这张樱花汉服，纱衣动得好看，眨眼却晚了。 我更喜欢这版的春风感：发梢向右飘，青金薄纱有透光层次，手还倚在花枝上。用 @Sirilee2026 的原横图直接做… · @superfang119 · TwKit

@superfang119

这张樱花汉服，纱衣动得好看，眨眼却晚了。 我更喜欢这版的春风感：发梢向右飘，青金薄纱有透光层次，手还倚在花枝上。用 @Sirilee2026 的原横图直接做8秒H3，没先重绘，也没重抽。 我写的是3.5–4.5秒眨眼，那一秒的密集采样里仍睁着眼；5秒和5.5秒的点才看到闭眼。笑容比计划更早出现，纱摆和笑意也比“轻微”更明显。画面我愿意拿来做春日人像短片，精确卡表情时点还不能直接交差。 这次看了22个时间点和6张原尺寸帧，没把抽帧当逐帧的手部接触或眨眼次数验证；音轨未试听。此前的四模型图帖保留，这次是原横图的新动画。 完整实际H3提示词、参数和来源在下面。作者公开了静态图词，本次动作、时间轴和声音是新写的。 Source / image reference: @Sirilee2026 https://x.com/Sirilee2026/status/2111467266709782775 Author's still-image prompt: https://x.com/Sirilee2026/status/2111467267158691906 The source is one still landscape AI portrait, not a video. The author provides a still-image prompt including Negative Prompt, but the source image model and sampling settings are unknown. That image prompt was not executed in this H3 run. The unchanged1200x671 medium-image download is the only H3 reference. No imagegen was used, and none of the four-model images from our earlier comparison were H3 inputs. Ref2VA is not an exact first-frame or3D geometry lock. The complete H3 prompt is newly written from the visible image; the smile/blink timing, hair/gauze motion, ambience and music are new. No source video/audio was supplied. 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: 16:9 (Widescreen), 1.5MP, original multiple32 -> 1664x928, 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: 211146726670978. AI-generated H3 video. The prompt below is copied from the actual submitted workflow. Full actual H3 prompt: subject_definitions: <Subject 1> is the single adult Chinese woman, age 25, in <Picture 1>. Preserve her recognizable face, dark eyes, red lips, glossy black coiled hairstyle with loose strands flowing right, pastel silk flowers, dangling pearl hair ornaments, pearl necklace and bracelet. She wears the same pastel cyan-and-gold hanfu, with translucent outer gauze over the existing inner garment. Keep the original neckline and clothing coverage. She leans diagonally toward screen-left, looks into the camera, and rests her visible hands around the diagonal cherry branch. This is a fashion portrait, not an undressing scene. <Subject 2> is the dark cherry branch crossing from the left edge toward the lower middle of <Picture 1>, beneath and between the woman's visible resting hands. Preserve the branch's main diagonal, bark, adjoining blossom clusters, and the existing hand-to-branch relationship. The branch is an anchored prop, not a moving arm or a second person. <Subject 3> is the surrounding white cherry blossom canopy and foreground clusters against the clear soft-blue sky in <Picture 1>. The upper background has crossing dark branches with white blossoms; the nearer left/lower clusters remain softly out of focus. Preserve this shallow-depth-of-field spring setting without adding a new building, path or landscape. <Picture 1> is the unchanged 1200x671 landscape source image and the only image reference for [Shot 1]. It anchors the woman's upper-body framing, diagonal branch, cyan-and-gold gauze, cherry blossoms and bright spring palette. Use it as an appearance and composition reference, not a guaranteed pixel-exact first frame. No source video or audio is provided. summary: [reference generation] Create one continuous 8-second landscape spring hanfu portrait using <Picture 1>. Keep <Subject 1> leaning on <Subject 2> amid <Subject 3>, with a locked camera and the original close portrait framing. Add gentle motion to the existing loose hair strands and loose outer gauze, a slight closed-mouth smile, and one natural blink. Keep the main pose, hands, branch and clothing coverage stable. All motion, ambience and music are newly written for this video rather than recovered from the static source or from the earlier four-model image comparison. retention_analysis: <Subject 1> (appears in [Shot 1] throughout): partially_preserved - retain identity, hairstyle, accessories, cyan/gold costume, upper-body pose, camera-directed gaze and existing clothing coverage; newly animate loose hair, gauze edges, a restrained smile and one blink. No age change or costume change occurs during the shot. <Subject 2> (appears in [Shot 1] throughout): fully_preserved - retain the anchored diagonal branch and the original hand contact arrangement. The hands do not release, slide along, squeeze, bend or pass through the branch. <Subject 3> (appears in [Shot 1] throughout): partially_preserved - retain the white blossom canopy, branch layout, blue sky, natural daylight and shallow depth of field; add only slight movement to existing petal edges and fine nearby twigs, with no large swaying tree or artificial shower of petals. <Picture 1> ([Shot 1] composition and appearance anchor): partially_preserved - keep the original landscape crop and visual relationships while introducing restrained new temporal motion. The source is a still image, not an original animation or a soundtrack reference. detailed_description: The video is a luminous, realistic-looking AI fashion portrait with soft spring daylight, fine silk and gauze texture, delicately separated white blossoms, and natural-looking facial shading. Preserve the source's soft blue/cyan/gold palette without turning the scene into an illustration, heavy glamour filter or glossy plastic render. [Shot 1] A locked landscape 16:9 close portrait establishes <Subject 1> leaning toward the left, her face near the upper-middle-left and her eyes already looking toward the lens. Her coiled black hair is decorated with pastel flowers and dangling pearls. <Subject 2> crosses diagonally beneath her resting hands from the left toward the lower middle. Her overlapping hand and wrist arrangement stays in the same place on that branch. Cyan outer gauze and pale golden fabric fill the lower and right portions of the frame. <Subject 3> surrounds her with white cherry blossoms, crossing upper branches and gaps of soft-blue sky. The near left/lower blossoms retain their soft focus. <Picture 1> supplies the visible identity, clothing, pose, branch geometry, framing and lighting. From 00:00.000 to 00:02.000, establish the same calm pose and direct gaze. Her torso and head do not shift toward the camera, and both visible hands remain resting against the branch. A gentle breeze carries only the already-loose hair strands toward screen-right. A small wave passes through the free outer gauze at the far right, while the garment stays attached at the shoulder and wrist. Pearl ornaments have only minute motion; they keep their shape and do not merge with the hair. The main branch remains anchored. From 00:02.000 to 00:03.500, add a very slight, closed-mouth smile, keeping the face direction and eye line toward the lens. The lips remain together, without speaking, teeth exposure or a large change of expression. The head does not nod or turn. Existing fine hair strands and the loose cyan-and-gold cloth edges continue one slow breeze-driven wave. Do not move the hands, lift fingers, grow new fingers or change the branch into a hand. Preserve the original neckline and all clothing coverage. From 00:03.500 to 00:04.500, she makes one brief natural blink, closing and reopening her eyelids, then resumes the same camera-directed gaze. Keep the head, hands and upper body in place during the blink. The white blossoms remain separate from her hair and face. Do not time a new gesture or head movement to the blink, and do not add an exaggerated wink or smile. From 00:04.500 to 00:06.000, let the smile settle into a soft, relaxed closed-mouth expression. Keep the original leaning pose and the fingers resting on the same bark area. The loose outer gauze briefly billows at the right edge and eases back, without sliding off, exposing new body areas, becoming a solid sheet or wrapping around a new object. Existing blossom edges and fine twigs move only slightly; the thick branches do not drift across the face. From 00:06.000 to 00:08.000, hold the same portrait and gaze. Hair ends and loose cloth edges continue a small natural movement, then grow gentler toward the end. Preserve the pale blue sky, blossom density, background branch layout and soft focus. End within the same shot without fading or changing scenes. No camera push-in, zoom, pan, orbit, reframing or rack focus occurs. Do not add people, birds, text, logos, extra limbs, a falling-petal storm or new props. Maintain the same face and clothing; the scene must not become a different woman or a different hanfu outfit. overall_soundscape: Newly generated quiet outdoor air and a very light silk rustle support the spring breeze. A few soft, distant garden-bird chirps may sit far behind the scene, with no bird added visually. Keep the ambience consistent across the eight seconds. There is no source soundtrack to copy, and the woman never speaks, sings or audibly breathes into the microphone. No impact, hand-contact sound or dramatic wind gust is added. non_diegetic_music: A quiet original instrumental bed of sparse plucked zither notes and a soft sustained airy tone, slow and unhurried with no strong beat or crescendo. Keep it unobtrusive beneath the natural ambience. No vocals or recognizable borrowed melody.

曝光 12 · 评论 0 · 点赞 0 · 书签 0