nsfw_moawxx_female_pleasure_moans_body_writhi_3': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Moawxx - Female Pleasure Moans + Body Writhing v0.1 2750e - MinimaxH3 - moawxx,0.6-0.85str.safetensors', "scale": 0.72, "label": 'Moawxx - Female Pleasure Moans + Body Writhing v0.1 2750e', "trigger": 'moawxx', "desc": "Moawxx pleasure LoRA β€” female moans plus body writhing during stimulation (2750e training variant). Use for pleasure and orgasm scenes. Not for calm scenes; it adds audible and visible arousal.", }, 'nsfw_mystic_xxx_v2_0': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Mystic XXX v2.0 - MinimaxH3 - 0.7str.safetensors', "scale": 0.7, "label": 'Mystic XXX v2.0', "trigger": 'mystic xxx', "desc": "General NSFW enhancer β€” improves overall explicitness and scene quality. Use as a base on top of specific act LoRAs. Not a specific act; it's a general boost.", }, 'nsfw_natural_sagging_breasts': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Natural Sagging Breasts - MinimaxH3 - She has naturally sagging breasts.safetensors', "scale": 0.7, "label": 'Natural Sagging Breasts', "trigger": 'She has naturally sagging breasts', "desc": "Renders naturally sagging breasts β€” realistic gravity and shape. Use for mature or natural figures. Not for perky or enhanced breasts; it adds sag.", }, 'nsfw_natural_saggy_boobs': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Natural Saggy Boobs - MinimaxH3 - hugesaggytits,shirt-lift titty drop,bra unhook from behind,then drop.safetensors', "scale": 0.7, "label": 'Natural Saggy Boobs', "trigger": 'hugesaggytits', "desc": "Saggy boobs with specific reveal actions β€” shirt-lift titty drop and bra unhook from behind. Use for breast-reveal scenes. Not for clothed scenes; it's about the reveal.", }, 'nsfw_nipple_play': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Nipple Play - MinimaxH3 - flicks her own nipples,gently pinch and slowly roll her own beautifully formed,fully erect,firm and symmetrical nipples.safetensors', "scale": 0.7, "label": 'Nipple Play', "trigger": 'nipple play', "desc": "Makes the subject play with her nipples β€” flicking, pinching, and rolling them. Use for nipple-play scenes. Not for scenes without breast focus; it centers on the nipples.", }, 'nsfw_pjslide_pussyjob': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'PJSLIDE - Pussyjob - MinimaxH3 - 0.5str,PJSLIDE,add HMPenis lora at 0.5str.safetensors', "scale": 0.5, "label": 'PJSLIDE - Pussyjob', "trigger": 'PJSLIDE', "desc": "Pussyjob LoRA β€” the penis slides along the vulva. Use for pussyjob scenes; the author recommends adding the penis LoRA at 0.5 strength alongside it. Not for penetration scenes; it's a sliding act.", }, 'nsfw_pk_parasyte_turbo_clean_up_the_models_vi': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'PK Parasyte Turbo (clean up the models visuals and audio) - MinimaxH3 - 3.5-5str,6-8steps,8-3 shift,euler+simple.safetensors', "scale": 4.25, "label": 'PK Parasyte Turbo (clean up the models visuals and audio)', "trigger": 'PK Parasyte Turbo', "desc": "Turbo LoRA that cleans up model visuals and audio β€” improves overall quality at low step counts. Use with turbo mode (6-8 steps, high strength 3.5-5). Not for standard 25-step generation; it's tuned for turbo.", }, 'nsfw_pov_missionary_insertion_v0_7': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'POV Missionary Insertion v0.7 - MinimaxH3.safetensors', "scale": 0.7, "label": 'POV Missionary Insertion v0.7', "trigger": 'missionary insertion', "format": "kohya", "desc": "POV missionary insertion β€” the insertion moment reads clearly from the male POV. Use for missionary insertion scenes. Not for other positions; it's position-specific.", }, 'nsfw_passionate_kissing': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Passionate Kissing - MinimaxH3 - They kiss passionately,1str.safetensors', "scale": 1.0, "label": 'Passionate Kissing', "trigger": 'They kiss passionately', "desc": "Makes passionate kissing read naturally β€” lip contact and head motion. Use for kissing scenes. Not for oral or sex acts; it's kissing-specific.", }, 'nsfw_photorealistic_image_generator': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Photorealistic Image Generator - MinimaxH3 - ph0t0r34l.safetensors', "scale": 0.7, "label": 'Photorealistic Image Generator', "trigger": 'ph0t0r34l', "format": "kohya", "desc": "Pushes output toward photorealism β€” more realistic textures and lighting. Use for photorealistic scenes. Not for stylized or animated content; it fights the style.", }, 'nsfw_pillow_humping': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Pillow Humping - MinimaxH3 - She is straddling a pillow.She is humping the pillow.safetensors', "scale": 0.7, "label": 'Pillow Humping', "trigger": 'She is straddling a pillow.She is humping the pillow', "desc": "Makes the subject straddle and hump a pillow β€” solo grinding on a pillow. Use for pillow-humping scenes. Not for partnered sex; it's a solo act.", }, 'nsfw_pokies': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Pokies - MinimaxH3 - covered nipples,pokies.safetensors', "scale": 0.7, "label": 'Pokies', "trigger": 'covered nipples,pokies', "format": "kohya", "desc": "Renders visible nipple outlines through clothing β€” pokies. Use for clothed scenes with tight tops. Not for nude scenes; it needs clothing.", }, 'nsfw_polyhedron_perfect_eyes_perfect_skin_per': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Polyhedron Perfect Eyes - Perfect Skin - Perfect Hands T2VA - MinimaxH3 - perfe8ct,perfect hands,perfect skin,perfect eyes.safetensors', "scale": 0.7, "label": 'Polyhedron Perfect Eyes - Perfect Skin - Perfect Hands T2VA', "trigger": 'perfe8ct', "desc": "Quality enhancer for eyes, skin, and hands β€” fixes common rendering flaws. Use as a general quality boost for any scene. Not for specific acts; it's a detail fixer.", }, 'nsfw_realism_people': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Realism People - MinimaxH3 - r34l1sm,0.6-1str.safetensors', "scale": 0.8, "label": 'Realism People', "trigger": 'r34l1sm', "desc": "Makes people look more realistic β€” natural proportions and skin. Use for photorealistic human scenes. Not for stylized content; it fights the style.", }, 'nsfw_sexgod_after_midnight_sexytime_r64': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'SexGod After Midnight Sexytime R64 - MinimaxH3 - euler+beta.safetensors', "scale": 0.7, "label": 'SexGod After Midnight Sexytime R64', "trigger": 'sexgod', "format": "kohya", "desc": "SexGod base LoRA β€” general NSFW scene quality. Use as a base enhancer on top of specific act LoRAs. Not a specific act; it's a general boost.", }, 'nsfw_sexgod_s_naughtytimes_v2_0_pruned_rank25': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": "SexGod's NaughtyTimes v2.0 pruned rank256 - MinimaxH3.safetensors", "scale": 0.7, "label": "SexGod's NaughtyTimes v2.0 pruned rank256", "trigger": 'naughtytimes', "format": "kohya", "desc": "SexGod's NaughtyTimes β€” general NSFW scene enhancer (rank-256 pruned). Use as a base enhancer. Not act-specific; pair it with act LoRAs.", }, 'nsfw_sexgod_s_naughtytimes_v2_0_pruned_rank51': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": "SexGod's NaughtyTimes v2.0 pruned rank512 - MinimaxH3.safetensors", "scale": 0.7, "label": "SexGod's NaughtyTimes v2.0 pruned rank512", "trigger": 'naughtytimes', "format": "kohya", "desc": "SexGod's NaughtyTimes β€” general NSFW scene enhancer (rank-512 pruned, slightly higher capacity). Use as a base enhancer. Pick one of the two NaughtyTimes variants; don't stack both.", }, 'nsfw_she_struts': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'She Struts - MinimaxH3 - she struts.safetensors', "scale": 0.7, "label": 'She Struts', "trigger": 'she struts', "desc": "Makes the subject strut β€” confident walking with attitude. Use for walk or entrance scenes. Not for static scenes; it needs locomotion.", }, 'nsfw_sideview_deepthroat_gag_audio': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'SideView Deepthroat Gag Audio - MinimaxH3.safetensors', "scale": 0.7, "label": 'SideView Deepthroat Gag Audio', "trigger": 'deepthroat', "format": "kohya", "desc": "Side-view deepthroat with gag audio β€” the gag reflex reads and sounds. Use for deepthroat scenes. Not for gentle oral; it's intense.", }, 'nsfw_skin_tight_bottoms_cameltoe': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Skin Tight Bottoms (Cameltoe) - MinimaxH3 - cameltoe,0.9str.safetensors', "scale": 0.9, "label": 'Skin Tight Bottoms (Cameltoe)', "trigger": 'cameltoe', "desc": "Skin-tight bottoms with a visible cameltoe. Use for tight-clothing scenes. Not for nude scenes; it needs the clothing.", }, 'nsfw_slop_twerk_v0_2_epoch6': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": "Slop Twerk v0.2 Epoch6 - MinimaxH3 - twerks causing her ass to jiggle and shake,0.5-1str,0.5 STR if you don't want Asians in T2V.safetensors", "scale": 0.75, "label": 'Slop Twerk v0.2 Epoch6', "trigger": 'twerk', "desc": "Twerk LoRA with strong ass jiggle and shake. Use for twerk scenes. Note: use 0.5 strength if you don't want Asian subjects in text-to-video.", }, 'nsfw_slop_bounce': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Slop bounce - MinimaxH3 - breasts bounce,shake,jiggle,sway.safetensors', "scale": 0.7, "label": 'Slop bounce', "trigger": 'breasts bounce', "desc": "Breast bounce, shake, jiggle, and sway during movement. Use for any scene with breast motion. Not for static shots; it only shows with movement.", }, 'nsfw_snow_white_lingerie': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Snow White Lingerie - MinimaxH3 - She is wearing slutty snow white lingerie.safetensors', "scale": 0.7, "label": 'Snow White Lingerie', "trigger": 'She is wearing slutty snow white lingerie', "desc": "Renders the subject in slutty Snow White lingerie. Use for scenes featuring this specific outfit. Not for other outfits; it's tied to this garment.", }, 'nsfw_strapon_dildo_v0_1': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Strapon Dildo v0.1 - MinimaxH3 - She is wearing a strapon dildo.safetensors', "scale": 0.7, "label": 'Strapon Dildo v0.1', "trigger": 'She is wearing a strapon dildo', "desc": "Makes the subject wear and use a strapon dildo. Use for strapon scenes. Not for scenes without the toy; it won't add anything.", }, 'nsfw_synth_pussy_v0_5': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Synth Pussy v0.5 - MinimaxH3.safetensors', "scale": 0.7, "label": 'Synth Pussy v0.5', "trigger": 'synth pussy', "format": "kohya", "desc": "Synth pussy β€” a stylized or synthetic vulva rendering. Use for stylized genital shots. Not for photorealistic anatomy; it's a stylized look.", }, 'nsfw_the_orgasm_button': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'The Orgasm button - MinimaxH3 - Remoteorgasm.safetensors', "scale": 0.7, "label": 'The Orgasm button', "trigger": 'Remoteorgasm', "desc": "Remote-control orgasm β€” the subject reacts to a remote trigger. Use for orgasm-control scenes. Not for natural orgasms; it's the remote-control concept.", }, 'nsfw_thumb_finger_puppets': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Thumb-Finger puppets - MinimaxH3 - A thumb person.safetensors', "scale": 0.7, "label": 'Thumb-Finger puppets', "trigger": 'A thumb person', "desc": "Thumb-finger puppets β€” a 'thumb person' novelty effect. Use for puppet or novelty scenes. Not for realistic content; it's a gag.", }, 'nsfw_thumbinbutt': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'ThumbInButt - MinimaxH3 - thum1n8utt,the man inserts his right thumb in her anus causing the woman to moan with pleasure.safetensors', "scale": 0.7, "label": 'ThumbInButt', "trigger": 'thum1n8utt', "desc": "Makes the man insert his thumb into the woman's anus, causing a pleasure moan. Use for anal-fingering scenes. Not for non-anal scenes; it forces the action.", }, 'nsfw_tiddies_and_realism_slider_vbeta': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Tiddies and Realism Slider vBeta - MinimaxH3 - 0.4-2str.safetensors', "scale": 1.2, "label": 'Tiddies and Realism Slider vBeta', "trigger": 'tiddies and realism', "desc": "Slider LoRA β€” adjust breast size and realism on a scale (0.4-2 strength; 1.2 is a good default). Use to fine-tune breast size. Not a fixed anatomy LoRA; it's a continuous slider.", }, 'nsfw_tiny_breasts_flat_chest': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Tiny Breasts - Flat Chest - MinimaxH3 - She is flat chested,she has a flat chest.safetensors', "scale": 0.7, "label": 'Tiny Breasts - Flat Chest', "trigger": 'She is flat chested,she has a flat chest', "desc": "Renders a flat chest or tiny breasts. Use for petite figures. Not for busty figures; it conflicts with breast-size LoRAs.", }, 'nsfw_tiny_panties': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Tiny Panties - MinimaxH3 - She is wearing tiny panties.safetensors', "scale": 0.7, "label": 'Tiny Panties', "trigger": 'She is wearing tiny panties', "desc": "Renders the subject in tiny panties β€” minimal coverage. Use for lingerie scenes. Not for nude scenes; the panties are the point.", }, 'nsfw_torpedo_tits_t2v': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Torpedo Tits T2V - MinimaxH3 - torpedo tits,a 20 year old woman,wearing nothing shows off her medium sized torpedo tit breasts.safetensors', "scale": 0.7, "label": 'Torpedo Tits T2V', "trigger": 'torpedo tits', "desc": "Torpedo-shaped breasts β€” a specific conical breast shape. Use for torpedo-tit scenes. Not for natural shapes; it imposes the torpedo look.", }, 'nsfw_transition_boomsplitro': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Transition BoomSplitRo - MinimaxH3 - ysp_spliroa12 a woman. hard cut with no smooth transition, jumps straight into spitroast.safetensors', "scale": 0.7, "label": 'Transition BoomSplitRo', "trigger": 'ysp_spliroa12', "format": "kohya", "desc": "Hard-cut transition straight into a spitroast scene β€” no smooth transition. Use for abrupt scene changes. Not for continuous scenes; it's a cut effect.", }, 'nsfw_transition_f2v_instant_cut_to_things': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Transition F2V - Instant Cut to Things - MinimaxH3 - ysp_smshcut.safetensors', "scale": 0.7, "label": 'Transition F2V - Instant Cut to Things', "trigger": 'ysp_smshcut', "format": "kohya", "desc": "Instant cut transition to a new scene. Use for abrupt scene changes. Not for continuous scenes; it's a cut effect.", }, 'nsfw_tribtime': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Tribtime - MinimaxH3 - tribtime.safetensors', "scale": 0.7, "label": 'Tribtime', "trigger": 'tribtime', "desc": "Tribbing β€” scissoring or grinding between two women. Use for tribadism scenes. Not for solo scenes; it needs two partners.", }, 'nsfw_upskirt_helper_v0_1': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Upskirt Helper v0.1 - MinimaxH3 - upsk1rt_panties,upsk1rt_no_panties.safetensors', "scale": 0.7, "label": 'Upskirt Helper v0.1', "trigger": 'upsk1rt_panties,upsk1rt_no_panties', "format": "kohya", "desc": "Upskirt view β€” with or without panties. Use for upskirt shots. Not for nude scenes; it needs a skirt.", }, 'nsfw_whispering': { "repo": "EllaPriest45/MinimaxH3_Actions", "file": 'Whispering - MinimaxH3 - 0.7str,she whisper XYZ.safetensors', "scale": 0.7, "label": 'Whispering', "trigger": 'she whisper XYZ', "desc": "Makes the subject whisper β€” replace XYZ in the trigger with what she whispers. Use for whispering scenes. Not for loud or vocal scenes; it's a quiet effect.", }, } # LoRAs that overlap in purpose and fight each other when stacked. The description box warns whenever 2+ from a # group are selected. Groups may overlap (e.g. a deepthroat LoRA is both a blowjob act and an audio LoRA). LORA_CONFLICT_GROUPS = { "penis anatomy": ["nsfw_penis", "nsfw_hmpenis_v2_0"], "innie pussy": ["nsfw_innie", "nsfw_innie_pussy"], "breast size/shape": ["nsfw_bustywomen", "nsfw_tiny_breasts_flat_chest", "nsfw_huge_breasts_large_breasts", "nsfw_torpedo_tits_t2v", "nsfw_tiddies_and_realism_slider_vbeta", "nsfw_natural_sagging_breasts", "nsfw_natural_saggy_boobs"], "breast motion": ["nsfw_boob_physics", "nsfw_slop_bounce", "nsfw_jiggle_walk"], "cameltoe": ["nsfw_cameltoe_v2_0", "nsfw_skin_tight_bottoms_cameltoe"], "twerk": ["nsfw_icy_twerk_pro_max_v2_0", "nsfw_slop_twerk_v0_2_epoch6"], "squirting": ["nsfw_hmmasturbation_squirting_and_female_mast", "nsfw_hmsquirt_squirting_v0_1"], "blowjob act": ["nsfw_blowjob_v3_0", "nsfw_blowjob_wearing_thong", "nsfw_chasing_blowjob_v1_2", "nsfw_sideview_deepthroat_gag_audio", "nsfw_deeper_hold_it", "nsfw_licking_penis_and_testicles_nsfw_v1_2"], "cumshot": ["nsfw_cumshot", "nsfw_facial_cumshot_v0_1"], "scene transition": ["nsfw_transition_boomsplitro", "nsfw_transition_f2v_instant_cut_to_things"], "moans (same LoRA, different training epochs)": ["nsfw_moawxx_female_pleasure_moans_body_writhi", "nsfw_moawxx_female_pleasure_moans_body_writhi_2", "nsfw_moawxx_female_pleasure_moans_body_writhi_3"], "SexGod NaughtyTimes (same LoRA, different ranks)": ["nsfw_sexgod_s_naughtytimes_v2_0_pruned_rank25", "nsfw_sexgod_s_naughtytimes_v2_0_pruned_rank51"], "general NSFW enhancer": ["nsfw_astro_nsfw", "nsfw_mystic_xxx_v2_0", "nsfw_sexgod_after_midnight_sexytime_r64", "nsfw_sexgod_s_naughtytimes_v2_0_pruned_rank25", "nsfw_sexgod_s_naughtytimes_v2_0_pruned_rank51"], "visual style": ["nsfw_disney_style_renaissance_era", "nsfw_photorealistic_image_generator", "nsfw_realism_people"], "missionary POV": ["nsfw_missionary_sex_pov", "nsfw_pov_missionary_insertion_v0_7"], "dildo": ["nsfw_dildo", "nsfw_dildo_masturbation_v2_0", "nsfw_strapon_dildo_v0_1"], "audio": ["nsfw_jav_moans_audio_only", "nsfw_whispering", "nsfw_sideview_deepthroat_gag_audio"], "hip motion": ["nsfw_grindtime", "nsfw_hip_gyration"], } # Must stay identical to the conditioner's table: the *label* goes over the wire, so a canvas that half does not know # is rejected there and surfaces as a failure here. CANVASES = { # 16:9 "960x544 Β· 16:9 fast": (544, 960), "1024x576 Β· 16:9 fast": (576, 1024), "1152x640 Β· 16:9": (640, 1152), "1280x704 Β· 16:9": (704, 1280), "1344x768 Β· 16:9 full": (768, 1344), # 9:16 "544x960 Β· 9:16 fast": (960, 544), "640x1152 Β· 9:16": (1152, 640), "768x1344 Β· 9:16 full": (1344, 768), # 1:1 "544x544 Β· 1:1 fast": (544, 544), "768x768 Β· 1:1 full": (768, 768), # 4:3 / 3:4 "768x576 Β· 4:3 fast": (576, 768), "1024x768 Β· 4:3 full": (768, 1024), "576x768 Β· 3:4 fast": (768, 576), "768x1024 Β· 3:4 full": (1024, 768), # 21:9 "1152x512 Β· 21:9 fast": (512, 1152), "1536x672 Β· 21:9 full": (672, 1536), } DEFAULT_CANVAS = "960x544 Β· 16:9 fast" FPS, FRAMES_PER_CHUNK, LATENTS_PER_CHUNK = 24, 17, 5 # It is the *snapped* frame count the ceiling has to hold for: 15 s is 360 frames, which rounds up to 362, i.e. # 15.083 s, and is refused. MIN_UI_DURATION, MAX_UI_DURATION = 2, 14 def snap_frames(seconds: float) -> int: """The frame count MiniMax-H3's video VAE can decode: the next `17 * n + 5` at 24 fps.""" frames = max(1, round(float(seconds) * FPS)) while frames % FRAMES_PER_CHUNK != LATENTS_PER_CHUNK: frames += 1 return frames def lower_duration_floor(seconds: float = MIN_UI_DURATION) -> None: """Let the pipeline generate below its 5 s floor. 56 frames (2.33 s) is fine on the released checkpoint.""" from diffusers.modular_pipelines.minimax_h3.modular_pipeline import MiniMaxH3ModularPipeline MiniMaxH3ModularPipeline.min_duration = property(lambda self: float(seconds)) PIPE = None MANAGER = None LOAD_ERROR: str | None = None LOADED_IN: float | None = None TURBO_AVAILABLE: bool = False _TURBO_LoRA: list | None = None # list of (module, A, B) tuples β€” A/B factors on CPU, tiny footprint _TURBO_merged = False _NSFW_LoRA: dict[str, list] = {} # key -> list of (module, A, B) tuples, A/B on CPU _NSFW_merged: set[str] = set() _CUSTOM_LoRA: dict[tuple[str, str], list] = {} # (repo, file) -> list of (module, A, B) tuples, A/B on CPU _CUSTOM_merged: set[tuple[str, str]] = set() def _convert_comfyui_h3_lora(sd): """Convert ComfyUI-format MiniMax-H3 LoRA keys to the module-tree keys the loader navigates. ComfyUI keys look like `diffusion_model.blocks.N.attn.out_proj.lora_A.weight` (and `diffusion_model.token_refiner.blocks.N.attn.out_proj.lora_A.weight` for the token refiner). The diffusers module tree uses `transformer_blocks.N.attn.to_out.0` / `token_refiner.refiner_blocks.N.attn.to_out.0`, `attn.to_q/to_k/to_v` (fused `qkv_proj` split 3-way), and `ff.net.0.proj` / `ff.net.2` (for `mlp.fc1` / `mlp.fc2`). """ out = {} for k, v in sd.items(): k = k.removeprefix("diffusion_model.").removeprefix("transformer.") parts = k.split(".") # Main blocks: [blocks, N, attn|mlp, target, lora_A|B, weight] # Token refiner: [token_refiner, blocks, N, attn|mlp, target, lora_A|B, weight] if parts[0] == "token_refiner": prefix = "token_refiner.refiner_blocks." block, layer, target = parts[2], parts[3], parts[4] suffix = ".".join(parts[5:]) # e.g. "lora_A.weight" else: prefix = "transformer_blocks." block, layer, target = parts[1], parts[2], parts[3] suffix = ".".join(parts[4:]) # e.g. "lora_A.weight" # Some ComfyUI exports also carry kohya-style `.alpha` scalars (alpha == rank here, so a no-op). # Skip any non-lora_A/B key so metadata never reaches the shape-sensitive qkv split below. if suffix not in ("lora_A.weight", "lora_B.weight"): continue if layer == "attn": if target == "out_proj": out[f"{prefix}{block}.attn.to_out.0.{suffix}"] = v elif target == "qkv_proj": if suffix == "lora_A.weight": # A is shared across q/k/v for name in ("to_q", "to_k", "to_v"): out[f"{prefix}{block}.attn.{name}.lora_A.weight"] = v else: # lora_B.weight β€” split the rows 3-way into q/k/v head_dim = v.shape[0] // 3 for i, name in enumerate(("to_q", "to_k", "to_v")): out[f"{prefix}{block}.attn.{name}.lora_B.weight"] = v[i * head_dim:(i + 1) * head_dim] elif layer == "mlp": if target == "fc1": out[f"{prefix}{block}.ff.net.0.proj.{suffix}"] = v elif target == "fc2": out[f"{prefix}{block}.ff.net.2.{suffix}"] = v return out def _convert_kohya_h3_lora(sd): """Convert kohya/musubi-tuner-format MiniMax-H3 LoRA keys to ComfyUI format, then to module-tree keys. kohya keys look like `lora_unet_blocks_N_attn_out_proj.lora_down.weight` / `.lora_up.weight` / `.alpha` (also `attn_qkv_proj`, `mlp_fc1`, `mlp_fc2`). Some exports are hybrids: the ComfyUI-style prefix (`diffusion_model.blocks.N.attn.out_proj`) with kohya-style suffixes β€” the prefix is already the ComfyUI layout, so only the suffix needs renaming. We first rename to the ComfyUI layout the loader understands (`diffusion_model.blocks.N.attn.out_proj.lora_A/lora_B.weight`), folding the per-module `alpha/rank` factor into the `lora_up` (B) tensor so the loader's `B @ A * scale` stays correct, then hand off to `_convert_comfyui_h3_lora`. kohya has no `token_refiner` keys β€” those are simply absent. """ import torch comfy = {} alphas: dict[str, float] = {} # First pass: collect alpha factors (alpha/rank) per module base name. `.alpha` is a 0-dim # scalar tensor (ndim == 0), so the rank must come from the lora_down weight's shape[0]. for k, v in sd.items(): if k.endswith(".alpha"): base = k[: -len(".alpha")] _down = sd.get(f"{base}.lora_down.weight") rank = _down.shape[0] if _down is not None else float(v) alphas[base] = float(v) / float(rank) for k, v in sd.items(): if k.endswith(".alpha"): continue if k.endswith(".lora_down.weight"): base = k[: -len(".lora_down.weight")] lora = "lora_A" elif k.endswith(".lora_up.weight"): base = k[: -len(".lora_up.weight")] lora = "lora_B" else: continue if base.startswith("diffusion_model.") or base.startswith("transformer."): # Hybrid export: ComfyUI-style prefix with kohya-style suffixes # (diffusion_model.blocks.N.attn.out_proj.lora_down.weight). The prefix is already the # ComfyUI layout β€” only the suffix needs renaming. token_refiner keys, if any, flow # through unchanged too. comfy_key = f"{base}.{lora}.weight" else: # base: lora_unet_blocks_N_attn_out_proj -> diffusion_model.blocks.N.attn.out_proj parts = base.split("_") # parts: [lora, unet, blocks, N, attn, out, proj] (mlp_fc1 -> [.., mlp, fc1]) if len(parts) < 6: continue # unknown key shape β€” skip rather than crash block = parts[3] layer = parts[4] target = "_".join(parts[5:]) comfy_key = f"diffusion_model.blocks.{block}.{layer}.{target}.{lora}.weight" if lora == "lora_B" and base in alphas: v = v * alphas[base] comfy[comfy_key] = v return _convert_comfyui_h3_lora(comfy) def _apply_turbo(): """Merge LoRA into transformer weights using A/B factors. The precomputed delta (B @ A) is ~57 MB per module Γ— 363 = ~20 GB β€” far more GPU memory than is free after the transformer is unpacked (~14 GB on an 80 GB card). Instead we store the LoRA factors A (~2.6 MB) and B (~2.6 MB) per module β€” total ~1.2 GB β€” and compute the delta on-the-fly: mod.weight += B @ A * scale. Peak extra GPU memory is one temporary delta tensor (~57 MB), freed each iteration. A/B factors stay CPU-resident in `_TURBO_LoRA`; local CUDA copies are made per merge so the store is never mutated and no CUDA memory is held between requests. """ global _TURBO_merged, _TURBO_LoRA if _TURBO_merged or _TURBO_LoRA is None: return _t0 = time.time() for mod, A, B in _TURBO_LoRA: # delta = B @ A is temporary (~57 MB), freed next iteration β€” never stored simultaneously. mod.weight.data += (B.to("cuda", non_blocking=True) @ A.to("cuda", non_blocking=True)) * TURBO_LORA_SCALE _elapsed = time.time() - _t0 _TURBO_merged = True _log(f"turbo LoRA merged ({len(_TURBO_LoRA)} modules in {_elapsed:.2f}s)") def _remove_turbo(): """Restore base weights by undoing the LoRA merge. Idempotent: only acts when `_TURBO_merged`, and rolls back in reverse order so a partial failure leaves the fewest modules touched.""" global _TURBO_merged if not _TURBO_merged or _TURBO_LoRA is None: return _t0 = time.time() _done = 0 try: for mod, A, B in reversed(_TURBO_LoRA): mod.weight.data -= (B.to("cuda", non_blocking=True) @ A.to("cuda", non_blocking=True)) * TURBO_LORA_SCALE _done += 1 except Exception as _e: _log(f"turbo LoRA partial unmerge ({_done}/{len(_TURBO_LoRA)}): {_e}", "ERROR") raise _elapsed = time.time() - _t0 _TURBO_merged = False import torch as _torch _torch.cuda.empty_cache() _log(f"turbo LoRA removed ({len(_TURBO_LoRA)} modules in {_elapsed:.2f}s)") # ── Debug logging ────────────────────────────────────────────────────────────── def _log(msg, level="INFO"): """Timestamped line to the Space logs (stdout). Levels: DEBUG / INFO / WARN / ERROR.""" print(f"[gen][{time.strftime('%H:%M:%S')}][{level}] {msg}", flush=True) # ── Memory / OOM diagnostics ─────────────────────────────────────────────────── # The Space has a history of OOM crashes. A hard kill leaves no traceback, so the phase # logs + bucket dumps below are the only evidence. `_log_mem` snapshots GPU + CPU at phase # boundaries; `_debug_dump` persists diagnostics to the bucket under /data/debug; # `_oom_dump` captures a CUDA OOM (traceback + torch memory summary) before re-raising. # `H3_MEM_SAMPLE_INTERVAL` (s, 0 = off) drives the in-generation sampler thread. _DEBUG_DIR = os.environ.get("H3_DEBUG_DIR", "/data/debug") _MEM_SAMPLE_INTERVAL = float(os.environ.get("H3_MEM_SAMPLE_INTERVAL", "5")) def _mem_info(): """Compact GPU + CPU memory snapshot for the logs. Never raises.""" _parts = [] try: import torch if torch.cuda.is_available(): _free, _total = torch.cuda.mem_get_info() _parts.append( f"gpu_alloc={torch.cuda.memory_allocated() / 2**30:.2f}G " f"gpu_resv={torch.cuda.memory_reserved() / 2**30:.2f}G " f"gpu_max={torch.cuda.max_memory_allocated() / 2**30:.2f}G " f"gpu_free={_free / 2**30:.2f}G/{_total / 2**30:.2f}G" ) except Exception: pass try: import psutil _parts.append(f"rss={psutil.Process().memory_info().rss / 2**30:.2f}G") except Exception: try: with open("/proc/self/status") as _f: for _line in _f: if _line.startswith("VmRSS:"): _parts.append(f"rss={int(_line.split()[1]) / 2**20:.2f}G") break except Exception: pass return " ".join(_parts) or "mem=n/a" def _log_mem(tag): """Log a memory snapshot under the greppable `MEM` level.""" _log(f"{tag}: {_mem_info()}", "MEM") def _debug_dump(tag, text): """Persist a diagnostic to the bucket debug folder (/data/debug). Never raises.""" try: os.makedirs(_DEBUG_DIR, exist_ok=True) _path = os.path.join(_DEBUG_DIR, f"{tag}_{time.strftime('%Y%m%d_%H%M%S')}.txt") _body = ( f"=== {tag} @ {time.strftime('%Y-%m-%d %H:%M:%S')} ===\n" f"{_mem_info()}\n" f"{text}\n" ) _tmp = _path + ".tmp" with open(_tmp, "w", encoding="utf-8") as _f: _f.write(_body) os.replace(_tmp, _path) try: shutil.copyfile(_path, os.path.join(_DEBUG_DIR, f"{tag}_latest.txt")) except OSError: pass _log(f"debug dump written: {_path}") except Exception as _e: _log(f"debug dump failed ({_e}); see logs", "WARN") def _oom_dump(tag, exc, extra=""): """Capture a CUDA OOM: exception + context + torch memory summary, saved to the bucket.""" _text = [f"exception: {type(exc).__name__}: {exc}"] if extra: _text.append(extra) _text.append(traceback.format_exc()) try: import torch if torch.cuda.is_available(): _text.append(torch.cuda.memory_summary()) except Exception: pass _debug_dump(f"oom_{tag}", "\n".join(_text)) _log(f"OOM captured ({tag}): {exc}", "ERROR") # ── Bucket-backed LoRA cache ─────────────────────────────────────────────────── # Every downloaded LoRA is copied into the mounted bucket (/data) so a Space restart # loads it from the bucket instead of re-downloading it. Falls back to the HF cache # whenever the bucket isn't mounted or a write fails. `H3_LORA_CACHE_DIR` overrides # the mount path (used by tests). _LORA_CACHE_DIR = os.environ.get("H3_LORA_CACHE_DIR", "/data/loras") def _lora_cache_path(repo, file): """Bucket path for a (repo, file) LoRA, or None if the bucket isn't usable.""" try: os.makedirs(_LORA_CACHE_DIR, exist_ok=True) _probe = os.path.join(_LORA_CACHE_DIR, ".probe") with open(_probe, "w") as _f: _f.write("ok") os.remove(_probe) except Exception: return None _key = re.sub(r"[^A-Za-z0-9._-]", "_", f"{repo.replace('/', '__')}__{file}") return os.path.join(_LORA_CACHE_DIR, _key) def _fetch_lora_file(repo, file): """Local path to a LoRA .safetensors, preferring the bucket cache over a fresh download.""" from huggingface_hub import hf_hub_download _cached = _lora_cache_path(repo, file) if _cached and os.path.exists(_cached): try: with open(_cached, "rb") as _f: _n = struct.unpack("<Q", _f.read(8))[0] json.loads(_f.read(_n)) _log(f"LoRA cache hit: {repo}/{file} ({os.path.getsize(_cached) / 1e6:.0f} MB)") return _cached except Exception: _log(f"LoRA cache corrupt, re-downloading: {repo}/{file}", "WARN") try: os.remove(_cached) except OSError: pass _t0 = time.time() _path = hf_hub_download(repo, file) _log(f"LoRA downloaded: {repo}/{file} ({os.path.getsize(_path) / 1e6:.0f} MB in {time.time() - _t0:.0f}s)") if _cached: try: _tmp = _cached + ".tmp" with open(_path, "rb") as _src, open(_tmp, "wb") as _dst: shutil.copyfileobj(_src, _dst, 1024 * 1024) os.replace(_tmp, _cached) _log(f"LoRA cached to bucket: {repo}/{file}") except Exception as _e: _log(f"LoRA bucket cache write failed ({_e}); using HF cache", "WARN") return _path def _ensure_nsfw_lora_loaded(key): """Download + convert + store a lazy NSFW LoRA's A/B factors on CPU. No-op if already loaded.""" if key in _NSFW_LoRA: return cfg = NSFW_LORA_CONFIGS[key] import torch from safetensors.torch import load_file as _load_lora _t0 = time.time() _log(f"loading NSFW LoRA {key} from {cfg['repo']} ...") _path = _fetch_lora_file(cfg["repo"], cfg["file"]) _raw = _load_lora(_path) if cfg.get("format") == "kohya": _sd = _convert_kohya_h3_lora(_raw) else: _sd = _convert_comfyui_h3_lora(_raw) _pairs: dict[str, dict[str, torch.Tensor]] = {} for _k, _v in _sd.items(): if ".lora_A." in _k: _pairs.setdefault(_k.split(".lora_A.")[0], {})["A"] = _v elif ".lora_B." in _k: _pairs.setdefault(_k.split(".lora_B.")[0], {})["B"] = _v _loaded = [] for _base, _pair in _pairs.items(): if "A" not in _pair or "B" not in _pair: continue _parts = _base.split(".") _obj = PIPE.transformer for _p in _parts: _obj = _obj[int(_p)] if _p.isdigit() else getattr(_obj, _p) _loaded.append((_obj, _pair["A"].to(_obj.weight.dtype), _pair["B"].to(_obj.weight.dtype))) del _sd, _pairs import gc as _gc; _gc.collect() _NSFW_LoRA[key] = _loaded _log(f"NSFW LoRA {key} parsed ({len(_loaded)} modules in {time.time() - _t0:.1f}s)") def _ensure_custom_lora_loaded(repo, file): """Download + convert + store a user-supplied Hugging Face LoRA's A/B factors on CPU. No-op if already loaded.""" key = (repo, file) if key in _CUSTOM_LoRA: return import torch from safetensors.torch import load_file as _load_lora _t0 = time.time() _log(f"loading custom LoRA {repo}/{file} ...") try: _path = _fetch_lora_file(repo, file) _raw = _load_lora(_path) except Exception as _e: raise gr.Error(f"Couldn't fetch custom LoRA `{repo}/{file}`: {_e}") _has_comfy = any(".lora_A." in _k for _k in _raw) _has_kohya = any(".lora_down.weight" in _k or _k.startswith("lora_unet_") for _k in _raw) if _has_comfy: _sd = _convert_comfyui_h3_lora(_raw) elif _has_kohya: _sd = _convert_kohya_h3_lora(_raw) else: raise gr.Error(f"`{repo}/{file}` doesn't look like a MiniMax-H3 LoRA (no lora_A/lora_down keys found).") _pairs: dict[str, dict[str, torch.Tensor]] = {} for _k, _v in _sd.items(): if ".lora_A." in _k: _pairs.setdefault(_k.split(".lora_A.")[0], {})["A"] = _v elif ".lora_B." in _k: _pairs.setdefault(_k.split(".lora_B.")[0], {})["B"] = _v _loaded = [] for _base, _pair in _pairs.items(): if "A" not in _pair or "B" not in _pair: continue _parts = _base.split(".") _obj = PIPE.transformer for _p in _parts: _obj = _obj[int(_p)] if _p.isdigit() else getattr(_obj, _p) _loaded.append((_obj, _pair["A"].to(_obj.weight.dtype), _pair["B"].to(_obj.weight.dtype))) if not _loaded: raise gr.Error(f"`{repo}/{file}` converted to no usable modules β€” is it a MiniMax-H3 LoRA?") del _sd, _pairs, _raw import gc as _gc; _gc.collect() _CUSTOM_LoRA[key] = _loaded _log(f"custom LoRA {repo}/{file} parsed ({len(_loaded)} modules in {time.time() - _t0:.1f}s)") def _parse_hf_link(url): """Parse a Hugging Face repo or file link into (repo, file). `file` is None for a bare repo link.""" url = (url or "").strip().split("?")[0].split("#")[0] if not url: raise gr.Error("Paste a Hugging Face link first.") if not url.startswith("http"): parts = [p for p in url.split("/") if p] if len(parts) >= 2: return parts[0] + "/" + parts[1], "/".join(parts[2:]) or None raise gr.Error(f"Couldn't parse `{url}` as a Hugging Face repo link.") if "huggingface.co" not in url: raise gr.Error("Only Hugging Face links are supported (huggingface.co).") parts = [p for p in url.split("huggingface.co", 1)[1].strip("/").split("/") if p] if parts and parts[0] in ("spaces", "models", "datasets"): parts = parts[1:] # https://huggingface.co/spaces/Owner/Repo β†’ Owner/Repo if len(parts) < 2: raise gr.Error(f"Couldn't parse `{url}` as a Hugging Face repo link.") repo = parts[0] + "/" + parts[1] if len(parts) == 2: return repo, None if parts[2] in ("resolve", "tree", "blob"): if len(parts) == 4 and parts[2] in ("tree", "blob"): return repo, None # tree/main or blob/main with no subpath β†’ treat as a repo link if len(parts) < 5: raise gr.Error(f"Couldn't parse a filename from `{url}`.") if parts[3] != "main": raise gr.Error("Only `main`-branch links are supported β€” use a `resolve/main/...` link.") return repo, "/".join(parts[4:]) return repo, "/".join(parts[2:]) def _resolve_repo_file(repo): """Pick the single .safetensors file of a repo link; error if there isn't exactly one.""" from huggingface_hub import list_repo_files try: files = [f for f in list_repo_files(repo) if f.endswith(".safetensors")] except Exception as _e: raise gr.Error(f"Couldn't list files in `{repo}`: {_e}") if len(files) == 1: return files[0] if not files: raise gr.Error(f"No .safetensors files found in `{repo}`.") raise gr.Error(f"`{repo}` has multiple .safetensors files β€” paste a direct file link instead: " + ", ".join(files[:5]) + ("…" if len(files) > 5 else "")) def _download_custom_lora(url, scale=1.0, trigger=""): """Download + verify a user-supplied Hugging Face LoRA and report status. Loads it for the next Generate.""" try: repo, file = _parse_hf_link(url) if file is None: file = _resolve_repo_file(repo) _ensure_custom_lora_loaded(repo, file) except gr.Error as _e: return gr.update(value=f"**❌ {_e}**", visible=True) except Exception as _e: return gr.update(value=f"**❌ Unexpected error loading the LoRA: {_e}**", visible=True) n = len(_CUSTOM_LoRA[(repo, file)]) note = f" (scale {scale:g}" + (f", trigger `{trigger}`)" if trigger else ")") return gr.update( value=f"**βœ… Loaded `{repo}/{file}`** β€” {n} modules, will apply on Generate{note}.", visible=True, ) def _apply_lora(key, scale, store, merged): """Merge a loaded LoRA into transformer weights on GPU. A/B factors stay CPU-resident in `store`; local CUDA copies are made per merge so the store is never mutated.""" if key in merged or key not in store: return _t0 = time.time() for _mod, _A, _B in store[key]: _mod.weight.data += (_B.to("cuda", non_blocking=True) @ _A.to("cuda", non_blocking=True)) * scale merged.add(key) _log(f"LoRA {key} merged ({len(store[key])} modules in {time.time() - _t0:.2f}s)") def _remove_lora(key, scale, store, merged): """Undo a LoRA merge. Idempotent: only acts when `key` is in `merged`, and rolls back in reverse order so a partial failure leaves the fewest modules touched.""" if key not in merged or key not in store: return _t0 = time.time() _done = 0 try: for _mod, _A, _B in reversed(store[key]): _mod.weight.data -= (_B.to("cuda", non_blocking=True) @ _A.to("cuda", non_blocking=True)) * scale _done += 1 except Exception as _e: _log(f"LoRA {key} partial unmerge ({_done}/{len(store[key])}): {_e}", "ERROR") raise merged.discard(key) import torch as _torch _torch.cuda.empty_cache() _log(f"LoRA {key} removed ({len(store[key])} modules in {time.time() - _t0:.2f}s)") def _apply_nsfw(key): _apply_lora(key, NSFW_LORA_CONFIGS[key]["scale"], _NSFW_LoRA, _NSFW_merged) def _remove_nsfw(key): _remove_lora(key, NSFW_LORA_CONFIGS[key]["scale"], _NSFW_LoRA, _NSFW_merged) def _apply_custom(key, scale): _apply_lora(key, scale, _CUSTOM_LoRA, _CUSTOM_merged) def _remove_custom(key, scale): _remove_lora(key, scale, _CUSTOM_LoRA, _CUSTOM_merged) def status() -> str: if LOAD_ERROR: return LOAD_ERROR if PIPE is None: return f"Loading `{MODEL_REPO}` (transformer + VAEs, 77.3 GB). Watch the Space logs." import h3_aoti return ( f"Ready Β· transformer + VAEs **bfloat16, unquantized** Β· placement `{PLACEMENT}` Β· attention `{ATTENTION}` Β· " f"{h3_aoti.status()} Β· loaded in {LOADED_IN:.0f}s Β· conditioner `{CONDITIONER_SPACE}`" ) def load_models() -> str | None: """Load the denoising half at startup. `MiniMaxH3GeneratorBlocks` declares `transformer`, `vae`, `audio_vae`, the two schedulers and `video_processor`, so `load_components` fetches exactly those subfolders β€” `text_encoder/` and `transformer_ref/` are never touched. Both autoencoders carry `_keep_in_fp32_modules` over every module and stay float32: a bfloat16 audio VAE decodes the soundtrack roughly 20 dB too quiet. """ global PIPE, MANAGER, LOAD_ERROR, LOADED_IN if PIPE is not None or LOAD_ERROR is not None: return LOAD_ERROR started = time.time() try: _log_mem("startup begin") import torch from diffusers import ComponentsManager from h3_split_blocks import MiniMaxH3GeneratorBlocks lower_duration_floor() manager = ComponentsManager() blocks = MiniMaxH3GeneratorBlocks() _log(f"loading {[c.name for c in blocks.expected_components]} from {MODEL_REPO} ...") pipe = blocks.init_pipeline(MODEL_REPO, components_manager=manager, collection="h3") pipe.load_components(dtype=torch.bfloat16) pipe.transformer.set_attention_backend(ATTENTION) _log(f"components loaded in {time.time() - started:.0f}s") _log_mem("startup components loaded") if TURBO_LORA_REPO: from safetensors.torch import load_file as _load_lora _t0 = time.time() _log(f"loading turbo LoRA from {TURBO_LORA_REPO} ...") _lora_path = _fetch_lora_file(TURBO_LORA_REPO, TURBO_LORA_FILE) _lora_sd = _load_lora(_lora_path) # Group LoRA A/B pairs by target module name. _lora_pairs: dict[str, dict[str, torch.Tensor]] = {} for _k, _v in _lora_sd.items(): if ".lora_A." in _k: _base = _k.split(".lora_A.")[0] _lora_pairs.setdefault(_base, {})["A"] = _v elif ".lora_B." in _k: _base = _k.split(".lora_B.")[0] _lora_pairs.setdefault(_base, {})["B"] = _v # Store LoRA A/B factors on CPU (~1.2 GB total). The full delta (B @ A) would be # ~20 GB and overflow GPU memory when moved to CUDA inside @spaces.GPU (~14 GB free). # Instead _apply_turbo computes the delta on-the-fly: one temporary ~57 MB tensor # per module, freed each iteration. global TURBO_AVAILABLE, _TURBO_LoRA _TURBO_LoRA = [] for _base_name, _pair in _lora_pairs.items(): if "A" not in _pair or "B" not in _pair: continue _base_name = _base_name.removeprefix("transformer.") _parts = _base_name.split(".") _obj = pipe.transformer for _p in _parts: _obj = _obj[int(_p)] if _p.isdigit() else getattr(_obj, _p) _A = _pair["A"].to(_obj.weight.dtype) # (rank, in_features) _B = _pair["B"].to(_obj.weight.dtype) # (out_features, rank) _TURBO_LoRA.append((_obj, _A, _B)) del _lora_sd, _lora_pairs import gc as _gc; _gc.collect() TURBO_AVAILABLE = True _log(f"turbo LoRA parsed ({len(_TURBO_LoRA)} modules in {time.time() - _t0:.0f}s, merge pending)") if _TURBO_LoRA: _m, _A, _B = _TURBO_LoRA[0] _log(f" sample: module={type(_m).__name__}, " f"A.shape={_A.shape}, B.shape={_B.shape}, " f"A.device={_A.device}, dtype={_A.dtype}") _log_mem("startup turbo lora parsed") # Still startup, still free: an AoTI package carries no weights and opens its archive lazily inside the GPU # worker. Off unless `H3_AOTI=1`. import h3_aoti h3_aoti.maybe_load(pipe.transformer) if PLACEMENT == "pack": # Scoped to the transformer. `spaces` packs every startup-resident CUDA tensor into a second on-disk copy, # and packing all 77.3 GB busts the 150 GB storage quota; the 61.7 GB transformer alone fits. The ~10 GB of # fp32 VAEs move on the first GPU call instead. pipe.transformer.to("cuda") import gc as _gc _gc.collect() torch.cuda.empty_cache() _log_mem("startup transformer packed") if PLACEMENT == "offload": manager.enable_auto_cpu_offload(device="cuda") _arm_decode_hooks(pipe) PIPE, MANAGER = pipe, manager LOADED_IN = time.time() - started _log(f"ready in {LOADED_IN:.0f}s") _log_mem("startup ready") _dump_text = [f"placement={PLACEMENT}, attention={ATTENTION}, model={MODEL_REPO}"] try: import torch as _torch if _torch.cuda.is_available(): _dump_text.append(_torch.cuda.memory_summary()) except Exception: pass _debug_dump("startup", "\n".join(_dump_text)) if _lora_cache_path("probe", "probe.safetensors"): _log(f"LoRA bucket cache ready: {_LORA_CACHE_DIR}") else: _log(f"LoRA bucket cache unavailable ({_LORA_CACHE_DIR} not writable); using HF cache only", "WARN") except Exception as error: traceback.print_exc() LOAD_ERROR = f"**Loading `{MODEL_REPO}` failed** after {time.time() - started:.0f}s: `{type(error).__name__}: {error}`" return LOAD_ERROR def _arm_decode_hooks(pipe): """Make the offload hooks fire for the two VAEs. `enable_auto_cpu_offload` wraps `forward`, and the decode blocks call `vae.decode(...)` directly, so the hook never runs and the VAE is still on the host when the latents arrive on the card. """ for name in ("vae", "audio_vae"): module = getattr(pipe, name) inner = module.decode def armed(*args, _module=module, _decode=inner, **kwargs): hook = getattr(_module, "_hf_hook", None) if hook is not None: hook.pre_forward(_module) return _decode(*args, **kwargs) module.decode = armed @cache def conditioner(): """The other half, over the gradio API. `gradio_client` attaches the caller's own ZeroGPU token per call, so the conditioner's booking is billed to whoever asked for the video.""" from gradio_client import Client return Client(CONDITIONER_SPACE) def encode_remote(prompt, image_path, last_image_path, canvas, num_frames, rewrite_prompt=False): """`/encode` on the conditioner Space: a safetensors file holding `prompt_embeds` + `text_token_tags`, with the resolved `height` / `width` / `num_frames` in its metadata, plus the plan. `canvas` is the label. The cached `Client` is dropped and rebuilt on a failed call (the conditioner may have restarted or the connection dropped); one retry before giving up.""" from gradio_client import handle_file from safetensors import safe_open _t0 = time.time() _log(f"encode_remote: conditioning on {CONDITIONER_SPACE} (canvas={canvas}, {num_frames}f, " f"rewrite={rewrite_prompt}, image={bool(image_path)}, last_image={bool(last_image_path)}) ...") for _attempt in (1, 2): try: path, plan = conditioner().predict( prompt=prompt, image_path=handle_file(image_path) if image_path else None, last_image_path=handle_file(last_image_path) if last_image_path else None, canvas=canvas, num_frames=num_frames, rewrite_prompt=bool(rewrite_prompt), api_name="/encode", ) break except Exception: conditioner.cache_clear() if _attempt == 2: raise _log("encode_remote: conditioner call failed; reconnecting ...", "WARN") time.sleep(2) with safe_open(path, framework="pt") as handle: metadata = handle.metadata() _log(f"encode_remote: done in {time.time() - _t0:.0f}s " f"({plan.get('num_text_tokens')} tokens, {metadata.get('height')}x{metadata.get('width')}, " f"{metadata.get('num_frames')}f)") _log_mem("encode_remote done") _embeds = handle.get_tensor("prompt_embeds") _tags = handle.get_tensor("text_token_tags") try: os.remove(path) except OSError: pass return _embeds, _tags, metadata, plan # Seconds of GPU one request needs, from the packed video rows it is about to denoise: linear in the rows for the # matmuls, quadratic for the attention, against the AoTI block package this Space runs. _DUR_B, _DUR_C = 1.1745e-4, 3.8396e-9 # The two resident decoders and the mux, which scale with the output rather than with the step count. _DECODE_BASE, _DECODE_PER_DEFAULT_CANVAS, _DEFAULT_CANVAS_PIXELS = 15, 15, 960 * 544 * 124 # `pack` mode: only the ~10 GB of VAEs move on a cold worker. _PLACEMENT_ALLOWANCE, _PAD = 12, 10 def get_duration(prompt_embeds, text_token_tags, image, last_image, height, width, num_frames, steps, seed, *a, **k): height, width, num_frames, steps = int(height), int(width), int(num_frames), int(steps) latent_frames = (num_frames - LATENTS_PER_CHUNK) // FRAMES_PER_CHUNK * LATENTS_PER_CHUNK + 2 patches = (height // 32) * (width // 32) rows = latent_frames * patches + (int(image is not None) + int(last_image is not None)) * patches denoise = steps * (_DUR_B * rows + _DUR_C * rows**2) decode = _DECODE_BASE + _DECODE_PER_DEFAULT_CANVAS * (height * width * num_frames) / _DEFAULT_CANVAS_PIXELS _dur = max(60, int(denoise + decode) + _PLACEMENT_ALLOWANCE + _PAD) _log(f"get_duration: {width}x{height}, {num_frames}f, {steps} steps, rows={rows} -> {_dur}s " f"(denoise {denoise:.0f}s + decode {decode:.0f}s + {_PLACEMENT_ALLOWANCE + _PAD}s allowance)") return _dur @spaces.GPU(duration=get_duration, size=GPU_SIZE) def _generate(prompt_embeds, text_token_tags, image, last_image, height, width, num_frames, steps, seed, turbo=False, nsfw_keys=(), custom_loras=()): """The only thing on GPU time: the packed-sequence denoise loop and the two decoders. Only the three generated outputs come back β€” a `@spaces.GPU` return crosses a process boundary by pickling, and the full `PipelineState` still holds the packed latents, the rotary grid and the row indices on the card. """ import torch _t0 = time.time() _log(f"_generate ENTER: turbo={turbo}, TURBO_AVAILABLE={TURBO_AVAILABLE}, " f"_TURBO_merged={_TURBO_merged}, steps={steps}, {width}x{height}, {num_frames}f, " f"nsfw={tuple(nsfw_keys)}, custom={[(r, f) for r, f, _s in custom_loras]}") _log_mem("_generate ENTER") # Intra-phase memory sampler: a hard OOM kill leaves no traceback, so the last sampler # lines before the crash are the only evidence of where memory grew. Queries only. _stop_sampler = threading.Event() def _sampler(): while not _stop_sampler.wait(_MEM_SAMPLE_INTERVAL): _log_mem("sampler") if _MEM_SAMPLE_INTERVAL > 0: threading.Thread(target=_sampler, daemon=True).start() try: if PLACEMENT == "lazy": PIPE.to("cuda") elif PLACEMENT == "pack": PIPE.vae.to("cuda") PIPE.audio_vae.to("cuda") _log_mem("_generate VAEs on cuda") # Merge LoRA inside @spaces.GPU where the real weights are live on CUDA (in pack mode, # ZeroGPU unpacks the real transformer weights before calling this function). if turbo and TURBO_AVAILABLE: _log("_generate: calling _apply_turbo ...") _apply_turbo() _log("_generate: _apply_turbo done") for _key in nsfw_keys: _apply_nsfw(_key) for _repo, _file, _scale in custom_loras: _apply_custom((_repo, _file), _scale) _log_mem("_generate LoRAs merged") _log(f"_generate: starting PIPE (steps={steps}) ...") _log_mem("_generate before PIPE") state = PIPE( prompt_embeds=prompt_embeds.to("cuda"), text_token_tags=text_token_tags, image=image, last_image=last_image, height=height, width=width, num_frames=num_frames, num_inference_steps=int(steps), generator=torch.Generator("cpu").manual_seed(int(seed) if int(seed) >= 0 else torch.randint(0, 2**32, (1,)).item()), ) _log(f"_generate: PIPE done in {time.time() - _t0:.0f}s") _log_mem("_generate after PIPE") # Pull the outputs out of the state before the finally block drops it: the state still holds # the packed latents, rotary grid and row indices on the card. _videos = state.get("videos")[0] _audio = state.get("audio")[0].cpu() _sampling_rate = state.get("sampling_rate") except torch.cuda.OutOfMemoryError as _e: _oom_dump("generate", _e, f"request: {width}x{height}, {num_frames}f, steps={steps}, " f"turbo={turbo}, nsfw={tuple(nsfw_keys)}, " f"custom={[(r, f) for r, f, _s in custom_loras]}") raise finally: _stop_sampler.set() # Each unmerge is attempted independently so one failure can't leave the others merged. for _repo, _file, _scale in custom_loras: try: _remove_custom((_repo, _file), _scale) except Exception as _e: _log(f"custom LoRA unmerge failed for {_repo}/{_file}: {_e}", "ERROR") for _key in nsfw_keys: try: _remove_nsfw(_key) except Exception as _e: _log(f"NSFW LoRA unmerge failed for {_key}: {_e}", "ERROR") if turbo and TURBO_AVAILABLE: try: _remove_turbo() except Exception as _e: _log(f"turbo LoRA unmerge failed: {_e}", "ERROR") # Aggressive worker-side release: drop the packed state and hand every cached CUDA block back # so the worker's GPU + RSS footprint is flat between requests. try: del state except Exception: pass import gc as _gc _gc.collect() torch.cuda.empty_cache() _log_mem("_generate EXIT") _log(f"_generate EXIT: {(time.time() - _t0):.0f}s total (denoise + decode)") return _videos, _audio, _sampling_rate def _generate_impl(prompt, image_path=None, last_image_path=None, canvas=DEFAULT_CANVAS, duration=5, steps=25, seed=-1, upsample=False, turbo=False, nsfw_lora_keys=(), custom_lora_url="", custom_lora_scale=1.0, custom_lora_trigger="", progress=gr.Progress(track_tqdm=True)): """The request pipeline: LoRA prep β†’ remote conditioning β†’ @spaces.GPU denoise β†’ mux.""" if LOAD_ERROR: raise gr.Error(LOAD_ERROR) if PIPE is None: raise gr.Error("The denoiser is still loading.") if not prompt or not prompt.strip(): raise gr.Error("MiniMax-H3 always takes a prompt, keyframes or not.") from PIL import Image, ImageOps from diffusers.utils import encode_video num_frames = snap_frames(duration) if turbo and TURBO_AVAILABLE: steps = 7 # 7 sigma points β†’ 6 transformer evaluations (sweet spot for larryvrh v4-step600-EMA) # LoRAs the user selected via the catalog dropdown; the AIO motion adapter is auto-loaded whenever any is on. nsfw_keys = tuple(k for k in (nsfw_lora_keys or ()) if k in NSFW_LORA_CONFIGS) # User-supplied LoRA from the custom URL box: parse the link and require it was downloaded & verified. custom = [] if custom_lora_url and str(custom_lora_url).strip(): _repo, _file = _parse_hf_link(custom_lora_url) if _file is None: _file = _resolve_repo_file(_repo) if (_repo, _file) not in _CUSTOM_LoRA: raise gr.Error(f"`{_repo}/{_file}` isn't downloaded yet β€” press 'Download & Verify' first.") try: _scale = float(custom_lora_scale) if custom_lora_scale not in (None, "") else 1.0 except (TypeError, ValueError): raise gr.Error("Custom LoRA scale must be a number.") if not 0.01 <= _scale <= 4.0: raise gr.Error("Custom LoRA scale must be between 0.01 and 4.") _trigger = str(custom_lora_trigger or "").strip() custom = [(_repo, _file, _scale, _trigger)] if nsfw_keys or custom: nsfw_keys += ("nsfw_aio",) for _key in nsfw_keys: _ensure_nsfw_lora_loaded(_key) for _key in nsfw_keys: # The trigger token must be in the prompt text for the conditioner to encode the LoRA's concept. prompt = f"{prompt}, {NSFW_LORA_CONFIGS[_key]['trigger']}" for _repo, _file, _scale, _trigger in custom: if _trigger: prompt = f"{prompt}, {_trigger}" progress(0.0, desc=f"Upsampling the prompt on {CONDITIONER_SPACE} ..." if upsample else f"Conditioning on {CONDITIONER_SPACE} ...") conditioned = time.time() prompt_embeds, text_token_tags, metadata, plan = encode_remote( prompt, image_path, last_image_path, canvas, num_frames, rewrite_prompt=upsample ) condition_seconds = time.time() - conditioned _log_mem("after conditioning") height, width, num_frames = (int(metadata[key]) for key in ("height", "width", "num_frames")) refined = plan.get("refined_prompt") or "" def keyframe(path): # The conditioning latents encoded here have to be of the image the conditioner looked at, which it prepares # exactly this way. return ImageOps.exif_transpose(Image.open(path)).convert("RGB") if path else None progress(0.1, desc=f"Denoising {steps} steps at {width}x{height}, {num_frames} frames ...") started = time.time() frames, audio, sampling_rate = _generate( prompt_embeds, text_token_tags, keyframe(image_path), keyframe(last_image_path), height, width, num_frames, steps, seed, turbo, nsfw_keys=nsfw_keys, custom_loras=[(r, f, s) for r, f, s, _t in custom], ) generate_seconds = time.time() - started directory = os.path.join(tempfile.gettempdir(), "h3-outputs") os.makedirs(directory, exist_ok=True) # Rotate: keep only the newest _MAX_OUTPUTS MP4s so /tmp never fills up. _MAX_OUTPUTS = 25 for _f in sorted(glob.glob(os.path.join(directory, "*.mp4")), key=os.path.getmtime)[:-_MAX_OUTPUTS]: try: os.remove(_f) except OSError: pass path = os.path.join(directory, f"h3-{int(time.time() * 1000)}.mp4") encode_video(frames, fps=FPS, output_path=path, audio=audio, audio_sample_rate=sampling_rate) _log_mem("after mux") # The frames list and audio tensor are the last big main-process objects; drop them now. del frames, audio import gc as _gc _gc.collect() lora_note = " Β· LoRA: " + ", ".join(NSFW_LORA_CONFIGS[k]["trigger"] for k in nsfw_keys) if nsfw_keys else "" lora_label_text = "**LoRAs applied:** " + ", ".join(NSFW_LORA_CONFIGS[k]["label"] for k in nsfw_keys) if nsfw_keys else "" if custom: _custom_note = ", ".join(f"{r}/{f} (Γ—{s:g})" for r, f, s, _t in custom) lora_note = (lora_note + " Β· custom: " + _custom_note) if lora_note else " Β· custom: " + _custom_note lora_label_text = (lora_label_text + " Β· " + _custom_note) if lora_label_text else "**LoRAs applied:** " + _custom_note report = ( f"`{width}x{height}`, {num_frames} frames ({num_frames / FPS:.3f} s), {int(steps)} steps Β· " f"conditioner {condition_seconds:.0f}s ({plan['num_text_tokens']} tokens" f"{', upsampled' if refined else ''}) Β· " f"denoise + decode {generate_seconds:.0f}s ({generate_seconds / int(steps):.1f} s/step) Β· seed {int(seed)}" f"{lora_note}" ) _log(report) return path, report, refined, gr.update(visible=bool(refined)), gr.DownloadButton(value=path, visible=True), gr.update(value=lora_label_text, visible=bool(nsfw_keys) or bool(custom)) def generate(prompt, image_path=None, last_image_path=None, canvas=DEFAULT_CANVAS, duration=5, steps=25, seed=-1, upsample=False, turbo=False, nsfw_lora_keys=(), custom_lora_url="", custom_lora_scale=1.0, custom_lora_trigger="", progress=gr.Progress(track_tqdm=True)): """One request. `upsample` is last among the pre-existing positional args, so a positional API client that predates it is unaffected; the custom-LoRA args come after it. Logs the full request to the Space logs, then runs the pipeline. Any failure is logged with a full traceback before it reaches the UI. """ _log(f"generate ENTER: prompt={prompt!r:.120}, canvas={canvas}, duration={duration}s -> {snap_frames(duration)}f, " f"steps={steps}, seed={seed}, upsample={upsample}, turbo={turbo}, " f"nsfw={tuple(nsfw_lora_keys or ())}, custom_url={custom_lora_url!r}, " f"custom_scale={custom_lora_scale}, custom_trigger={custom_lora_trigger!r}") _t_req = time.time() try: return _generate_impl( prompt, image_path, last_image_path, canvas, duration, steps, seed, upsample, turbo, nsfw_lora_keys, custom_lora_url, custom_lora_scale, custom_lora_trigger, progress, ) except gr.Error as _e: _log(f"generate rejected: {_e}", "WARN") raise except Exception as _e: _log(f"generate FAILED after {time.time() - _t_req:.0f}s:\n{traceback.format_exc()}", "ERROR") try: import torch as _torch if isinstance(_e, _torch.cuda.OutOfMemoryError): _oom_dump("request", _e) except Exception: pass raise def export_last_frame(video_path): """Extract the last frame from a generated video and return it as an image file.""" if not video_path: raise gr.Error("No video to extract from.") import subprocess, tempfile out = os.path.join(tempfile.gettempdir(), f"h3-lastframe-{int(time.time() * 1000)}.png") result = subprocess.run( ["ffmpeg", "-sseof", "-0.1", "-i", video_path, "-frames:v", "1", "-y", out], capture_output=True, text=True, ) if result.returncode != 0: raise gr.Error(f"ffmpeg failed: {result.stderr[-300:]}") return out def _fit_keyframe(image_path, current_canvas): """Cover-crop an uploaded keyframe to the closest supported aspect ratio and select that ratio's smallest (fastest) canvas, unless the user already picked a matching ratio.""" if not image_path: return gr.update(), gr.update() from PIL import Image as _Image img = _Image.open(image_path) aspect = img.width / img.height fastest = {} for label, (h, w) in CANVASES.items(): r = w / h if r not in fastest or w * h < fastest[r][1][0] * fastest[r][1][1]: fastest[r] = (label, (h, w)) ratio = min(fastest, key=lambda r: abs(r - aspect)) label, (h, w) = fastest[ratio] cur_h, cur_w = CANVASES[current_canvas] if abs(cur_w / cur_h - aspect) <= abs(ratio - aspect): label = current_canvas h, w = cur_h, cur_w target = w / h if abs(img.width / img.height - target) <= 1e-3: return gr.update(value=image_path), gr.update(value=label) if img.width / img.height > target: new_w = int(img.height * target) left = (img.width - new_w) // 2 img = img.crop((left, 0, left + new_w, img.height)) else: new_h = int(img.width / target) top = (img.height - new_h) // 2 img = img.crop((0, top, img.width, top + new_h)) img.save(image_path) return gr.update(value=image_path), gr.update(value=label) load_models() INTRO = """# minimax h3 turbo + lora's <div align="center"> <a href="https://huggingface.co/MiniMaxAI/MiniMax-H3" target="_blank" rel="noopener"><strong>[ model ]</strong></a> &nbsp; <a href="https://www.minimax.io/blog/minimax-h3" target="_blank" rel="noopener"><strong>[ blog ]</strong></a> &nbsp; <a href="https://huggingface.co/spaces/multimodalart/minimax-h3-reference" target="_blank" rel="noopener"><strong>[ reference to video ]</strong></a> </div> **MiniMax-H3** is a 33B parameter state of the art video generation model that produces video and a fully synchronized soundtrack (ambience, foley, speech). **Turbo mode** (optional) applies the [larryvrh distillation LoRA](https://huggingface.co/InstantX/MiniMax-H3-Turbo-Lora-Diffusers) at generation time, reducing inference from 25 steps to 7 for ~4Γ— faster generation with minimal quality loss. """ CSS = """ .main.fillable { max-width: 1250px !important; } /* ── Root gradient background ── */ .dark .gradio-container { background: linear-gradient(135deg, #0a1628 0%, #0d2137 25%, #0f2b3a 50%, #0e3330 75%, #0a2a1f 100%) !important; color: #e8f0f8 !important; min-height: 100vh; /* ── Override Gradio's CSS variables for dark theme ── */ --background-fill-primary: rgba(10, 22, 40, 0.6) !important; --background-fill-secondary: rgba(10, 22, 40, 0.4) !important; --block-background-fill: rgba(10, 22, 40, 0.5) !important; --block-border-color: rgba(79, 195, 247, 0.15) !important; --block-label-text-color: #a8c4d8 !important; --block-title-text-color: #a8c4d8 !important; --background-text-color: #e0edf5 !important; --body-text-color: #e0edf5 !important; --body-text-color-subdued: #8aa8bc !important; --border-color-primary: rgba(79, 195, 247, 0.15) !important; --color-accent: #26a69a !important; --color-accent-soft: rgba(38, 166, 154, 0.15) !important; --button-primary-background-fill: linear-gradient(135deg, #0288d1 0%, #00897b 50%, #2e7d32 100%) !important; --button-primary-text-color: #ffffff !important; --button-secondary-background-fill: rgba(255, 255, 255, 0.06) !important; --button-secondary-border-color: rgba(255, 255, 255, 0.1) !important; --button-secondary-text-color: #b0c8dc !important; --input-background-fill: rgba(10, 22, 40, 0.6) !important; --input-background-fill-focus: rgba(10, 22, 40, 0.7) !important; --input-border-color: rgba(79, 195, 247, 0.15) !important; --input-border-color-focus: rgba(79, 195, 247, 0.45) !important; --slider-color: #26a69a !important; --accordion-border-color: rgba(255, 255, 255, 0.06) !important; } /* ── Intro markdown ── */ .gradio-container .prose h1, .gradio-container .prose h2 { background: linear-gradient(135deg, #4fc3f7, #26a69a) !important; -webkit-background-clip: text !important; -webkit-text-fill-color: transparent !important; background-clip: text !important; } .gradio-container .prose { color: #c8dde8 !important; } .gradio-container .prose a { color: #4dd0e1 !important; } /* ── Glass card panels (no backdrop-filter on .block β€” it traps fixed-position dropdown popups) ── */ .gradio-container .block { background: rgba(255, 255, 255, 0.04) !important; border: 1px solid rgba(255, 255, 255, 0.08) !important; border-radius: 14px !important; box-shadow: 0 4px 30px rgba(0, 0, 0, 0.25), inset 0 1px 0 rgba(255,255,255,0.05) !important; } /* ── Text / Number inputs (Gradio 6: label.container > .input-container > textarea/input) ── */ .gradio-container label.container textarea, .gradio-container label.container input[type="text"], .gradio-container input[type="number"] { background: rgba(10, 22, 40, 0.6) !important; border: 1px solid rgba(79, 195, 247, 0.15) !important; color: #e0edf5 !important; border-radius: 10px !important; } .gradio-container label.container textarea:focus, .gradio-container label.container input[type="text"]:focus, .gradio-container input[type="number"]:focus { border-color: rgba(79, 195, 247, 0.45) !important; box-shadow: 0 0 0 2px rgba(79, 195, 247, 0.12) !important; } /* ── Labels (Gradio 6: BlockTitle renders <span data-testid="block-info">) ── */ .gradio-container [data-testid="block-info"], .gradio-container label { color: #a8c4d8 !important; font-weight: 500 !important; } /* ── Slider track ── */ .gradio-container input[type="range"] { accent-color: #26a69a !important; } /* ── Checkbox (Gradio 6: <label class="checkbox-container"> β†’ <input> + <span class="label-text">) ── */ .gradio-container .checkbox-container { color: #c0d4e4 !important; } .gradio-container .checkbox-container .label-text { color: #c0d4e4 !important; } /* ── Dropdown (Gradio 6: div.container > div.wrap > div.wrap-inner > input[role="combobox"]) ── */ .gradio-container .wrap { background: rgba(10, 22, 40, 0.6) !important; border: 1px solid rgba(79, 195, 247, 0.15) !important; border-radius: 10px !important; } .gradio-container .wrap:focus-within { border-color: rgba(79, 195, 247, 0.45) !important; box-shadow: 0 0 0 2px rgba(79, 195, 247, 0.12) !important; } .gradio-container .wrap-inner input[role="combobox"] { color: #e0edf5 !important; background: transparent !important; } /* Dropdown popup options (Gradio 6: <ul class="options"> with position:fixed) */ .gradio-container ul.options { background: rgba(10, 22, 40, 0.95) !important; border: 1px solid rgba(79, 195, 247, 0.25) !important; border-radius: 10px !important; box-shadow: 0 8px 32px rgba(0, 0, 0, 0.5) !important; z-index: 1000 !important; } .gradio-container ul.options li.item { color: #e0edf5 !important; } .gradio-container ul.options li.item:hover, .gradio-container ul.options li.item.active { background: rgba(79, 195, 247, 0.15) !important; } /* ── Image upload area ── */ .gradio-container .image-container, .gradio-container .upload-container { background: rgba(10, 22, 40, 0.35) !important; border: 2px dashed rgba(79, 195, 247, 0.2) !important; border-radius: 12px !important; } .gradio-container .image-container:hover, .gradio-container .upload-container:hover { border-color: rgba(38, 166, 154, 0.5) !important; } /* ── Video output ── */ .gradio-container video { border-radius: 12px !important; border: 1px solid rgba(255,255,255,0.06) !important; } /* ── Generate button β€” hero gradient ── */ .gradio-container .lg.primary, .gradio-container button.primary, .gradio-container button.lg.primary { background: linear-gradient(135deg, #0288d1 0%, #00897b 50%, #2e7d32 100%) !important; color: #ffffff !important; border: none !important; border-radius: 12px !important; font-weight: 700 !important; letter-spacing: 0.03em !important; text-transform: uppercase !important; box-shadow: 0 4px 20px rgba(0, 137, 123, 0.35), 0 0 40px rgba(2, 136, 209, 0.15) !important; transition: all 0.3s cubic-bezier(0.4, 0, 0.2, 1) !important; position: relative; overflow: hidden; } .gradio-container .lg.primary::before { content: ""; position: absolute; inset: 0; background: linear-gradient(135deg, rgba(255,255,255,0.15) 0%, transparent 50%, rgba(255,255,255,0.05) 100%); pointer-events: none; } .gradio-container .lg.primary:hover, .gradio-container button.primary:hover { box-shadow: 0 6px 28px rgba(0, 137, 123, 0.5), 0 0 60px rgba(2, 136, 209, 0.25) !important; transform: translateY(-1px) !important; } /* ── Secondary buttons ── */ .gradio-container button.secondary, .gradio-container button.lg.secondary { background: rgba(255, 255, 255, 0.06) !important; border: 1px solid rgba(255, 255, 255, 0.1) !important; color: #b0c8dc !important; } .gradio-container button.secondary:hover { background: rgba(255, 255, 255, 0.1) !important; border-color: rgba(79, 195, 247, 0.3) !important; } /* ── Accordion (Gradio 6: Index.svelte adds "gr-accordion" to elem_classes; header is <button class="label-wrap">) ── */ .gradio-container .gr-accordion { background: rgba(255, 255, 255, 0.03) !important; border: 1px solid rgba(255, 255, 255, 0.06) !important; border-radius: 12px !important; } .gradio-container .gr-accordion .label-wrap { color: #a8c4d8 !important; font-weight: 600 !important; } .gradio-container .gr-accordion .label-wrap:hover { color: #4fc3f7 !important; } .gradio-container .gr-accordion [data-testid="accordion-content"] { padding-top: var(--size-2, 8px) !important; } /* ── Examples ── */ .gradio-container .examples { background: rgba(10, 22, 40, 0.3) !important; border: 1px solid rgba(255, 255, 255, 0.06) !important; border-radius: 12px !important; } /* ── Markdown output ── */ .gradio-container .prose { color: #b8ccda !important; } .gradio-container .prose code { background: rgba(79, 195, 247, 0.1) !important; color: #4dd0e1 !important; border-radius: 4px !important; padding: 2px 6px !important; } /* ── Scrollbar ── */ .gradio-container ::-webkit-scrollbar { width: 6px !important; } .gradio-container ::-webkit-scrollbar-track { background: rgba(0,0,0,0.2) !important; } .gradio-container ::-webkit-scrollbar-thumb { background: rgba(79, 195, 247, 0.3) !important; border-radius: 3px !important; } """ def _lora_descriptions(keys): """Markdown of each selected LoRA's label + description, plus conflict warnings; hidden when nothing is selected.""" keys = tuple(k for k in (keys or ()) if k in NSFW_LORA_CONFIGS) if not keys: return gr.update(visible=False) blocks = [f"**{NSFW_LORA_CONFIGS[k]['label']}** β€” {NSFW_LORA_CONFIGS[k]['desc']}" for k in keys] conflicts = [] for group, members in LORA_CONFLICT_GROUPS.items(): hit = [k for k in keys if k in members] if len(hit) > 1: conflicts.append(f"**{group}**: " + " + ".join(NSFW_LORA_CONFIGS[k]['label'] for k in hit)) if conflicts: blocks.append("**⚠️ Conflicts:** " + " ".join(conflicts) + " β€” these overlap in purpose and may fight each other; consider picking one from each group.") return gr.update(value="\n\n".join(blocks), visible=True) with gr.Blocks(title="minimax h3 turbo + lora's") as demo: gr.Markdown(INTRO) with gr.Row(): with gr.Column(): prompt = gr.Textbox( label="Prompt", lines=3, value="A red fox trotting through a snowy pine forest at dawn, snow crunching underfoot", ) with gr.Row(): upsample = gr.Checkbox(label="Upsample prompt", value=False) turbo = gr.Checkbox(label="Turbo (7-step LoRA)", value=False, visible=TURBO_AVAILABLE, info="Apply larryvrh distillation LoRA, reduce steps from 25 to 7") with gr.Accordion("NSFW LoRAs", open=False): nsfw_lora_keys = gr.Dropdown( label="Select LoRAs to apply", choices=[(cfg["label"], key) for key, cfg in NSFW_LORA_CONFIGS.items() if key != "nsfw_aio"], value=[], multiselect=True, filterable=True, info="Searchable β€” type to filter. Selected LoRAs show as removable tags; their trigger tokens are auto-appended to the prompt. The HMNSFW AIO motion adapter auto-loads whenever any LoRA is selected.", ) lora_desc = gr.Markdown(visible=False) with gr.Accordion("Custom LoRA (Hugging Face)", open=False): gr.Markdown("Paste a Hugging Face link β€” a repo page or a direct `.safetensors` file link β€” then press " "**Download & Verify**. The verified LoRA is applied on Generate. " "ComfyUI and kohya formats are auto-detected.") custom_lora_url = gr.Textbox( label="HF repo or .safetensors link", placeholder="https://huggingface.co/Owner/Repo or https://huggingface.co/Owner/Repo/resolve/main/file.safetensors", ) with gr.Row(): custom_lora_scale = gr.Number(label="Scale", value=1.0, minimum=0.01, maximum=4.0, step=0.1) custom_lora_trigger = gr.Textbox(label="Trigger (optional)", placeholder="Auto-appended to the prompt if given") custom_lora_btn = gr.Button("Download & Verify", size="sm") custom_lora_status = gr.Markdown(visible=False) with gr.Row(): image = gr.Image(label="First frame (optional)", type="filepath") last_image = gr.Image(label="Last frame (optional)", type="filepath") run = gr.Button("Generate", variant="primary") with gr.Accordion("Advanced options", open=False): canvas = gr.Dropdown(label="Canvas", choices=list(CANVASES), value=DEFAULT_CANVAS) duration = gr.Slider(label="Duration (s)", minimum=MIN_UI_DURATION, maximum=MAX_UI_DURATION, step=1, value=5) steps = gr.Slider(label="Steps", minimum=7, maximum=40, step=1, value=25) seed = gr.Number(label="Seed (βˆ’1 = random)", value=-1, precision=0) with gr.Column(): video = gr.Video(label="Video + soundtrack") with gr.Row(): last_frame_btn = gr.Button("Export Last Frame", size="sm") download_btn = gr.DownloadButton("Download Video", size="sm", visible=False) last_frame_output = gr.Image(label="Last frame") lora_label = gr.Markdown(visible=False) report = gr.Markdown(visible=False) # An output, so it can be revealed only for a request that asked for a rewrite. with gr.Accordion("Upsampled prompt", open=False, visible=False) as upsampled_panel: upsampled = gr.Textbox(show_label=False, lines=8, interactive=False) image.upload(_fit_keyframe, [image, canvas], [image, canvas]) run.click( generate, [prompt, image, last_image, canvas, duration, steps, seed, upsample, turbo, nsfw_lora_keys, custom_lora_url, custom_lora_scale, custom_lora_trigger], [video, report, upsampled, upsampled_panel, download_btn, lora_label], api_name="generate", ) custom_lora_btn.click(_download_custom_lora, [custom_lora_url, custom_lora_scale, custom_lora_trigger], [custom_lora_status]) nsfw_lora_keys.change(_lora_descriptions, [nsfw_lora_keys], [lora_desc]) last_frame_btn.click( export_last_frame, [video], [last_frame_output], ) download_btn.click( lambda: gr.Info("Downloading video..."), [], [], ) if __name__ == "__main__": demo.launch(show_error=True, theme=gr.themes.Base(), css=CSS)
"""MiniMax-H3 `t2va` / `fl2va`, split deployment β€” the denoising half."""
from __future__ import annotations
import glob
import json
import os
import re
import shutil
import struct
import tempfile
import threading
import time
import traceback
from functools import cache
# Before anything that could initialize CUDA: `import spaces` patches `torch.cuda` so the 72 GiB load can happen at
# startup rather than on GPU time.
import spaces
import gradio as gr
MODEL_REPO = os.environ.get("H3_MODEL_REPO", "MiniMaxAI/MiniMax-H3")
CONDITIONER_SPACE = os.environ.get("H3_CONDITIONER", "multimodalart/qwen3vl-conditioner")
# `pack` places the transformer at startup, `lazy` moves everything on the first GPU call, `offload` hands placement to
# `ComponentsManager.enable_auto_cpu_offload`.
PLACEMENT = os.environ.get("H3_PLACEMENT", "pack").lower()
# cuDNN's fused attention is 10-20% faster than the SDPA default on this pool and needs nothing installed.
ATTENTION = os.environ.get("H3_ATTENTION", "_native_cudnn").lower()
GPU_SIZE = os.environ.get("H3_GPU_SIZE", "xlarge")
# larryvrh distillation LoRA (diffusers conversion): A/B factors stored on CPU, merged on-demand inside @spaces.GPU.
TURBO_LORA_REPO = os.environ.get("H3_TURBO_LORA", "InstantX/MiniMax-H3-Turbo-Lora-Diffusers")
TURBO_LORA_FILE = "minimax_h3_turbo_4step_ckpt500_diffusers.safetensors"
TURBO_LORA_SCALE = 1.0 # alpha == rank
# NSFW / anatomy LoRAs (ComfyUI format, converted at load). Lazy: downloaded + converted on CPU in generate(),
# merged on GPU inside @spaces.GPU only when selected. Standard LoRA only β€” LoKR (.lokr_w1/.lokr_w2) is unsupported.
NSFW_LORA_CONFIGS = {
"nsfw_innie": {
"repo": "Hearmeman/minimax-h3-loras",
"file": "HMInnie_v1_e50.safetensors",
"scale": 1.0,
"label": "NSFW Innie Pussy (HMInnie V1)",
"trigger": "inniepussy",
"desc": "Renders the vulva as a neat 'innie' β€” labia tucked in with no protruding inner lips. Use for close-up genital shots where you want a clean, compact vulva look. Not for scenes needing prominent outer labia or detailed pussy anatomy β€” pair with HMPussy for that.",
},
"nsfw_penis": {
"repo": "observantdistressed/plora-h3", # CivitAI "Penis Lora - H3" v1.1, mirrored to HF
"file": "PLORA_H3_V2-step00006300.safetensors",
"scale": 0.7,
"label": "NSFW Penis (Penis Lora - H3 v1.1)",
"trigger": "PENISLORA",
"format": "kohya",
"desc": "Adds a detailed, erect penis to male subjects β€” the core anatomy LoRA for any scene with a visible penis. Use whenever a penis is in frame. Not for solo female scenes, and don't stack it with HMPenis v2.0 (same purpose, conflicting weights).",
},
"nsfw_dildo": {
"repo": "EllaPriest45/MinimaxH3_Actions", # CivitAI "Dildoing" model 2903074, backup repo
"file": "Dildoing - MinimaxH3 - pumping a dildo insider her vagina.safetensors",
"scale": 0.7,
"label": "NSFW Dildoing (Dildo Insertion)",
"trigger": "pumping a dildo insider her vagina",
"desc": "Makes dildo insertion read naturally β€” the toy slides into the vagina with proper depth and motion instead of clipping or floating. Use for any dildo penetration scene. Not for scenes without a dildo, and don't combine it with other insertion LoRAs at full strength.",
},
# Motion/action adapter β€” auto-loaded (not user-selectable) whenever any anatomy LoRA is on, so the
# anatomy renders sharply AND the act moves naturally. The Hearmeman README says to stack anatomy
# adapters underneath this action adapter.
"nsfw_aio": {
"repo": "Hearmeman/minimax-h3-loras",
"file": "HMNSFW-AIO-V2.5.safetensors",
"scale": 0.7,
"label": "HMNSFW AIO V2.5 (motion)",
"trigger": "hmmotion",
"desc": "Motion/action adapter that makes anatomy LoRAs move naturally during sex acts β€” auto-loaded whenever any NSFW LoRA is selected. Use stacked underneath anatomy LoRAs for sharp rendering plus natural motion. Not for standalone use; it's a motion enhancer, not an anatomy fix.",
},
# ── EllaPriest45/MinimaxH3_Actions catalog (t2va-compatible, standard LoRA only) ──
# 93 LoRAs, classified by header inspection: 80 ComfyUI + 13 kohya. LoKR, I2V, REF2V, R2V, FL2VA excluded.
# Triggers/scales extracted from the CivitAI filenames (best effort β€” refine per-LoRA as needed).
'nsfw_ahegao_face': {
"repo": "EllaPriest45/MinimaxH3_Actions",
"file": 'Ahegao Face - MinimaxH3 - She is making the ahegao face - eyes rolled back and tongue hanging out.safetensors',
"scale": 0.7,
"label": 'Ahegao Face',
"trigger": 'ahegao face',
"desc": "Makes the subject's face contort into the ahegao expression β€” eyes rolled back, tongue hanging out β€” for intense reaction shots. Use during orgasm or overstimulation scenes. Not for normal expressions or dialogue scenes; it overrides the face.",
},
'nsfw_all_tied_up': {
"repo": "EllaPriest45/MinimaxH3_Actions",
"file": 'All Tied Up - MinimaxH3 - 0.8str.safetensors',
"scale": 0.8,
"label": 'All Tied Up',
"trigger": 'all tied up',
"desc": "Renders the subject bound and restrained, with rope or ties visible and functional. Use for bondage and restraint scenes. Not for scenes where the subject needs to move freely; it locks them into restraint poses.",
},
'nsfw_astro_nsfw': {
"repo": "EllaPriest45/MinimaxH3_Actions",
"file": 'Astro NSFW - MinimaxH3 - improves human movement and anatomy.safetensors',
"scale": 0.7,
"label": 'Astro NSFW',
"trigger": 'astro nsfw',
"format": "kohya",