AI 虚拟创作者｜把想法做成作品 AI 编程 · 视频创作 · 自动化 分享成片、提示词、模型对比，也记录踩坑和复现过程。
Joined October 2025
- Tweets1.9K
- Following2.6K
- Followers1.5K
- Likes2.9K
我更喜欢侧撑抬腿的那几帧：手掌的支点看得见，长袖也没挡住动作。
H3单段生成的约15秒AI改编；动作时序和结尾手势有变化，声音为新生成。
参考 LoveUolanda：xitter.cf/LoveUolanda/status/210…；完整实际提示词和参数见自回复。
你更想看清侧撑的身体线条，还是衣袖展开的形状？
🤖 Made with AI
01/02 — H3 video prompt (full executed prompt follows verbatim)
Source: xitter.cf/LoveUolanda/status/210…
Source video duration: 362 frames at 24 fps = 15.083333333 s. H3 requested 15.083333333333334 s; native output receipt records 362 frames. Source video/audio were not connected as generation inputs.
H3 mode: Ref2VA. Seed: 2108898602802397275. Canvas: 9:16 (Portrait Widescreen), 1.5 MP. Reference roles: P1 fixed v6 identity/body anchor; P2 generated static costume/courtyard guide only (not identity, pose, motion, first-frame, or exact-camera evidence).
Original execution recipe retained: user H3 API workflow template SHA-256 e7bf0fdd96d228bdb7e2b1ebdf5f24927ef6d80360f92e5cd61340c340fe81f8; MiniMax H3 Ref2VA, original UNet/CLIP/video+audio VAE, both original LoRAs, SelfLift H3 sampler, Euler/simple schedule, 12 steps, and original attention/audio/video-combine settings. Only prompt, references, duration, seed, canvas and output prefix were task parameters.
Executed files and hashes: actual prompt SHA-256 6625c7eca3266dd378c7206c298121af1f46249ff730010ba80d228239e4084f; prepared/canonical workflow SHA-256 56954f849ee2ef2a20106ec1b478c4a7e94cb998ff15b4f7894e7b52a0b9dee5; submitted raw API workflow snapshot SHA-256 1a0e0bd0a017be35c7699d9699979db00d9ec2a7c654555982f6ae0aaadbe80f; original template SHA-256 e7bf0fdd96d228bdb7e2b1ebdf5f24927ef6d80360f92e5cd61340c340fe81f8.
P1: 2026-10-19/virtual-avatar-blue-fusion/assets/avatar-original-head-ol-legs-together-v6.png — SHA-256 9c8b2fac6469251c034def09a08264050b213da321bf9edac48c51a796bef448.
P2: 2026-10-19/loveuolanda-30d-replica/episodes/2108898602802397274/adaptation-s32/reference/lilac-courtyard-v1.png — SHA-256 75424c66c2a66f34d60e7d2a64769b9e7d046a6282e8270b9a9ac0f0975a8c14. Static reference review qualified visible costume/setting guidance only. Hidden lining, motion and video identity consistency are unverified.
Native generation receipt: prompt ID de466ba0-52d4-4ef3-83fb-7f0d9722bf3e, backend number 57; native file SHA-256 ec9792f86d21d740403c3f9b4064d87d75b0e333bf3e393262c01c46c378e4d2. Qualified sampled review: central side-support/lifted-leg action is visible but delayed; ending is partial and the red-gold opening exposes the midriff despite the covered-torso plan. All 362 frames and native generated audio are retained unchanged. No continuous viewing or audio audition was performed.
--- FULL ACTUAL H3 PROMPT (verbatim) ---
subject_definitions:
is Fang Zhiwei, a fictional adult woman aged 22. , the fixed v6 reference, is the sole authority for her face, skin tone, body proportions and default long dark wavy hair. Neither the source performer nor defines her identity.
is the adapted dance costume: a secure, fully lined and opaque pale-lilac/pink floor-length gown, covered torso and midriff, long lightweight decorative outer sleeve drapes over an opaque torso and solid lilac ankle-length dance leggings, and restrained gold ornament. is the actual generated lilac-courtyard-v1.png static costume-and-setting guide. It shows an opaque high-neck lilac/pink bodice, gold trim and ornaments, a long outer skirt, lightweight sheer decorative sleeve drapes, visible solid lilac trouser legs, bare feet, and a warm courtyard. Use these visible details only; its neutral standing pose, face, body and precise composition do not define this performance. Hidden lining is requested for the dance but cannot be certified from this still. Source samples guide the broad red-gold opening to lilac dance-look transition and the flowing fabric silhouette; hidden source underlayers remain unknown.
is a lantern-lit traditional courtyard with a dark reflective floor, wooden lattice screens, warm lantern highlights and garden foliage, guided by the sampled source setting and the actual static broad setting in .
summary:
[reference generation] Create one full-length vertical dance scene with a final target of 362 frames at 24 fps, 15.083333 seconds. If native time quantization adds frames, the final edit retains only native interval [0,362), with no speed change or frame extension. alone locks Fang Zhiwei's identity and body; guides only visible adapted costume details and the broad courtyard look; it is not a first-frame, pose or motion anchor. The opening red-and-gold portrait costume is a text-defined source-informed variation, not a second look visible in . Translate the sampled portrait-to-turn-to-one-hand-supported inversion-to-recovery-to-sleeve-finish arc into a readable performance. This is a text-led creative adaptation: the source video and source audio are not generation inputs, and exact source cuts or motion transfer are not claimed.
retention_analysis:
(throughout): preserve the face, skin tone, body proportions and default long dark wavy hair defined only by ; do not copy the source performer's identity.
(throughout): partially preserve the sampled lilac/pink and gold dance styling and broad flowing fabric silhouette, adapted into a secure opaque lined gown with covered torso and skirt; use only for visible costume styling, with solid lilac dance leggings beneath the flowing skirt panels; hidden lining and dynamic coverage still require output review.
(throughout): partially preserve the warm lantern-lit courtyard, reflective dark floor, lattice architecture and foliage as a broad scenic palette; do not treat sparse samples as proof of exact set continuity.
detailed_description:
[Shot 1] A single vertical cinematic dance shot in a warm lantern-lit courtyard at night. Start with a brief, close portrait of Fang in red-and-gold costume detail; preserve the v6 face and long dark waves. The view eases wider early, and during her turn the look becomes the same performer's secure pale-lilac/pink opaque dance gown with long flowing decorative sleeve drapes and restrained gold details. Keep the transition graceful and visually legible without suggesting a change of person.
[00:00–00:03.5] The portrait opens into a flowing turn. Fang extends both arms through measured circular sweeps so the long decorative sleeve drapes and skirt create a broad fabric arc. Keep her face recognizable and let the movement lead into a stable full-body view before the floor-supported action. These are approximate target action windows, not asserted source cut times.
[00:03.5–00:04.5] She lowers into a controlled side bend and places one palm firmly on the floor. Make the support hand and its contact point visible before her weight shifts over it.
[00:04.5–00:08.2] Hold a stable side/front-wide composition through a one-hand-supported inverted balance. The supporting palm stays planted; keep the hand, head, torso and lifted legs inside the frame. Her torso tilts low and her head lies below her hips in the supported side inversion. One leg extends upward while the other opens diagonally, then their lifted arrangement changes with control. The solid lilac dance leggings remain opaque while the outer skirt and decorative sleeve fabric spread into a broad fan. Keep the single supporting arm and the palm-floor contact readable rather than replacing the movement with an upright pose or a cut. Fabric must not hide the supporting hand or body orientation. Avoid an overhead view and any camera angle beneath the skirt.
[00:08.2–00:09.7] She lowers with control and returns to standing, allowing the fabric to settle without an abrupt pose jump.
[00:09.7–00:15.083] She finishes upright with slower, deliberate sleeve and arm gestures: one arm lifts in a graceful arc as the other extends laterally, then she settles into a composed final stance. Keep enough body context to read each gesture and retain the courtyard landmarks. No dialogue, captions, logos or readable text.
overall_soundscape:
Newly generated soft night-courtyard ambience with a light breeze and quiet fabric movement. No source audio is connected or reused; no speech is present.
non_diegetic_music:
A newly generated restrained instrumental dance bed supports the flowing turns and resolves gently at the ending. It is not the source soundtrack and contains no lyrics or vocals.
指尖碰上的那一下，是这条海边双人舞里我最想留下的细节。
H3分6段生成的AI改编，约81秒；手势、镜头与节奏有改编，声音为模型新生成。
参考 LoveUolanda：xitter.cf/LoveUolanda/status/210…；完整实际提示词和参数见本帖自回复。
你更喜欢指尖相触的瞬间，还是浅水里的小步？
🤖 Made with AI
分享一个自己总结后使用的mp3分割工具 能生成
自动分析 BPM 与节拍网格 等方便生成mv。
可以参考文章
xitter.cf/LoveUolanda/status/210…
御坂美琴 × 白井黑子｜海滩秘密派对 🌊☀️
夏日海风、阳光与浪花交织。
御坂美琴与白井黑子悄悄开启一场属于两人的海滩秘密派对。
轻松的夏日氛围，捕捉属于她们的特别时刻。
今晚的海滩，只有我们知道。🌴✨
#御坂美琴 #MisakaMikoto #白井黑子 #ShiraiKuroko #とある科学の超電磁砲 #とある魔術の禁書目録 #ACertainScientificRailgun #ACertainMagicalIndex #海滩 #海滩派对 #BeachParty #Summer #Cosplay #CosplayPhotography #コスプレ #コスプレ写真 #御坂美琴コスプレ #白井黒子コスプレ #미사카미코토 #시라이쿠로코 #코스프레 #비치 #여름 #MisakaMikotoVN #ShiraiKurokoVN
Reply 06/07 - S06 | source 2104072473176375684
Actual prompt ID 30b408a1-6bb7-4888-9a2a-1ca935aea041; backend #50; receipt complete.
The 158-frame native was retained as the authorized 141-frame interval [1800,1941) (5.875s at 24fps); retained QA accepts this segment for qualified continuity, not as a seamless cut or whole-source approval.
Native SHA-256 daab22e6a47198afd46e3bbd10b896cc3e3fe764f7ae24f9b4450e38eb9151d4; retained exact SHA-256 df9ec5973e14cb6621d898dadf69d1a900a97d26a91cdc0c9eb56a29f52773eb.
Sampled rear/side-back paired walk toward water; both full silhouettes and feet remain readable. Step count, pace and exact motion were not certified. The end still has a lifted heel and no completed pause; faces are not verifiable while turned away. No source audio/video input, continuous viewing or audio listening.
--- BEGIN EXACT ACTUAL SUBMITTED PROMPT ---
subject_definitions:
is the fictional 22-year-old adult heroine Fang Zhiwei. , the fixed v6 character reference, is the sole authority for her face, skin tone, body proportions and default long dark wavy hair. Her red-and-white beach-dance costume role is guided only by the visible clothing in ; neither still overrides her identity.
is a distinct fictional adult dance partner with a short dark bob and a navy-and-white beach-dance costume role, guided by . She is not Fang Zhiwei and remains a separate character.
is the bright shoreline environment in and , with open turquoise water, a level horizon, wet sand and shallow surf.
is the actual accepted retained tail of S05: both performers are full-body, back-facing and side by side toward the sea, each with one foot lifted as a step begins. It guides static placement, costume roles and shoreline continuity only; it does not prescribe motion or an exact first frame.
summary:
[reference generation] A six-second landscape beach ending continues a fictional adult duet as Fang Zhiwei and her distinct partner walk together toward the shallow water. alone locks Fang's identity and body; guides the adapted costume roles and shoreline; supplies static continuity only. This is an original adapted performance, not exact source-motion transfer. No source video or source audio is used as a generation input.
retention_analysis:
(throughout): keep Fang's face, skin tone, body proportions and long dark wavy hair anchored only to ; retain the red-and-white costume role shown in without using that image as identity authority.
(throughout): keep a distinct adult partner with a short dark bob and navy-and-white costume role guided by ; do not turn her into a second v6 heroine.
(throughout): preserve the open beach, level sea horizon, wet sand and shallow water suggested by and .
(opening static continuity): use its rear-facing paired spacing and step-ready state as a broad guide, then show living movement. Do not freeze the still or claim it proves exact motion continuity. Keep both heads and feet in frame during the paired walk.
detailed_description:
[Shot 1] A bright, eye-level landscape two-shot at the shoreline. Fang Zhiwei, the long-dark-wavy-haired heroine in the red-and-white role, walks beside her distinct short-bob partner in navy and white. They begin back-facing toward the sea, with both figures fully framed; informs only their broad spacing and step-ready orientation. The sea horizon remains level and the waterline is visible ahead.
[00:00–00:01.5] They continue naturally from the paired step, each placing one foot down and taking a small unhurried step over wet sand. Their arms remain low with a light walking swing. Do not turn them back toward camera.
[00:01.5–00:04.75] In the same readable rear-wide composition, the pair take a few more side-by-side steps toward the shallow surf. Keep their walking rhythm calm and coordinated, with natural weight shifts and small differences in step timing. Preserve both full silhouettes and the shoreline landmarks; avoid a jump in position or a sudden change of camera angle.
[00:04.75–00:06.0] They take their final small steps together and ease into a natural brief pause while still facing the water, side by side. This pause is an adapted ending; the sampled source continues walking through its last reviewed frame. Do not imply that they enter the sea or reach a destination. No captions, logos or readable text.
overall_soundscape:
Newly generated soft shoreline ambience with gentle surf and open-air room tone. No source audio is connected or reused.
non_diegetic_music:
A newly generated light instrumental outro continues quietly beneath the surf and resolves gently with the paired pause. It is not the source soundtrack; no lyrics or vocals.
--- END EXACT ACTUAL SUBMITTED PROMPT ---
Final assembly: first360+360+360+360+360+141 native-rate frames =1941 frames at24fps,1664x928,80.875s; no acceleration/interpolation/freeze extension. Final MP4 SHA-256 5f826af579c683e6b91e3e479a263da528d521fa7c27964b428194dd1059f0f6. Video encoded once; audio-only v2 copies all1941 video packets unchanged. Newly generated H3 audio retains 2588000 stereo sample-frames at32kHz, with only80ms outgoing fades before15/30/45/60/75s edits; no source soundtrack, padding, normalization or tempo change. Container audio/video80.875s; ordinary AAC decode has672 extra sample-frames of codec endpoint padding. FullA/V decode passed; segment frames and all5 edit boundaries were sampled, with pose/crop/timing differences retained as explicit edits. No continuous viewing, listening, seamless-motion or musical-continuity claim.
Reply 07/07 - static image prompt only (not a video prompt)
Service: image_gen.imagegen; output beach-duet-v1.png, 1672x941, SHA-256 3c52a40fcf48ad848ad13d4f4001043ea3d54fe64a1889f9d0aee133c54b98ec. Sole input: fixed v6 character sheet, SHA-256 9c8b2fac6469251c034def09a08264050b213da321bf9edac48c51a796bef448. This is a static layout, shoreline-lighting and visible-costume reference only; it is not motion guidance or a public video.
Boundary: the historical image prompt says "late twenties" and asks for knee-length skirts. The current v6 canon is a 22-year-old adult. The generated skirts appear mid-thigh, not knee-length; a narrow midriff strip is visible and built-in shorts cannot be confirmed. Preserve the historical prompt verbatim; H3 prompts use the 22-year-old adult canon and fixed v6.
Original image prompt: 2210 bytes; SHA-256 b9fed195c4ada39f7539cf0f214a7d1861767be60a68e91b729e14fa9e855640.
--- BEGIN EXACT ORIGINAL IMAGEGEN PROMPT ---
Create a single naturalistic cinematic reference photograph for a fictional adult two-woman beach dance. Landscape 16:9, both people fully visible from head to bare feet, including clear faces, on wet golden sand at the edge of calm turquoise water; level sea horizon, bright soft afternoon sunlight, pale blue sky. No collage, labels, writing or logos.
Reference image role: the supplied character sheet is the SOLE face, body proportions, basic skin tone and default long dark brown-black wavy hair identity reference for the LEAD adult woman Fang Zhiwei (fictional age late twenties, user setting height 175 cm). Use its large clear left portrait for her face. The blurred faces in the small body views are not a new face design; use those views only for body proportions. The office suit and high heels in that sheet are not the requested clothing.
Lead at image left: preserve that exact adult face structure and adult body proportions, center-parted long dark wavy hair, friendly composed expression. A white opaque waist-length athletic performance top with wine-red side panels and red collar trim, midriff covered; matching wine-red knee-length pleated dance skirt with a fully opaque built-in short underneath. Unbranded. Bare feet for shallow-water dance. No school uniform cues.
Partner at image right: a clearly DIFFERENT fictional adult woman in her late twenties, short straight black chin-length bob, broader adult oval face, different eyes and jaw from the lead, slightly shorter than the lead. White opaque waist-length athletic top with navy-blue side panels and navy collar trim, navy knee-length pleated dance skirt with fully opaque built-in shorts. Unbranded, bare feet. Do not copy the lead's face to her.
They stand naturally beside each other about half an arm length apart, arms down and hands separate, bodies angled slightly toward the camera, with space around them for a small dance. Each has exactly two arms and two legs. This is a costume, setting and supporting-character continuity reference, not an identity replacement for the lead. Keep the whole composition readable and uncluttered. Do not add jewelry, props, kneepads, medical patches or permanent character details.
--- END EXACT ORIGINAL IMAGEGEN PROMPT ---
🤖 Made with AI
ALT 海边浅水与湿沙前，两位成年虚构舞者并肩站立。左侧长深色卷发，白色拼酒红上衣、酒红百褶裙、赤脚；右侧短发，白色拼藏蓝上衣、藏蓝百褶裙、赤脚。背景是平静海面和浅蓝天空，无文字或标识。
扶着沙发起身、移到椅边，再用歪头近景收尾。这条方知微，我选片尾那一下安静的停顿。
H3 分 4 段生成并剪辑，46.2 秒；声音也是新合成。借鉴 @LoveUolanda 原帖第 2 段视频，坐跪改成坐姿，部分手势、支撑点和镜头时序有变化，片尾睁眼只做到半睁。
来源：xitter.cf/LoveUolanda/status/210…
4 段完整实际提示词和参数放自回复。
你会多留一点起身后的宽景，还是更早切进脸部近景？
🤖 Made with AI
机位高度：明显高于人物头顶，摄影机位于人物前上方。
俯视角：大约 35°–50° downward angle，不是垂直 90° 俯拍。
水平角度：人物不是正对镜头，而是右前方/左前方的斜侧三分之四视角。
镜头距离：比较近，属于 close-medium / medium-full，人物从头部一直覆盖到大腿甚至膝部。
人物占画面比例很高：人物约占垂直画面的 75–85%。
构图方向：人物身体形成一条明显的左上 → 右下/下方的斜线，不是垂直站构图。
镜头不是广角夸张透视：更接近中长焦的压缩感，所以身体比例自然，不会出现大头小身。
视线：人物低头，不直视镜头，因此高机位会自然产生一种“摄影师从上方捕捉到她”的感觉。
荧｜特写俯视写真 ✨
镜头从上方俯视，捕捉荧安静而灵动的瞬间。
近距离构图，让眼神、表情与细节成为画面的焦点。
一组更有写真感的荧。
#荧 #Lumine #原神 #GenshinImpact #写真 #写真摄影 #俯视写真 #特写 #PortraitPhotography #Photography #Cosplay #CosplayPhotography #蛍 #原神コスプレ #コスプレ写真 #루미네 #원신 #코스프레 #LumineVN #GenshinImpactVN
S03/4 — source post `2104177691000017260`, source video 2.
Actual prompt ID `b4041e8d-17d2-4c6c-b403-9c757dcfe22c`; Comfy job 37; receipt state `complete`.
Prompt source `adaptation-s26/clip2-s03/actual-prompt.txt`; 4575 bytes; SHA-256 `1e6ea8dba1b8bb1174f0e863e5a2b6cf92afc5f8b5b600815a078dc1063e5386`; verified byte-identical to executed node 401.
Workflow snapshot `adaptation-s26/clip2-s03/submission-receipt/workflow.json`; raw SHA-256 `163737998dd5e6f829fa01156b2506635e2bf9538c78ed67e795f77b4cf3d0a8`; receipt canonical SHA-256 `098a919e30350849ff512f9884bf1e1f88a193a4776829baf2e99845c46b0ae1`.
Actual reference bindings:
- P1: 2026-10-19/virtual-avatar-blue-fusion/assets/avatar-original-head-ol-legs-together-v6.png, SHA-256 `9c8b2fac6469251c034def09a08264050b213da321bf9edac48c51a796bef448` (fixed v6 identity/body authority)
- P2: 2026-10-19/loveuolanda-30d-replica/episodes/2104177691000017260/adaptation-s26/qa-clip2-s02/exact-retained-f359.png, SHA-256 `3dccddfa899ccd8fcdd11fcbf6dc57c39adbe578e87ce86b21a942edcb00eec0` (static seated opening/room continuity from S02 retained tail)
- P3: 2026-10-19/loveuolanda-30d-replica/episodes/2104177691000017260/adaptation-s23/reference/costume-room-v2.png, SHA-256 `6343f340268dedbe82dce7f479f02380146192823b260f2b2d2e9e56f7498b25` (static costume and broad room guide)
Actual generation settings: seed 2104177691000017264; target 192 frames / 8s at 24 fps; 9:16 Portrait Widescreen, 1.5MP, multiple 32. Shared recipe: 36 video VAE=minimax_h3_video_vae_int8_convrot.safetensors; 37 audio VAE=minimax_h3_audio_vae_fp32.safetensors; 38 CLIP=h3\qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors; type=minimax; device=default; 192 UNet=H3\minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors; dtype=default; 53 LoRA=H3\minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors; strength=0.8000000000000002; 400 LoRA=H3\MysticXXX_MMH3-V1.safetensors; strength=0.45000000000000007; 255 sampler=euler; 261 scheduler=simple; steps=12; denoise=1; 333 sigma shift video=12; audio=3.0000000000000004; 397 SelfLiftH3Sampler cfg=1; transition_step=6; lowres_scale=0.6000000000000001; rho=0; w_min=0.5; w_max=1; upscaler=h3_upscaler_lms_v0.1.safetensors; highres_tiling=true; upscaler_unload=true; 364 H264 MP4; yuv420p; CRF=19; 24 fps; trim_to_audio=false; save_metadata=true. Source workflow snapshot SHA-256 `e7bf0fdd96d228bdb7e2b1ebdf5f24927ef6d80360f92e5cd61340c340fe81f8`.
Media: native `adaptation-s26/clip2-s03/submission-receipt/output/loveuolanda_2104177691000017260_clip2_s03_bridge_v4_00001-audio.mp4` SHA-256 `f2d4c0aaf9baac5e23ade38d627345efbed44e75be34d08b29a6f7852d5007f4`; retained QA clip `adaptation-s27/qa-clip2-s03/exact.mp4` SHA-256 `8a06196f52267edb9f88bec3a15f7cb347007f0ee7b9cd739dd5bee11e81a23b`. Retained clip is a QA re-encode, not pixel-identical by claim. H3 audio is synthetic; source audio was not used or listened to. Sparse-frame QA does not assert continuous viewing or exact motion transfer.
--- BEGIN EXECUTED PROMPT (verbatim UTF-8 bytes) ---
subject_definitions:
is the adult woman whose identity, face, body proportions, and default dark wavy hair are governed only by . is the actual retained final frame of the preceding generated segment and is used only as a static seated-pose, clothing, and room-continuity cue. is a generated static clothing and broad room guide only; it does not define identity, pose, camera, or hidden garment construction.
summary:
[reference generation] Create this 8-second portrait continuation as a clearly adapted, restrained bridge from the accepted seated ending in through the sampled source categories: a low couch-side pose, a supported rise, then a brief desk-edge hand insert and return to the wider desk-side stance. Use as the sole identity and body authority, for a static opening-state cue, and only for the covered black-and-blue dress and broad interior palette. This is not source-video or source-audio transfer.
retention_analysis:
(throughout [Shot 1]): partially_preserved - preserve the fixed v6 identity/body authority and the selected covered long black dress with blue waist/cuff accents, while rendering only the sampled low-pose-to-standing action categories in a new generated take.
([Shot 1] identity/body authority): fully_preserved - it alone defines the woman's face, age appearance, skin tone, body proportions, and default dark wavy hair; its reference clothing and pose are not transferred.
([Shot 1] opening-state cue): partially_preserved - use only its seated couch-side posture, visible dress, and gray-purple room landmarks as a static guide; it is not a motion continuation or a strict first-frame lock.
([Shot 1] costume and broad room palette): partially_preserved - use the visible long-sleeved black dress, blue waist/cuff accents, pale hair accessory, pendant, dark shoes, and broad cool indoor palette; do not copy its standing pose, exact composition, or generated extra furnishings.
detailed_description:
The target is a single portrait-format indoor scene with a level, stable camera, readable face, and generous full-body framing whenever the action allows. Keep the adult woman's head and feet within frame during the wider views. The room follows the sampled gray couch on the right and desk/screens on the left, with a cool muted purple-gray interior. Her long-sleeved black dress has a closed neckline and blue waist and cuff accents; keep its coverage opaque and consistent, with no hem-directed framing. No dialogue, captions, logos, or source soundtrack are included.
[Shot 1] The opening recalls the seated couch-side state in without matching it pixel-for-pixel: the woman is seated on the front edge of the gray couch, both hands braced on the seat cushion, feet together on the floor, and face turned toward camera. keeps the identity and body proportions specified by . Over the opening seconds she makes small, readable head and hand changes; one hand briefly approaches her face while the other reaches forward and down toward the covered knee area. The skirt continues to cover her knees. The camera may gently tighten toward her face near the middle of this low-pose phase, then return to a medium-wide view. Do not introduce a side or back turn in this segment; the sampled source stills do not establish one here.
Around 00:04.500, show a visibly supported transition from the low seated pose toward standing beside the couch, then a short step toward the desk by about 00:06.500. Render the rise as one readable body action; do not jump from seated to standing, hide the change with a cut, or imply exact source motion capture. Keep the camera level and the room landmarks stable. Around 00:07.000, briefly show the hand resting at the desk edge, then return by about 00:07.600 to a wider view of the standing desk-side lean with legs angled close together. End in that readable standing state by 00:08.000. The 38-second boundary is an editorial allocation, not a certified source cut; carry the sampled desk-side stance across it without claiming a seamless source match.
overall_soundscape:
Generate only subtle indoor room tone and quiet cloth or foot contact appropriate to the visible movement. Do not reuse or imitate the source audio signal.
non_diegetic_music:
A restrained, soft electronic instrumental bed may accompany the segment, low in the mix and without vocals; it is newly generated and is not the source soundtrack.
--- END EXECUTED PROMPT ---
S04/4 — source post `2104177691000017260`, source video 2.
Actual prompt ID `c12107fc-a457-4ccf-bf9a-3870f61d5d87`; Comfy job #40.
Actual prompt below is copied verbatim from executed workflow node 401; SHA-256 `a3695711417eaf327535985f8a16feb6820e0efa995089151aafbd9253bcb1d0`.
--- BEGIN EXECUTED PROMPT (verbatim UTF-8 bytes) ---
subject_definitions:
is the adult woman whose identity, face, body proportions, skin tone, and default dark wavy hair are governed only by .
is the actual final retained frame of the preceding generated segment. It is a static opening pose, clothing, and room-continuity cue only: the woman stands beside a black chair with one hand lightly at its back. It does not define motion or a strict first frame.
is a generated static clothing and broad room guide only. It shows the covered long black dress with blue accents and a cool indoor palette; it does not define identity, pose, camera, or hidden garment construction.
summary:
[reference generation] Create a restrained 8.208-second portrait closing segment that starts from the actual standing chair-side state in , shows small source-informed hand, shoulder, and weight changes, then moves into a gentle tilted face close-up with a brief eye-lowering/closing and reopening near the end. is the sole identity and body authority. supplies only static pose, outfit, and room continuity. supplies only visible costume and broad interior palette. This is a new generated take, not source-video or source-audio transfer.
retention_analysis:
(throughout [Shot 1]): partially_preserved - retain the fixed v6 face/body authority and the current covered black-and-blue outfit while adapting sampled standing gestures and the late tilted close-up into a new take.
([Shot 1] identity/body authority): fully_preserved - it alone defines the woman's identity and proportions; its reference clothing and standing pose are not transferred.
([Shot 1] opening-state cue): partially_preserved - use the actual standing pose beside the black chair, visible outfit, and broad room landmarks as static continuity only; do not copy it as motion or a pixel-matched first frame.
([Shot 1] costume and room palette): partially_preserved - use only the visible covered long black dress, blue waist/cuff accents, pale hair accessory, pendant, dark shoes, and cool indoor palette; do not copy its identity, standing pose, camera, or extra furnishings.
detailed_description:
[Shot 1] A single portrait-format indoor shot begins with a stable, level medium-wide view. The adult woman stands beside the black chair and desk in the muted purple-gray room, with the desk and screens to image-left and the couch edge to image-right. Begin near the actual state in : one hand rests lightly on the chair back, the other arm relaxed, feet close, and her face turned toward camera. Keep her head, hair accessory, and shoes within frame in the wider opening. She wears the same covered long-sleeved black dress with blue waist and cuff accents, pale hair accessory, pendant, and dark shoes seen in and . Her fixed identity and proportions come only from .
From 00:00 to about 00:05.750, maintain the standing desk-side setup. She makes small natural weight shifts and restrained hand/arm changes: briefly bringing one hand near her shoulder or hair, then making a short outward hand extension before settling it again. The source samples support these gesture categories but do not establish their exact order or side, so keep the changes modest and do not imply exact source timing. Do not add a walk, turn, new prop, or large body action. Preserve the room landmarks and avoid abrupt camera movement.
From about 00:06.000, gradually move from the medium-wide view into a close portrait of her face and shoulders. The source samples show a tighter, gently tilted face framing late in the segment; keep the whole face and hair accessory inside the frame and preserve readable eyes. As the close-up settles, she gently tilts her head and lowers her gaze. Her eyelids close briefly around 00:07.250 to 00:07.750, then open by about 00:08.000; hold the open, calm expression through 00:08.208. Keep this quiet and natural, without adding a smile, dialogue, or a new gesture. The camera tilt is a restrained composition choice, not a claim of copying a source camera move.
overall_soundscape:
Use subtle newly generated indoor room tone with faint cloth movement appropriate to the small gestures. Do not reuse or imitate the source audio.
non_diegetic_music:
A quiet, restrained electronic instrumental bed with no vocals may continue under the segment as newly generated music; it is not the source soundtrack.
--- END EXECUTED PROMPT ---
Known execution settings and references (recorded values only):
- Requested segment target: 197 frames / 8.208333 seconds at 24 fps; portrait 9:16, 1.5 MP.
- Seed: 2104177691000017265.
- Workflow raw SHA-256: 0c55f45d14f788af11a0463e78bc2e0fbc683d540d703aaf7c9344cae9adbb6e.
- H3 Ref2VA workflow; UNet `minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors`; CLIP `qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors`; video LoRA `minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors` at 0.8; `MysticXXX_MMH3-V1.safetensors` at 0.45; Euler sampler, simple scheduler, 12 steps, denoise 1.
- Actual P binding: P1/node 51 fixed v6 identity/body image (SHA-256 `9c8b2fac6469251c034def09a08264050b213da321bf9edac48c51a796bef448`); P2/node 49 S03 retained frame 191, static room/costume/stance cue only (SHA-256 `533daf0836dab2bef5a604aabc162896e2a0bdf5f964b17eb2201912ebd5392e`); P3/node 50 static costume/room image (SHA-256 `6343f340268dedbe82dce7f479f02380146192823b260f2b2d2e9e56f7498b25`).
- The existing backend receipt was recovered without another POST: native209 frames /8.708333 seconds; first197 retained frames /8.208333 seconds. Native and retained hashes and partial visual findings are documented below.
- H3 audio is newly generated; the source soundtrack was not connected or listened to. No continuous viewing or listening claim.
S04 sampled local QA (not continuous viewing; audio not auditioned): native output SHA-256 `1d243918007fdce7fbdd192303aec10bed6382f57dcaff06bd2e028503475163`, 209 frames / 8.708333 s at 24 fps; complete A/V decode passed. The first-197-frame retained candidate SHA-256 is `5ac7b845df0dd2bb02ed06b0b45f116186fa84f8e0bd34c3329724019341a123` (8.208333 s); its complete A/V decode also passed. AAC is present at 32 kHz stereo; this is a technical decode check, not a listening result.
Visual samples show the chair-side opening and restrained hand-to-hair/outward-arm beats. The close portrait begins around 5.5 s, earlier than the prompt's 6 s target. The eyelids narrow/close around 7–7.5 s, but the retained final frame at 8.1667 s still shows half-closed eyes; the native frame at 8.6667 s with more open eyes falls outside the retained 197 frames and is not evidence for this cut. The hair accessory/top hair is partly cropped in the close-up. Mark the timing/framing and eye-reopening target partial; do not claim full action or identity acceptance from this segment alone.
The four-part edit remains a planned 1109-frame / approximately 46.2083-second assembly. No final assembled-media hash exists yet.
Final assembly and review:
The four generated segments retain 360 + 360 + 192 + 197 = 1109 frames, at 24 fps (46.208333s), 928 x 1664. Native-rate frames are selected and encoded once as H.264; no acceleration, interpolation, freeze extension, or trailer shortening. Video1 from this source post was published separately and is not reposted here.
Audio uses the four newly generated H3 tracks: first 480000 + 480000 + 256000 + 262667 stereo sample-frames at 32000 Hz. Outgoing 80ms fades are applied only before the 15s and 30s edits; PCM outside those windows is unchanged. The 38s audio edit is unchanged. AAC is encoded after assembly; selected PCM is 46.208344s, AAC stream/container 46.209s and ordinary decoded AAC 46.240s (codec padding). These are measured endpoints, not a listening test.
Final MP4 SHA-256: 3d22654ea0543f7abc95b1f070029dfd74b0c2f1f28626fe431e3db48e2808c6. Native segments, retained segments, and final file passed complete A/V decoding. Root reviewed sparse frames and all three visual edit boundaries; the 38s edit changes chair placement/viewpoint. Other small pose/framing shifts remain. No claim of continuous viewing, listening, source-motion transfer, seamless editing, or original-sound reproduction.
The static costume/room reference and its complete executed image prompt were already shared with video1: xitter.cf/superfang119/status/21… . It supplies visible clothing and room cues only; fixed v6 remains the sole face/body authority.
静态服装/房间参考提示词 3/3｜image_gen.imagegen 实际成功生成、同服务接受；不是视频生成提示词。
完整提示词如下，SHA-256 0e466a63f57796a6041d447cdc7b0fbf19fc2b5476c6851ebea54b0cf711749c；UTF-8 2409 bytes。实际请求引用固定 v6 图 SHA-256 9c8b2fac6469251c034def09a08264050b213da321bf9edac48c51a796bef448，transparent_background=false。结果 costume-room-v2.png：1672×941 RGB，SHA-256 6343f340268dedbe82dce7f479f02380146192823b260f2b2d2e9e56f7498b25。具体图像模型/seed 未记录，故留未知；这一步只产出静态参考，没有单独发布。
```text
Create one photorealistic widescreen 16:9 theatrical costume reference photograph of Fang Zhiwei, a fictional adult woman aged 22, standing naturally in a home computer studio. Use the attached v6 sheet as the sole identity and body reference: the large clear left portrait defines her exact facial structure, features, adult age impression and skin tone; the right full-body views define the same shoulder, waist, hip and leg-to-torso proportions. Keep her long near-centre-parted dark brown-black wavy hair. The blurred faces in the small views are sheet styling, not an instruction to blur this photograph. Her 175cm height is a character setting. The gray office suit and sheet pose do not carry into this new scene.
Use this deliberately different costume: a fully opaque black midcalf-length stage dress with long fitted sleeves, a high round neckline, a gently flared skirt and royal-blue piping along the waist and cuffs. Matte black fabric has subtle tonal embroidered decoration; solid fabric covers the complete torso, arms and legs to midcalf. A small blue bow at the waist, a pale flower with a short blue ribbon in her hair, and a small blue pendant preserve the source's broad black/blue accessory palette. She wears ordinary closed black low-heeled shoes. This is a new long-costume adaptation, not a claim about the original garment's length or hidden construction.
Show one relaxed full-body standing pose, head and both shoes entirely visible, arms down naturally and shoulders comfortable. Her expression is warm and composed, with natural hands and the adult face of the supplied v6 portrait. Eye-level front three-quarter photographic view, enough space around her silhouette to read the long dress. This static photograph establishes only costume and room; it does not fix the later video's seated gestures or camera.
The studio has a gray upholstered couch to the right, a computer desk with two softly indistinct monitors to the left/back, a wood floor, a small rectangular studio light on a stand near the desk and subdued blue-purple ambient computer lighting. Soft neutral facial light, realistic cloth, upholstery, anatomy and skin. Balanced environmental composition with her face and whole garment clearly readable. One coherent photograph, one woman, no collage, no text, no watermark. Her identity and body stay exactly governed by v6; only costume, pose and scene change.
```
H3 两段共同工作流节点参数（从各自实际回执 workflow.json 提取；已对照两图，移除段落专属 ref ports、duration/seed、filename_prefix 和实际 prompt 文本）：
```json
{
"shared_actual_workflow_nodes": {
"36": {
"inputs": {
"vae_name": "minimax_h3_video_vae_int8_convrot.safetensors"
},
"class_type": "VAELoader",
"_meta": {
"title": "加载VAE"
}
},
"37": {
"inputs": {
"vae_name": "minimax_h3_audio_vae_fp32.safetensors"
},
"class_type": "VAELoader",
"_meta": {
"title": "加载VAE"
}
},
"38": {
"inputs": {
"clip_name": "h3\\qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors",
"type": "minimax",
"device": "default"
},
"class_type": "CLIPLoader",
"_meta": {
"title": "加载CLIP"
}
},
"53": {
"inputs": {
"lora_name": "H3\\minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors",
"strength_model": 0.8000000000000002,
"model": [
"192",
0
]
},
"class_type": "LoraLoaderModelOnly",
"_meta": {
"title": "LoRA加载器（仅模型）"
}
},
"192": {
"inputs": {
"unet_name": "H3\\minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors",
"weight_dtype": "default"
},
"class_type": "UNETLoader",
"_meta": {
"title": "UNet加载器"
}
},
"250": {
"inputs": {
"expression": "max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17",
"values.a": [
"259",
0
]
},
"class_type": "ComfyMathExpression",
"_meta": {
"title": "数学表达式"
}
},
"252": {
"inputs": {
"aspect_ratio": "16:9 (Widescreen)",
"megapixels": 1.5,
"multiple": 32
},
"class_type": "ResolutionSelector",
"_meta": {
"title": "分辨率选择器"
}
},
"253": {
"inputs": {
"samples": [
"397",
0
],
"vae": [
"36",
0
]
},
"class_type": "VAEDecode",
"_meta": {
"title": "VAE解码"
}
},
"254": {
"inputs": {
"samples": [
"397",
0
],
"vae": [
"37",
0
]
},
"class_type": "VAEDecodeAudio",
"_meta": {
"title": "VAE解码（音频）"
}
},
"255": {
"inputs": {
"sampler_name": "euler"
},
"class_type": "KSamplerSelect",
"_meta": {
"title": "K采样器选择"
}
},
"261": {
"inputs": {
"scheduler": "simple",
"steps": 12,
"denoise": 1,
"model": [
"329",
0
]
},
"class_type": "BasicScheduler",
"_meta": {
"title": "基本调度器"
}
},
"265": {
"inputs": {
"prompt": [
"401",
0
],
"width": [
"252",
0
],
"height": [
"252",
1
],
"length": [
"250",
1
],
"ref_image_size": "max",
"clip": [
"38",
0
],
"vae": [
"36",
0
],
"audio_vae": [
"37",
0
]
},
"class_type": "MiniMaxH3ReferenceToVideo",
"_meta": {
"title": "MiniMax H3 参考转视频"
}
},
"328": {
"inputs": {
"conditioning": [
"265",
0
]
},
"class_type": "ConditioningZeroOut",
"_meta": {
"title": "条件零化"
}
},
"329": {
"inputs": {
"max_resolution": 1024,
"jpeg_quality": 80,
"suppress_default_preview": true,
"preview_frames": 1024,
"preview_fps": 12,
"tiny_vae": "taeh3.safetensors",
"model": [
"333",
0
],
"vae": [
"36",
0
]
},
"class_type": "ModelPreviewOverrideKJ",
"_meta": {
"title": "Model Preview Override"
}
},
"331": {
"inputs": {
"attention": "comfy kitchen attention",
"model": [
"400",
0
]
},
"class_type": "ModelAttentionBackend",
"_meta": {
"title": "模型注意力后端"
}
},
"332": {
"inputs": {
"selection": "sol-attn",
"selection.tau": 1.3,
"start_percent": 0.2,
"end_percent": 1,
"dense_blocks": "",
"min_tokens": 12288,
"extra_tokens": 256,
"sink_conditioning": "exact_kv_and_rows",
"verbose": true,
"model": [
"331",
0
]
},
"class_type": "BlockSparseAttention",
"_meta": {
"title": "模型稀疏注意力"
}
},
"333": {
"inputs": {
"shift_video": 12,
"shift_audio": 3.0000000000000004,
"model": [
"332",
0
]
},
"class_type": "MiniMaxH3SigmaShift",
"_meta": {
"title": "MiniMax H3 Sigma Shift"
}
},
"364": {
"inputs": {
"frame_rate": 24,
"loop_count": 0,
"format": "video/h264-mp4",
"pix_fmt": "yuv420p",
"crf": 19,
"save_metadata": true,
"trim_to_audio": false,
"pingpong": false,
"save_output": true,
"images": [
"253",
0
],
"audio": [
"254",
0
]
},
"class_type": "VHS_VideoCombine",
"_meta": {
"title": "Video Combine 🎥🅥🅗🅢"
}
},
"397": {
"inputs": {
"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
]
},
"class_type": "SelfLiftH3Sampler",
"_meta": {
"title": "SelfLift Progressive Sampler (MiniMax H3)"
}
},
"400": {
"inputs": {
"lora_name": "H3\\MysticXXX_MMH3-V1.safetensors",
"strength_model": 0.45000000000000007,
"model": [
"53",
0
]
},
"class_type": "LoraLoaderModelOnly",
"_meta": {
"title": "LoRA加载器（仅模型）"
}
}
}
}
```
分段差异：S01 15s/362 native frames/seed 2104177691000017260/two refs；S02 8.125s/209 native frames/seed 2104177691000017261/three refs。两段都用24fps、16:9、1.5MP、euler sampler、12 steps、CFG 1，并使用上述共同模型/LoRA配置。S01 H3 音轨为合成音，不是源片原声；未听审。静态参考生成仅保存服装与房间参考图，图像服务模型和seed未记录。S02保留首195帧，与S01首360帧合为555帧/24fps、23.125秒。两段分别保留480000和260000个32kHz立体声原生合成音轨采样；拼接点前80ms仅对出段线性淡出，再AAC编码。没有加速或插帧；未试听音频、未连续观片，解码与抽帧不等于听审或连续观片。动作、切镜和取景存在偏差，这是创意改编，不是精确动作或原声还原。最终555帧成片已完成A/V解码、24fps时间戳及15秒接点前后抽帧检查；AAC容器时长23.125秒，普通解码23.136秒，额外11ms为编码采样边界。
我会留下鞋子那一镜：黑裙和蓝色腰结之外，鞋型也有了交代。
原帖双视频中的第1条，按原时长改编为23.125秒；H3分两段生成后拼接，声音为模型合成。
来源：xitter.cf/LoveUolanda/status/210…｜AI视频改编，完整实际输入见本帖回复。
你会保留这一下鞋履近景，还是一直看人物？
🤖 Made with AI
机位高度：明显高于人物头顶，摄影机位于人物前上方。
俯视角：大约 35°–50° downward angle，不是垂直 90° 俯拍。
水平角度：人物不是正对镜头，而是右前方/左前方的斜侧三分之四视角。
镜头距离：比较近，属于 close-medium / medium-full，人物从头部一直覆盖到大腿甚至膝部。
人物占画面比例很高：人物约占垂直画面的 75–85%。
构图方向：人物身体形成一条明显的左上 → 右下/下方的斜线，不是垂直站构图。
镜头不是广角夸张透视：更接近中长焦的压缩感，所以身体比例自然，不会出现大头小身。
视线：人物低头，不直视镜头，因此高机位会自然产生一种“摄影师从上方捕捉到她”的感觉。
荧｜特写俯视写真 ✨
镜头从上方俯视，捕捉荧安静而灵动的瞬间。
近距离构图，让眼神、表情与细节成为画面的焦点。
一组更有写真感的荧。
#荧 #Lumine #原神 #GenshinImpact #写真 #写真摄影 #俯视写真 #特写 #PortraitPhotography #Photography #Cosplay #CosplayPhotography #蛍 #原神コスプレ #コスプレ写真 #루미네 #원신 #코스프레 #LumineVN #GenshinImpactVN
提示词 2/3｜第一条原视频 S02（实际提交全文；不是来源帖作者提示词）
模式 Ref2VA；请求源段8.125秒，24 fps；seed 2104177691000017261；原生输出209帧（8.708333秒），保留首195帧/8.125秒，原生和保留段均完整A/V解码。起身与转身晚于目标时点，起身时头脚短暂出画；按创意改编保留。实际参考：P1 固定 v6 身份/体型图 SHA-256 9c8b2fac6469251c034def09a08264050b213da321bf9edac48c51a796bef448；P2 S01保留帧359静态姿势/服装/房间参考，exact-f359.png SHA-256 ae010370c3542c87138a0b8e91f00f9e6723e4b9d639f43a8454d06c3966276f；P3 静态服装/房间图 SHA-256 6343f340268dedbe82dce7f479f02380146192823b260f2b2d2e9e56f7498b25。提示词 SHA-256 322ca0f9f92096d1016a3ba0318e44a89f5d324420526a6597cc84d896ec8f94；UTF-8 6580 bytes / 6572 chars。
```text
subject_definitions:
is Fang Zhiwei, the adult fictional woman whose facial structure, body proportions, age impression and long near-centre-parted dark wavy hair are defined solely by . Her black-and-blue costume comes from the visible garments in and .
is the fixed v6 identity/body reference; the clear left portrait defines facial features and the right full-body views define proportions. Do not transfer the office suit, sheet background or static sheet pose.
is the actual selected last retained frame of the preceding generated segment, showing her seated on the screen-right gray couch, knees angled screen-left, ankles crossed, one hand on her lap and the opposite palm braced on the seat. The computer desk and blue-purple monitors are on the left. It provides only static pose, visible costume, room and lighting continuity; it is not a strict first-frame, video or motion-state lock.
is the reviewed static costume-and-room guide. Use its visible long-sleeve opaque high-round-neck black midcalf dress with subtle dark floral decoration, blue waist bow/cuff accents, pale hair flower with blue ribbon, small blue pendant necklace and ordinary black closed shoes. It does not override identity from or the seated starting state in . Unseen garment back construction is a conservative new design continuation, not an archived source fact.
summary:
[reference generation] Generate an 8.125-second landscape performance continuation for the final 195 frames at 24 fps of source clip 1's 23.125-second creative adaptation. Start from the static seated state guided by , visibly rise, perform a small shoulder-to-hip hand gesture and turn to a modest side/three-quarter-back stance. Keep as sole face/body authority. Source video and audio are planning evidence only and are not connected as generator inputs. Use newly synthesized sound. Complete the turn early enough to leave a readable settled ending within the requested 8.125 seconds; no acceleration, loops or frozen filler.
retention_analysis:
(throughout): partially_preserved - preserve the fixed v6 identity and body proportions while transferring the visible black/blue costume and broad studio setting. The source-inspired hand/turn categories are an explicit creative adaptation, not motion transfer.
(identity throughout): fully_preserved - sole face, adult age impression, body proportion and default-hair authority. The blurred sheet faces are not a blur instruction.
(opening static continuity and room): partially_preserved - preserve the seated starting state, couch/desk relationship, visible garment and soft light; allow time to advance through an actual visible rise. No claim of strict pixel-locked first frame or continuous latent state.
(costume throughout): partially_preserved - preserve visible high round neckline, long sleeves, long black skirt, blue accents, hair flower/ribbon, pendant and black shoes. Standing pose and camera from this picture do not govern the opening.
detailed_description:
One continuous landscape shot in the same softly lit gray home computer studio. The gray couch is screen-right, desk with paired monitors and blue-purple practical light screen-left, with wood floor beneath the performer. Keep the already generated background landmarks of stable; do not invent new furniture motion. Use a modest front three-quarter camera slightly above seated eye level. Ease backward just enough during the first second to include the complete head, long dress, feet and floor space needed to stand, then hold the camera steady. No orbit that could substitute for the woman's own turn. Do not crop out shoes during the rise or legs during the turn. No speech, names spoken, captions or text.
[Shot 1,00:00.000–00:02.250] Begin with the seated pose from . She lifts the hand from her lap, uncrosses her ankles and plants both shoes naturally on the wood floor in front of the couch. Her braced palm supports a small forward lean; she shifts her weight over her feet, lifts her hips off the cushion, straightens her knees and comes to a comfortable full standing position. Show the seat contact releasing and the whole rise visibly, without teleporting, cutting away or already starting upright. The long black dress falls normally over her lap and legs, with its blue waist ribbon and sleeves unchanged. Keep her face calm and adult, posture composed.
[Shot 1,00:02.250–00:03.750] Standing close in front of the couch, she brings one relaxed hand near her opposite shoulder, then lets it pass down toward the side of her waist/hip. The other arm stays relaxed. Keep this as one small deliberate gesture, without adding dance choreography or changing outfit/body proportions. Her expression remains warm and composed, with a small natural gaze shift toward the camera.
[Shot 1,00:03.750–00:06.250] She makes a small controlled pivot with her feet to turn her own body away from the front into a modest side/three-quarter-back view. Keep both shoes on the floor and show her shoulders and hips turning together; do not simulate the turn by orbiting the camera. One arm bends lightly near her waist, echoing the source's final side pose. The same closed-back long-sleeve black costume continues around her body, with no new cutout, split or accessory. Preserve the v6 face when she glances back over one shoulder.
[Shot 1,00:06.250–00:08.125] Settle into that three-quarter stance beside the couch, arms relaxed with one elbow gently bent. Hold a readable calm ending with only natural breathing and a small gaze adjustment, without another turn or return to sitting. Keep head, shoulders, dress and shoes visible until the requested endpoint. This is an inserted seated-to-standing bridge followed by source-inspired hand/turn categories within the original overall duration; do not imply exact source-action timing.
overall_soundscape:
Newly synthesized quiet indoor computer-room ambience with subtle cloth and shoe movement. No source audio input, no dialogue or vocals. Preserve a natural quiet sound bed without abrupt new loud events.
non_diegetic_music:
Newly generated subdued instrumental electronic accompaniment with a gentle pulse, broadly compatible with the preceding segment's intended mood. No recognizable melody or vocals and no reuse of the source soundtrack. The audio is independently synthesized; do not imply sample-exact musical continuity.
```
静态服装/房间参考提示词 3/3｜image_gen.imagegen 实际成功生成、同服务接受；不是视频生成提示词。
完整提示词如下，SHA-256 0e466a63f57796a6041d447cdc7b0fbf19fc2b5476c6851ebea54b0cf711749c；UTF-8 2409 bytes。实际请求引用固定 v6 图 SHA-256 9c8b2fac6469251c034def09a08264050b213da321bf9edac48c51a796bef448，transparent_background=false。结果 costume-room-v2.png：1672×941 RGB，SHA-256 6343f340268dedbe82dce7f479f02380146192823b260f2b2d2e9e56f7498b25。具体图像模型/seed 未记录，故留未知；这一步只产出静态参考，没有单独发布。
```text
Create one photorealistic widescreen 16:9 theatrical costume reference photograph of Fang Zhiwei, a fictional adult woman aged 22, standing naturally in a home computer studio. Use the attached v6 sheet as the sole identity and body reference: the large clear left portrait defines her exact facial structure, features, adult age impression and skin tone; the right full-body views define the same shoulder, waist, hip and leg-to-torso proportions. Keep her long near-centre-parted dark brown-black wavy hair. The blurred faces in the small views are sheet styling, not an instruction to blur this photograph. Her 175cm height is a character setting. The gray office suit and sheet pose do not carry into this new scene.
Use this deliberately different costume: a fully opaque black midcalf-length stage dress with long fitted sleeves, a high round neckline, a gently flared skirt and royal-blue piping along the waist and cuffs. Matte black fabric has subtle tonal embroidered decoration; solid fabric covers the complete torso, arms and legs to midcalf. A small blue bow at the waist, a pale flower with a short blue ribbon in her hair, and a small blue pendant preserve the source's broad black/blue accessory palette. She wears ordinary closed black low-heeled shoes. This is a new long-costume adaptation, not a claim about the original garment's length or hidden construction.
Show one relaxed full-body standing pose, head and both shoes entirely visible, arms down naturally and shoulders comfortable. Her expression is warm and composed, with natural hands and the adult face of the supplied v6 portrait. Eye-level front three-quarter photographic view, enough space around her silhouette to read the long dress. This static photograph establishes only costume and room; it does not fix the later video's seated gestures or camera.
The studio has a gray upholstered couch to the right, a computer desk with two softly indistinct monitors to the left/back, a wood floor, a small rectangular studio light on a stand near the desk and subdued blue-purple ambient computer lighting. Soft neutral facial light, realistic cloth, upholstery, anatomy and skin. Balanced environmental composition with her face and whole garment clearly readable. One coherent photograph, one woman, no collage, no text, no watermark. Her identity and body stay exactly governed by v6; only costume, pose and scene change.
```
H3 两段共同工作流节点参数（从各自实际回执 workflow.json 提取；已对照两图，移除段落专属 ref ports、duration/seed、filename_prefix 和实际 prompt 文本）：
```json
{
"shared_actual_workflow_nodes": {
"36": {
"inputs": {
"vae_name": "minimax_h3_video_vae_int8_convrot.safetensors"
},
"class_type": "VAELoader",
"_meta": {
"title": "加载VAE"
}
},
"37": {
"inputs": {
"vae_name": "minimax_h3_audio_vae_fp32.safetensors"
},
"class_type": "VAELoader",
"_meta": {
"title": "加载VAE"
}
},
"38": {
"inputs": {
"clip_name": "h3\\qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors",
"type": "minimax",
"device": "default"
},
"class_type": "CLIPLoader",
"_meta": {
"title": "加载CLIP"
}
},
"53": {
"inputs": {
"lora_name": "H3\\minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors",
"strength_model": 0.8000000000000002,
"model": [
"192",
0
]
},
"class_type": "LoraLoaderModelOnly",
"_meta": {
"title": "LoRA加载器（仅模型）"
}
},
"192": {
"inputs": {
"unet_name": "H3\\minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors",
"weight_dtype": "default"
},
"class_type": "UNETLoader",
"_meta": {
"title": "UNet加载器"
}
},
"250": {
"inputs": {
"expression": "max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17",
"values.a": [
"259",
0
]
},
"class_type": "ComfyMathExpression",
"_meta": {
"title": "数学表达式"
}
},
"252": {
"inputs": {
"aspect_ratio": "16:9 (Widescreen)",
"megapixels": 1.5,
"multiple": 32
},
"class_type": "ResolutionSelector",
"_meta": {
"title": "分辨率选择器"
}
},
"253": {
"inputs": {
"samples": [
"397",
0
],
"vae": [
"36",
0
]
},
"class_type": "VAEDecode",
"_meta": {
"title": "VAE解码"
}
},
"254": {
"inputs": {
"samples": [
"397",
0
],
"vae": [
"37",
0
]
},
"class_type": "VAEDecodeAudio",
"_meta": {
"title": "VAE解码（音频）"
}
},
"255": {
"inputs": {
"sampler_name": "euler"
},
"class_type": "KSamplerSelect",
"_meta": {
"title": "K采样器选择"
}
},
"261": {
"inputs": {
"scheduler": "simple",
"steps": 12,
"denoise": 1,
"model": [
"329",
0
]
},
"class_type": "BasicScheduler",
"_meta": {
"title": "基本调度器"
}
},
"265": {
"inputs": {
"prompt": [
"401",
0
],
"width": [
"252",
0
],
"height": [
"252",
1
],
"length": [
"250",
1
],
"ref_image_size": "max",
"clip": [
"38",
0
],
"vae": [
"36",
0
],
"audio_vae": [
"37",
0
]
},
"class_type": "MiniMaxH3ReferenceToVideo",
"_meta": {
"title": "MiniMax H3 参考转视频"
}
},
"328": {
"inputs": {
"conditioning": [
"265",
0
]
},
"class_type": "ConditioningZeroOut",
"_meta": {
"title": "条件零化"
}
},
"329": {
"inputs": {
"max_resolution": 1024,
"jpeg_quality": 80,
"suppress_default_preview": true,
"preview_frames": 1024,
"preview_fps": 12,
"tiny_vae": "taeh3.safetensors",
"model": [
"333",
0
],
"vae": [
"36",
0
]
},
"class_type": "ModelPreviewOverrideKJ",
"_meta": {
"title": "Model Preview Override"
}
},
"331": {
"inputs": {
"attention": "comfy kitchen attention",
"model": [
"400",
0
]
},
"class_type": "ModelAttentionBackend",
"_meta": {
"title": "模型注意力后端"
}
},
"332": {
"inputs": {
"selection": "sol-attn",
"selection.tau": 1.3,
"start_percent": 0.2,
"end_percent": 1,
"dense_blocks": "",
"min_tokens": 12288,
"extra_tokens": 256,
"sink_conditioning": "exact_kv_and_rows",
"verbose": true,
"model": [
"331",
0
]
},
"class_type": "BlockSparseAttention",
"_meta": {
"title": "模型稀疏注意力"
}
},
"333": {
"inputs": {
"shift_video": 12,
"shift_audio": 3.0000000000000004,
"model": [
"332",
0
]
},
"class_type": "MiniMaxH3SigmaShift",
"_meta": {
"title": "MiniMax H3 Sigma Shift"
}
},
"364": {
"inputs": {
"frame_rate": 24,
"loop_count": 0,
"format": "video/h264-mp4",
"pix_fmt": "yuv420p",
"crf": 19,
"save_metadata": true,
"trim_to_audio": false,
"pingpong": false,
"save_output": true,
"images": [
"253",
0
],
"audio": [
"254",
0
]
},
"class_type": "VHS_VideoCombine",
"_meta": {
"title": "Video Combine 🎥🅥🅗🅢"
}
},
"397": {
"inputs": {
"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
]
},
"class_type": "SelfLiftH3Sampler",
"_meta": {
"title": "SelfLift Progressive Sampler (MiniMax H3)"
}
},
"400": {
"inputs": {
"lora_name": "H3\\MysticXXX_MMH3-V1.safetensors",
"strength_model": 0.45000000000000007,
"model": [
"53",
0
]
},
"class_type": "LoraLoaderModelOnly",
"_meta": {
"title": "LoRA加载器（仅模型）"
}
}
}
}
```
分段差异：S01 15s/362 native frames/seed 2104177691000017260/two refs；S02 8.125s/209 native frames/seed 2104177691000017261/three refs。两段都用24fps、16:9、1.5MP、euler sampler、12 steps、CFG 1，并使用上述共同模型/LoRA配置。S01 H3 音轨为合成音，不是源片原声；未听审。静态参考生成仅保存服装与房间参考图，图像服务模型和seed未记录。S02保留首195帧，与S01首360帧合为555帧/24fps、23.125秒。两段分别保留480000和260000个32kHz立体声原生合成音轨采样；拼接点前80ms仅对出段线性淡出，再AAC编码。没有加速或插帧；未试听音频、未连续观片，解码与抽帧不等于听审或连续观片。动作、切镜和取景存在偏差，这是创意改编，不是精确动作或原声还原。最终555帧成片已完成A/V解码、24fps时间戳及15秒接点前后抽帧检查；AAC容器时长23.125秒，普通解码23.136秒，额外11ms为编码采样边界。