{ "1": { "class_type": "UNETLoader", "inputs": { "unet_name": "minimax_h3_ref2va_pruned_int8_convrot.safetensors", "weight_dtype": "default" } }, "m0": { "class_type": "MiniMaxH3TurboLoRA", "inputs": { "lora_name": "minimax_h3_fl2v_lightx2v_turbo_8step_v1.0_comfy.safetensors", "strength": 1.0, "low_vram": false, "model": [ "1", 0 ] }, "_comment": "lightx2v Minimax-h3 Turbo **8step v1.0** 단독 @1.0 (2026-08-12 `t_4e40bb6cf50d` · CEO 승격 지시). 구 `ckpt850`(4step) 에서 교체 — 격자 4판 실측에서 **8step·캐시ON 이 최적**이었다: 6step·캐시OFF 대비 **80초 빠르고**(521~525s vs 601s) 오디오·화질이 같거나 낫다. ★6step·캐시ON 은 **오디오가 갈라진다**(CEO 육안 \"목소리 갈라지네\") — 속도만 보고 6step 으로 내리지 말 것." }, "L2": { "class_type": "LoraLoaderModelOnly", "inputs": { "lora_name": "MysticXXX_MMH3-V4.safetensors", "strength_model": 0.4, "model": [ "m0", 0 ] }, "_comment": "Mystic XXX `MysticXXX_MMH3-V4` **0.4** — 2026-09-06 `t_0ae98911f9cb` 신설 슬롯. 체인은 `m0(터보) → L2(mystic 0.4) → m1(sage)`. ★★2026-09-10 CEO 판정으로 **realism(구 L1) 을 이 판에서 걷었다**(`t_2ac5b03c5db2`) — 사유 = *\"mystic 과 realism 이 같이 적용되면 영상 품질이 망가지더라. 각각 쓰면 괜찮은데\"*. ⇒ 종전 이 자리에 있던 *\"realism 을 빼고 그 칸을 쓰지 않았다\"* 는 **그 판정 이전의 입장**이라 걷었다 (이 필드는 실행 `workflow.json` 에까지 실리므로, 낡은 입장을 현재형으로 남기면 읽는 사람이 **누락으로 읽고 되넣는다**). ★realism 자체가 기각된 게 아니라 **Mystic 과의 동시 적용**이 기각됐다. ★값 0.4 의 근거 = 형제 `ref2va-minimax-h3-mystic-5090-1` 의 저작값과 같게 둔다(같은 파일을 두 판이 다른 강도로 쓰면 「어느 쪽이 정본인가」가 생긴다). 제작자 권장은 0.5~0.9 라 우리 값은 그 아래이고, 강도 확정은 CEO 육안 별건이다. ★끄려면 `spec_override.toggles {\"lora_2\": false}` — 그러면 러너 splice 가 `m1.model` 을 `[\"m0\", 0]` 로 되돌려 **LoRA 0장(터보만)** 체인이 된다." }, "m1": { "class_type": "MiniMaxH3MemoryEfficientSageAttentionPatch", "inputs": { "model": [ "L2", 0 ] } }, "2": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors", "type": "minimax", "device": "default" } }, "3": { "class_type": "VAELoader", "inputs": { "vae_name": "minimax_h3_video_vae_fp16.safetensors" } }, "4": { "class_type": "VAELoader", "inputs": { "vae_name": "minimax_h3_audio_vae_fp32.safetensors" } }, "101": { "class_type": "LoadImage", "inputs": { "image": "dockaiq-input-453411-input_image-373b128b-4135-48ab-ba6b-d2284eb23fe3.webp" } }, "102": { "class_type": "LoadImage", "inputs": { "image": "dockaiq-input-453411-picture_2-ref-3shot.webp" } }, "103": { "class_type": "LoadImage", "inputs": { "image": "dockaiq-input-453411-picture_3-3880888d-5900-4a22-9749-3f92d848a988-outer.webp" } }, "6": { "class_type": "MiniMaxH3ReferenceToVideo", "inputs": { "prompt": " is the location; keep its layout and furnishings, surfaces clear of small objects.\n is — this exact person; keep the same face and hair. is 's clothing sheet.\n\n is the woman with dark brown long wavy hair. She is kneeling on the bed with her hips raised, her back and the back of her head toward the viewer, her arms folded under her as she grips the bedsheet, her cheek resting against the pillow, wearing the black sheer thigh-high stockings in , pulled down and bunched at her ankles with her toes brushing the bedsheet; her chest is bare and she is bare from the waist down.\n\nThe viewer, a man, is behind her, his hands on her waist at the lower edge of the frame; only his hands and forearms enter the frame, wearing the worn black hoodie and the worn leather bracelet with a faded brown strap, bare from the waist down. His face, head, torso and hips are never in frame — the camera is where his eyes are.\n\nThe camera is the viewer's own point of view, from his two eyes behind her: a medium shot from behind at hip height, her back and raised hips filling the frame, her body close against him and facing away, her back and the back of her head toward the camera. The camera moves only as much as the viewer's head actually moves — a slow rhythm with her, a slight tilt, a small lean in — and never pushes in on her face or changes to a close-up.\n\n[SHOT 1] The scene opens already in progress: is kneeling on the bed with her hips raised, the viewer penetrating her from behind, his hands on her waist at the lower edge of the frame. Her back arches and her hips rock slowly against him, her fingers gripping the bedsheet, her cheek pressed to the pillow. Daylight from the window falls across her smooth back, and the black stockings bunched at her ankles shift on the sheet. Her tone is breathless and startled, a little strained: [Korean] 읏… 이 자세는, 처음인데.\n\n[SHOT 2] At 00:02.600, the shot stays in the same position as her hips lift higher and her arms fold under her, her fingers clenching the sheet. Her shoulders tighten and a short breath escapes her, her damp nape hair sticking to her skin, the inner walls gripping and holding. Her tone is breathy and pleased, softening into a small moan: [Korean] 하아… 유준 씨, 거기… 아, 거기 좋아.\n\nSubtle early reflections from nearby walls. The dialogue is naturally embedded in the environment rather than sounding close-miked or studio-recorded. Soft environmental ambience continuously fills the space. Very subtle natural reverberation. No artificial narration quality.\n\nNo text overlays, no subtitles, no watermark, no extra people in the background, no camera shake, no jump cuts, no third-person angle, no over-the-shoulder view, no mirror, no face turning toward the camera.", "width": 384, "height": 672, "length": 124, "ref_image_size": "match", "clip": [ "2", 0 ], "vae": [ "3", 0 ], "audio_vae": [ "4", 0 ], "ref_images.ref_image_0": [ "101", 0 ], "ref_images.ref_image_1": [ "102", 0 ], "ref_images.ref_image_2": [ "103", 0 ] } }, "7": { "class_type": "BasicGuider", "inputs": { "model": [ "mS", 0 ], "conditioning": [ "6", 0 ] } }, "8": { "inputs": { "sampler_name": "er_sde" }, "class_type": "KSamplerSelect", "_meta": { "title": "KSampler (선택)" } }, "9": { "class_type": "BasicScheduler", "inputs": { "scheduler": "beta", "steps": 8, "denoise": 1.0, "model": [ "mS", 0 ] } }, "10": { "class_type": "RandomNoise", "inputs": { "noise_seed": 1952783644 } }, "11": { "class_type": "SamplerCustomAdvanced", "inputs": { "noise": [ "10", 0 ], "guider": [ "7", 0 ], "sampler": [ "8", 0 ], "sigmas": [ "9", 0 ], "latent_image": [ "6", 1 ] } }, "12": { "class_type": "VAEDecode", "inputs": { "samples": [ "11", 0 ], "vae": [ "vb", 0 ] } }, "13": { "class_type": "VAEDecodeAudio", "inputs": { "samples": [ "11", 0 ], "vae": [ "4", 0 ] } }, "14": { "class_type": "CreateVideo", "inputs": { "fps": 24, "bit_depth": 8, "images": [ "wm", 0 ], "audio": [ "13", 0 ] } }, "15": { "class_type": "SaveVideo", "inputs": { "filename_prefix": "h3_r2v_v384_j453411", "format": "auto", "video": [ "14", 0 ], "codec": "auto" } }, "vb": { "class_type": "H3VideoVAEBatchDecode", "inputs": { "vae": [ "3", 0 ], "enabled": true } }, "mL": { "inputs": { "head_chunks": 4, "model": [ "m1", 0 ] }, "class_type": "MiniMaxLowVRAMAttention", "_meta": { "title": "MiniMax H3 Low VRAM Attention" } }, "mC": { "inputs": { "chunks": 2, "seq_threshold": 4096, "model": [ "mL", 0 ] }, "class_type": "MiniMaxChunkFeedForward", "_meta": { "title": "MiniMax H3 Chunk FeedForward" } }, "mS": { "inputs": { "shift_video": 12.0, "shift_audio": 6.0, "model": [ "mC", 0 ] }, "class_type": "MiniMaxH3SigmaShift", "_meta": { "title": "ModelSamplingMiniMaxH3" } }, "wm": { "inputs": { "images": [ "12", 0 ], "line1": "AI", "line2": "yusti.net", "height_px": 12, "alpha": 0.85, "margin_px": 13, "shadow_px": 1, "line3": "33e70695" }, "class_type": "DockContentWatermark", "_meta": { "title": "법정 표시 워터마크 (t_88479c752c44)" } } }