← 返回提示词图库

PROMPT RECORD图像记录

【完整实际 H3 提示词|minimax H3 Ref2VA】 以下为已提交工作流节点401的完整原文;prompt-v1.txt与其规范化文本一致,全文只列一次。 subject_defin...

AI 图像 superfang119 Tue Oct 20 00:39:39 +0000 2026
一键生图 查看来源

中文说明

【完整实际 H3 提示词|minimax H3 Ref2VA】 以下为已提交工作流节点401的完整原文;prompt-v1.txt与其规范化文本一致,全文只列一次。 subject_definitions: 是方知微,一位虚构的成年女性;其面部结构、五官比例、年龄观感、自然基础肤色、身体比例以及默认的近中分深棕黑色长卷发,全部仅由 (用户固定的 v6 身份参考图)控制。她由用户设定的身高为 175 cm,并非从图像中测量得出。 仅提供浅象牙色/米色的宽船领上衣(带细窄深色横条)以及小吊坠项链。其身份、面部、身体与头发不得覆盖 。设定图上的深色长裤与鞋子属于推定补全,仅因源裁切画面中未显示,并将始终位于镜头取景之外。 是由 引导的泛化空夜巷环境;该图为源自源帧203的AI生成场景底板。它仅贡献柔焦街道纵深、湿感路...

原始 Prompt

【完整实际 H3 提示词|MiniMax H3 Ref2VA】
以下为已提交工作流节点401的完整原文;prompt-v1.txt与其规范化文本一致,全文只列一次。

subject_definitions:
 is Fang Zhiwei, the fictional adult woman whose facial structure, feature proportions, age appearance, natural base skin tone, body proportions, and default near-center-parted long dark brown-black waves are controlled exclusively by , the user's fixed v6 identity reference. Her user-defined height is 175 cm and is not measured from the image.  supplies only the light ivory/beige wide boat-neck top with narrow dark horizontal stripes and the small pendant necklace. Its identity, face, body and hair do not override . Dark trousers and shoes on the sheet are inferred completion only because the source crop does not show them and will stay outside the shot framing.
 is a generalized empty nighttime lane environment guided by , an AI-generated scene plate derived from source frame 203. It contributes only a soft-focused street depth, wet-looking road reflections, cyan-teal light bokeh and amber lighting. A distant left-rear pedestrian visible in the source is deliberately omitted in this adaptation; the AI plate does not authenticate a real location or exact architecture.
 is a background-only shot-planning reference for [Shot 1], defining a quiet night-street palette and eye-level perspective. It is not a first-frame lock. Do not reproduce text, signage, subtitles, a person, or graphic overlays from the source.

summary:
[reference generation] Create one 2.25-second target at 24 fps as a single continuous horizontal upper-body shot of Fang Zhiwei on a softly lit night street.  alone controls her identity, face, body and default hair;  controls only the visible striped wide-neckline top and small necklace;  guides only the generalized empty street environment and palette. Adapt the observed action of raising her right hand from chest level and bringing the index finger to the lips, then holding the pose. This is a static-reference adaptation, not video-conditioned generation or frame-exact motion transfer. The source audio has not been listened to; no source dialogue, lyrics, voice, ambience or music is inferred or reused.

retention_analysis:
 (throughout [Shot 1]): fully_preserved for the fictional identity - preserve the canonical face proportions, age appearance, base skin tone, body proportions and default long dark waves from . Never copy or imitate the source performer’s face, hairstyle or identity. Outfit is adapted separately from .
's clothing (throughout [Shot 1]): partially_preserved - use a modest ivory/beige wide boat-neck top with narrow dark horizontal stripes and a small simple pendant. Keep the top consistent; do not add a bag or unobserved accessories. The lower body is not visible in the source and remains out of frame; sheet trousers and footwear are inferred only.
 (throughout [Shot 1]): partially_preserved - reuse the broad nighttime street depth, cyan-teal lights at frame left and warm amber light toward frame right from . The empty scene and removal of the distant soft-focus pedestrian are adaptation choices. Avoid added signs, readable text, traffic, a second person or a specific identifiable location.
: reused as a static scene and composition guide only, not as a concrete frame or motion source.

detailed_description:
[Shot 1] 00:00.000–00:02.250 - One uninterrupted horizontal eye-level medium close-up, from Fang Zhiwei’s upper torso to a small margin above her head. Keep her head, both shoulders and the entire right hand in frame throughout; do not crop the hand during its rise. Use a stable lens and camera position with no zoom, pan, tilt, reframing, reverse angle or cut. Behind her, maintain a softly blurred nighttime lane with distant cyan-teal points of light toward frame left, restrained amber illumination toward frame right, and a dim reflective road in the depth.  is only an environmental guide. Keep background shapes generic and soft; omit the source’s indistinct left-rear pedestrian as an intentional adaptation. Do not add clear shop lettering, subtitles, title text, logos, signs, vehicles or another person.
, Fang Zhiwei, is the sole foreground person. Her visible face, facial proportions, age appearance, base skin tone, body identity and near-center-parted long dark waves come only from .  changes wardrobe only: she wears a relaxed ivory/beige wide boat-neck top with narrow dark horizontal stripes and a small simple pendant necklace. Keep hair loose over the shoulders, without copying the source performer’s exact hairstyle or face. The upper-body framing matches the visible source scope; do not invent clothing below the crop.
At 00:00.000 she faces generally forward with a mild upward/forward gaze, shoulders quiet and both hands below the visible chest line. From 00:00.250 to 00:00.850, she smoothly raises her own right forearm from chest level toward her face in one continuous, anatomically natural movement. By 00:00.950, her right index finger reaches just in front of her lips and pauses; her hand must not cover her eyes, nose or most of her face. Hold this final position steadily through 00:02.250, with only tiny natural finger and breathing movement. Keep the face readable above and beside the hand so identity can be reviewed. The gesture describes the visible hand-to-lips pose only: do not assign a “shush” intention, invent dialogue, lip-sync, singing, tears, laughter or a story explanation. No second gesture and no interaction occur. The ending remains on the held pose rather than cutting away. The action timings are target guidance only and require frame-by-frame post-generation verification; the final 54-frame delivery must show the hand reaching the lip and holding before the last frame.

overall_soundscape:
For the newly generated target, use low, generic nighttime street room tone with soft distant traffic wash and a faint cloth movement as the arm rises. Do not include intelligible dialogue, lyrics, a recognizable voice or a distinctive source sound. The source audio has not been listened to, so its sound content and subjective A/V sync are unknown; these words describe only a new H3 audio target.

non_diegetic_music:
Add a newly composed, restrained audience-only cue: a low sustained string-like synth bed with one or two sparse felt-piano notes at a slow pulse, held quietly without a dramatic swell or recognizable melody. This is a new target score, not a claim about or reuse of the source music.

【实际模型与采样参数】
下列JSON直接读取本次已提交graph,含原底模、双LoRA、SelfLift、Euler/simple、12步与CFG1。浮点数按实际存值保留。
{
  "192": {
    "class_type": "UNETLoader",
    "inputs": {
      "unet_name": "H3\\minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors",
      "weight_dtype": "default"
    }
  },
  "53": {
    "class_type": "LoraLoaderModelOnly",
    "inputs": {
      "lora_name": "H3\\minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors",
      "strength_model": 0.8000000000000002,
      "model": [
        "192",
        0
      ]
    }
  },
  "400": {
    "class_type": "LoraLoaderModelOnly",
    "inputs": {
      "lora_name": "H3\\MysticXXX_MMH3-V1.safetensors",
      "strength_model": 0.45000000000000007,
      "model": [
        "53",
        0
      ]
    }
  },
  "36": {
    "class_type": "VAELoader",
    "inputs": {
      "vae_name": "minimax_h3_video_vae_int8_convrot.safetensors"
    }
  },
  "37": {
    "class_type": "VAELoader",
    "inputs": {
      "vae_name": "minimax_h3_audio_vae_fp32.safetensors"
    }
  },
  "38": {
    "class_type": "CLIPLoader",
    "inputs": {
      "clip_name": "h3\\qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors",
      "type": "minimax",
      "device": "default"
    }
  },
  "255": {
    "class_type": "KSamplerSelect",
    "inputs": {
      "sampler_name": "euler"
    }
  },
  "261": {
    "class_type": "BasicScheduler",
    "inputs": {
      "scheduler": "simple",
      "steps": 12,
      "denoise": 1,
      "model": [
        "329",
        0
      ]
    }
  },
  "397": {
    "class_type": "SelfLiftH3Sampler",
    "inputs": {
      "seed": 62863194,
      "cfg": 1,
      "transition_step": 6,
      "lowres_scale": 0.6000000000000001,
      "rho": 0,
      "w_min": 0.5,
      "w_max": 1,
      "upscaler_model": "h3_upscaler_lms_v0.1.safetensors",
      "highres_tiling": true,
      "upscaler_unload": true,
      "model": [
        "329",
        0
      ],
      "positive": [
        "265",
        0
      ],
      "negative": [
        "328",
        0
      ],
      "vae": [
        "36",
        0
      ],
      "latent_image": [
        "265",
        1
      ],
      "sampler": [
        "255",
        0
      ],
      "sigmas": [
        "261",
        0
      ]
    }
  }
}

【源片、原生输出与交付】
源素材:https://t.co/2PV0MdbSGL
源完整原件SHA-256:d0f4dc8d698253a581195027b9a74f5027eee6b4b563f27f50aa2f33aea7ca1c
源帧[203,270),PTS[103936,138240),time_base 1/15360;67帧@30fps=2.233333秒。
H3请求2.25秒,实际原生56帧@24fps=2.333333秒,1664×928。原生SHA-256:87d971c81459816041c50a05123f92687365ed958ce3c95682316934c8f36da3
交付取原生视频帧[0,54),54帧@24fps=2.25秒,不变速、不插帧;比源长0.016667秒。H3生成音轨对应前2.25秒裁切后AAC重编码,未替换源AAC。
最终SHA-256:b96e14fac522468e9529f3a6593dadcb35ad45ee2985c8d8bd7d6d9529c5dd78
工作流SHA-256:7cabbd60515de53e3a541802e28a84824baeb54f4f5e9ccb54ea41e483860d39
prompt ID:a70c0ccc-4aa2-413e-83d9-06c0618e65e6。原生和交付完整音视频解码通过。

【实际输入图顺序】
Picture 1:references/fang-zhiwei-identity-v6.png
SHA-256:9c8b2fac6469251c034def09a08264050b213da321bf9edac48c51a796bef448
用途:sole authority for Fang Zhiwei face, age appearance, base skin tone, body proportions and default long dark wavy hair
ALT:方知微固定身份基准图,左侧清晰正脸、右侧正侧背全身视图;此图负责身份与身材。
Picture 2:characters/fang-zhiwei-striped-top-sheet-v1.png
SHA-256:e8908e894d0fa908bbf41634d4a2339dd25c08ebabeb1a57930f6ccc478ff96a
用途:wardrobe only: modest ivory/beige wide boat-neck top with narrow dark horizontal stripes and small pendant. Lower-body trousers/shoes are inferred and out of frame. Do not use this image for identity or body.
ALT:方知微米白底细深色横条宽领上衣及小吊坠的角色设定板;仅作服装参考,图中下装为补全且不入镜。
Picture 3:references/scene-empty-night-v1.png
SHA-256:c81e1514dc223489468930f5957fa81bb8187027ebdefa92e8e5b7ec527649e7
用途:AI-generated generalized night-street palette/composition reference only; source foreground woman and left-rear pedestrian omitted by design.
ALT:AI生成的空夜街参考图,蓝青散景与暖琥珀灯光、湿路面反光;移除了源画面左后方模糊行人。

Picture 1是既有、已确认的方知微v6身份资产,本轮没有重新生成它;左清晰头像锁脸,右三视图锁体型与默认长卷发。175cm为用户设定,成片下半身不可见,不能据此验证身高或全身体型。Picture2只供服装,裤鞋属源片不可见部位的推定。Picture3是AI派生空景,不是首帧硬锁或真实地点证明;源左后方模糊行人被明确省略。

【角色服装板的完整实际 ImageGen 原词】
实际参考为固定v6图,本次生成一张条纹上衣角色板。

Use case: photorealistic-natural
Asset type: production character identity and wardrobe reference board for an AI video shot
Input images: Image 1 is the user's fixed Fang Zhiwei identity reference and the sole authority for face, facial proportions, age appearance, base skin tone, body proportions, and default near-center-parted long dark brown-black wavy hair. Do not copy the source film actress.
Primary request: Create exactly one clean horizontal 3:2 four-panel character sheet for the same fictional adult Chinese woman Fang Zhiwei. Preserve the identity in Image 1; only change the task-specific clothes to match the source-shot wardrobe.
Scene/backdrop: Plain pale cool-gray studio background, no movie setting.
Subject: One adult woman, Fang Zhiwei. Keep her canonical facial structure, feature proportions, age appearance, skin tone, same body proportions and loose long dark waves from Image 1. Task wardrobe: a relaxed beige/ivory wide boat-neck off-shoulder top with thin dark horizontal stripes, matching the visible source garment; modest and nonsexual, shoulders naturally covered enough for a production reference, long sleeves, no cleavage emphasis. A small simple pendant necklace is visible in the source crop and may be included. Lower body is not visible in the source, so use a plain dark straight-leg trouser and dark flat shoes as inferred completion, clearly outside the shot crop.
Composition/framing: Left half has a large sharp front-facing head-and-shoulders portrait, hairline/crown to chin fully visible, neutral natural expression. Right half has three equal-width, same-scale complete head-to-shoes views: front, strict 90-degree side facing image-right and showing her own right side, and strict 180-degree back view with no look-back. Same garment, hair and body identity across views. Exactly four clearly separated panels; short Chinese labels outside figures only: 大头照, 正面, 侧面, 背面. Keep the large portrait and rear view sharp. Apply soft facial-feature-only blur to the FRONT full-body and SIDE full-body faces; do not blur hair, ears, neck, body or clothing. Keep face and body crisp in the large portrait.
Style/medium: Restrained high-quality realistic photographic character reference board, coherent natural anatomy and fabric texture, neutral studio light, no glamor retouching.
Constraints: Do not change or reinterpret the attached identity. No source performer face or hairstyle, no copied movie frame, no background people, no extra accessories except the small pendant, no extra panels, no props, no subtitles, no readable text beyond the four labels, no watermark, no signature, no exaggerated proportions.

【空景板的完整实际 ImageGen 原词】
实际参考为源n203的正确ROI(x=0,y=350,w=576,h=324);仅用于空景生成,不把源演员脸作为H3图像输入。

Create a clean, photorealistic empty night-street scene reference plate based on the supplied frame. Preserve the recognizable street perspective, distant storefront/road depth, cyan-teal bokeh and amber street lighting, soft night haze, and the same calm color palette. Remove the foreground woman and remove the out-of-focus pedestrian at the far left rear; this is an intentional adaptation choice, not a claim that the source had no pedestrian. Reconstruct the background naturally where they were. No people, no readable text, no subtitles, no watermark, no logos, no graphic overlays. Keep the camera angle and medium street framing suitable as a background-only reference for an AI video scene. This is an environmental reference, not a new shot from the source.

【观察、偏差与未核验项】
检查全部56帧索引缩略联系表,另看交付n0/n16/n53的1664×928全幅帧。抬手约在n11开始,n16(约0.667秒)食指已竖在唇前,持姿延续到交付n53(2.208秒);时点来自输出抽帧,不是把提示目标当结果。指尖偏高、接近鼻底,领口比源片覆盖更多;未做更低指尖位置的对照试验。
有限样帧里脸部结构、肤色与深色长卷发大体符合固定身份,未见明显换脸;角度和光线有变化。不能据此保证逐像素身份或全身体型一致。无可见切镜、额外人物、字幕或设定板标签。
源音轨和H3生成音轨均未听审;不判断对白、配乐、环境声、听感或主观声画同步。静图Ref2VA是动作和构图改编,不是逐帧视频条件迁移。