我写眼神动、手不动，H3把持杯手也挪了。 用 @maplejohn01《月汐银砂》的第1张夕照咖啡桌图，做了8秒本地 H3，只跑一版，没另做参考图。 我会留… · @superfang119 · TwKit

@superfang119

我写眼神动、手不动，H3把持杯手也挪了。 用 @maplejohn01《月汐银砂》的第1张夕照咖啡桌图，做了8秒本地 H3，只跑一版，没另做参考图。 我会留下这版的眼神变化：从看远处到看镜头，托腮、交叉坐姿、发丝逆光和海面夕照还能认出来。 但细看动作有两处偏差。我指定3.5—4.5秒眨眼，3秒抽帧就已闭眼；我要求持杯手原位不动，后段手却从杯肚移到了杯脚、桌面附近。杯子仍在桌边，不把这个变化直接写成“松手”或“抬杯”。 拿来做轻动作人像展示，我喜欢这版。要做手与道具位置固定、动作按秒执行的镜头，还得修。这里只做了抽帧和关键帧细看，没逐帧证明杯底接触、吸管连接或准确眨眼次数，音轨也未试听。 原帖有四张静态图，本轮只用了第1张。下面是实际提交的完整 H3 提示词、参数和来源；转眼、眨眼和声音都是我新写的。 Source / image reference: @maplejohn01 https://x.com/maplejohn01/status/2111245813548478933 The source post contains four still fashion images, not a video. Only photo1, the sunset waterfront cafe portrait, was supplied to this H3 run as an unchanged800x1200 medium-image download. The other three images were not model inputs. No imagegen was used in this sample. Ref2VA is an image reference, not an exact first-frame or3D geometry lock. The complete H3 prompt is newly written from the visible image, not the source author's original execution prompt. Eye movement, blink timing, ambience and music are newly designed; no source audio was provided. Local ComfyUI / H3 Ref2VA; one generated sample, no reroll, one image reference. Model: H3\minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors Text/image encoder: h3\qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors LoRA: H3\minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors x 0.80; H3\MysticXXX_MMH3-V1.safetensors x 0.45. SelfLift / Euler / simple, 12 steps, CFG1; lowres_scale0.6, transition_step6, rho0, w_min0.5/w_max1. Upscaler h3_upscaler_lms_v0.1.safetensors. Comfy kitchen attention + sol-attn (tau1.3, start0.2/end1); sigma shift video12/audio3. Negative conditioning: original ConditioningZeroOut. Canvas: 2:3 (Portrait Photo), 1.5MP, original multiple32 -> 1024x1536, 24fps; requested8s, native192frames/8.000000s. Delivery is an unchanged copy of the native H.264/AAC file, with no cut, speed change or interpolation. Audio: newly generated AAC 32000Hz, 2 channels. No source audio was provided; generated audio has not been subjectively listened to. Video VAE: minimax_h3_video_vae_int8_convrot.safetensors Audio VAE: minimax_h3_audio_vae_fp32.safetensors Seed: 211124581354847. AI-generated H3 video. The prompt below is copied from the actual submitted workflow. Full actual H3 prompt: subject_definitions: <Subject 1> is the adult woman in <Picture 1>, with long loose wavy brown hair, pearl drop earrings, a small pearl pendant, the same pearl-toned opaque satin fashion dress and pale blue-grey lace wrap. Preserve her face, seated crossed-knee pose, existing clothing coverage and the portrait framing. One hand supports her cheek at screen right; the other rests around the glass beside the table. She remains the only foreground person. <Subject 2> is the existing stemmed glass in <Picture 1>, with pale golden liquid, ice and a thin dark straw, beside a small lit candle on the round cafe table at frame right. Preserve one glass, its stem, circular base and the visible hand-to-glass relationship. Its base stays on the tabletop, without a lift or sip. <Subject 3> is the sunset waterfront cafe in <Picture 1>: wicker chair and pale cushion, round reflective table, terrace plants and awning at left, distant softly blurred patrons, harbor water, far hillside buildings and a low orange sun at upper right. Preserve the same warm late-day lighting and composition. <Picture 1> is the source post's first still photo, an unchanged 800x1200 image downloaded from the visible page. It is the sole image reference for this new eight-second shot, not a source video, pixel-exact first-frame requirement or 3D geometry lock. summary: [reference generation] Make an eight-second continuous editorial portrait of <Subject 1> seated at the cafe in <Subject 3>. Keep her body pose, crossed knees, cheek-supporting hand and the glass placement of <Subject 2> nearly unchanged. She gently shifts her eyes from the distant sky toward the camera, makes one natural blink, then holds a relaxed gaze. Only loose hair strands, the lace edge, candlelight and distant water have small ambient movement. All timing and sound are newly designed; no video or audio reference is supplied. retention_analysis: <Subject 1> (00:00.000-00:08.000): fully_preserved - retain her adult identity, brown hair, pearl jewelry, pearl satin dress, blue-grey lace wrap, existing coverage and seated pose. Add only a small eye shift, one blink, subtle breathing and gentle wind response, without changing the outfit or body framing. <Subject 2> (00:00.000-00:08.000): fully_preserved - keep the same single glass, golden drink, straw, stem, base, table and candle. Preserve the visible hand position; do not lift, tilt, duplicate, refill or exchange the glass. <Subject 3> (00:00.000-00:08.000): fully_preserved - keep the terrace, chair, plants, distant buildings, horizon and sunset in their approximate positions. Let distant reflections and warm light flicker subtly; no new time of day, new foreground person or replacement waterfront. <Picture 1> (00:00.000-00:08.000): partially_preserved - use the portrait composition and seated interaction as image guidance, retaining the original lower-edge crop near the knees rather than inventing a full-body view. The new motion is not recovered source motion. detailed_description: [Shot 1] (00:00.000-00:08.000), one uninterrupted eye-level seated portrait in a vertical 2:3 frame. Retain the initial framing from loose hair above her head through the crossed-knee area, with the round table, stemmed glass, lit candle and sunset harbor visible. The woman sits relaxed on the pale cushion; her screen-right hand continues supporting her cheek, while the other visible hand remains around the glass. Shoulders, hips, legs and the two arms do not change pose. Keep the original fashion outfit intact and its coverage unchanged throughout. The camera is virtually fixed, without an orbit, zoom into the body, cut, tilt or dramatic push; use only imperceptibly small natural camera settling. Golden rim light outlines the hair, soft amber reflections remain on the table and drink, and the sun stays low at upper right. From 00:00.000 to 00:01.500, hold the reference-like pose with her gaze directed upward toward frame right. A mild sea breeze shifts a few loose brown hair strands toward frame left and moves the free lace edge very slightly. Do not move the cheek-supporting fingers or change the contact point. The glass base remains visibly supported by the round table; the dark straw stays inside the glass. Distant harbor reflections shimmer gently, and the small candle flame flickers without flaring. From 00:01.500 to 00:03.500, she slowly brings her eyes toward the camera, with at most a tiny accompanying chin adjustment. Keep the cheek on the same resting hand, shoulders still and mouth softly closed in a neutral, relaxed expression. This is an eye-direction change, not a head turn away, a pose change or a lean toward the viewer. Her glass-holding hand remains in the same place and the glass does not leave the table. Maintain the same hair length, earrings, pendant, lace pattern and candle position. From 00:03.500 to 00:04.500, perform one short natural blink: the eyelids close and reopen, then the eyes return to the camera. Do not add a wink, mouth movement, speech, drinking gesture or hand adjustment. Keep the face structurally consistent while the pearl earring and a few loose strands move a little with the breeze. From 00:04.500 to 00:06.000, hold the new relaxed gaze into the lens. Hair and the free lace edge settle into small wind-driven movement. The body, crossed legs, cheek-hand contact and glass remain in their reference positions. Keep the glass transparent and structurally coherent, with one bowl, one stem and one base; preserve the straw, ice and liquid rather than turning them into fingers or extra ornaments. From 00:06.000 to 00:08.000, end in the same seated portrait and calm camera-facing gaze, with no additional gesture. Let the distant water and candlelight continue softly. Keep the sun, horizon, terrace architecture, glass base on the table and original lower-frame crop through the final frame. No transition to night, standing up, new foreground person, wardrobe transformation, changed clothing coverage, duplicated hand, extra glass, text, watermark or cut. overall_soundscape: Newly generated quiet waterfront cafe ambience: soft sea breeze, low distant water movement and faint indistinct terrace room tone, with a tiny rustle of lace. No intelligible conversation, vocalization, glass clink synchronized to an invented movement, sipping sound or copied source audio. Keep the ambience continuous for all eight seconds, without abrupt edits or loud impacts. non_diegetic_music: A newly composed very quiet evening instrumental with sparse warm electric-piano notes and a soft acoustic-guitar texture. Slow, unhurried and unobtrusive beneath the cafe atmosphere, easing down gently under the final held gaze. No vocals, recognizable borrowed melody or heavy beat.

曝光 75 · 评论 0 · 点赞 1 · 书签 0