这轮我最喜欢金眼特写。收尾却没照单执行：要求低机位全身、角色看向镜头，H3给了裁脚的侧脸。 白辫、黑衣、铜铃和四人伏击都出来了，腾空倒转的姿态也有。皮肤比参… · @superfang119 · TwKit

@superfang119

这轮我最喜欢金眼特写。收尾却没照单执行：要求低机位全身、角色看向镜头，H3给了裁脚的侧脸。 白辫、黑衣、铜铃和四人伏击都出来了，腾空倒转的姿态也有。皮肤比参考更平滑，原图的粗笔触弱了；做角色预告，我会留下近景，收尾构图还要再修。 参考 @itsPixieVerse 的30秒分镜，我拆成2段15秒，用本地H3各生成一次。第二段加了第一段真实末帧作为状态参考，没把右侧静态角色表复制进视频。 画面按首、中、尾和动作节点抽帧复核，未做逐帧连续性验收；生成音轨已完整解码，未做主观听审。下面是两段完整实际提示词和参数。 本次实测参数（两段各生成一次，没有补抽）： 本地 ComfyUI / H3 图片 Ref2VA。A 段3张参考：作者肖像、原片左侧0秒环境、原片左侧29秒造型。B 段前3张相同，第4张是A原生第359帧，作为姿态参考；不是首帧硬锁，也未输入原视频或原声。 模型：H3\minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors 文字/图像编码器：h3\qwen3vl_32b_minimax_h3_ultra_uncensored_heretic_int8_convrot.safetensors LoRA：H3\minimax_h3_dmad_full_dareties_v4_step600_comfy_fro.safetensors × 0.80；H3\MysticXXX_MMH3-V1.safetensors × 0.45。 SelfLift，Euler / simple，12步，CFG 1；lowres_scale 0.6，transition_step 6，rho 0，w_min 0.5 / w_max 1。注意力 comfy kitchen attention + sol-attn（tau 1.3，start 0.2 / end 1）；sigma shift video 12 / audio 3。负面条件沿用原图。 画布：9:16选择器 / 1.5MP，原生928×1664、24fps。每段请求15秒，原生362帧；各取前360帧拼接为30秒/720帧，未变速或补帧。保留模型原生音轨；视频VAE minimax_h3_video_vae_int8_convrot，音频VAE minimax_h3_audio_vae_fp32。 Seed A：211134473952632；Seed B：211134473952633。以下两份提示词均使用各段内部0–15秒时间轴，全文来自实际提交快照。 原作、角色与构图参考：@itsPixieVerse。原帖及作者提示词： https://x.com/itsPixieVerse/status/2111344739526328676 https://x.com/itsPixieVerse/status/2111344749689061616 本视频为AI生成的H3改编样片。 A段完整实际提示词（成片0–15秒）： subject_definitions: <Subject 1> is Shiro-Kane, the adult swordswoman in <Picture 1>, with a platinum-white messy bun and thick braid tied with black ribbons, narrow golden eyes, an angular face, gold hoop earrings, a black choker, a black halter wrap top, an off-shoulder black coat with white feather-plume trim, pleated black hakama trousers, platform sandals, large brass waist bells, and one back-slung katana. <Picture 3> supplies a full-body costume silhouette and a later hero-composition reference; in this first-half segment it guides costume continuity only, not a static pose or shot. <Subject 2> is the wet black-stone bridge, bamboo silhouettes, and pale mist canyon shown in <Picture 2>. It guides the environment, surface, and light, not a frozen first frame. summary: [reference generation] The first 15 seconds of a 30-second vertical painterly action introduction follow <Subject 1> through <Subject 2>: she approaches through shallow water, rings her brass waist bells, begins drawing her katana, faces masked attackers, and ends mid-pivot while ducking one incoming strike. <Picture 1> guides her identity, <Picture 2> guides the setting, and <Picture 3> guides costume continuity; none is a hard first-frame lock. Generate new synchronized sound; no source video or source audio is supplied to the model. retention_analysis: <Subject 1> (appears in [Shot 1] through [Shot 8] and [Shot 10], not the attacker insert in [Shot 9]): fully_preserved - retain her adult identity, white braid and black ribbons, golden eyes, earrings, black outfit, feather trim, brass bells, platform sandals, and single katana through every view. <Subject 2> (appears in [Shot 1] through [Shot 10]): partially_preserved - retain the wet dark bridge, bamboo, pale canyon mist, and warm directional highlights while allowing new camera positions, attackers, and action. detailed_description: The target is a full-frame vertical 9:16 painterly 2.5D action film with matte digital-gouache texture, broad opaque shapes, hard-edged warm ivory-gold light from the upper left, cool slate fill, a black-white-gold palette, and atmospheric chalky fog. Keep one coherent view of the same adult swordswoman and canyon; no glossy CGI, photoreal skin, comic outlines, captions, logos, signatures, split screens, character sheets, or static reference panels. [Shot 1] An extreme-wide tracking shot pushes quickly along <Subject 2>, a wet obsidian bridge framed by black bamboo, toward <Subject 1> walking through thin mirror water at center. Her pale head and feathered collar separate from the dark grove. White petals drift across the foreground as pale canyon mist catches the upper-left warm light. Deep bamboo wind establishes the canyon; light footsteps approach. [Shot 2] At 00:01.500, cut to a low macro tracking view of one black platform sandal planting in shallow water. Its weight creates one readable white painted ripple and a small splash; the trouser hem follows with a delayed swing. Keep the foot attached to the same leg and grounded on the bridge. The sandal splash and soft wet footfall are audible. [Shot 3] At 00:02.700, cut to waist height. A fast short lateral pan catches the large brass bells on her black belt swinging against the pleats. They strike once at approximately 00:03.000 with a crisp metallic chime, then settle. Do not multiply or detach them. [Shot 4] At 00:04.000, cut to an extreme facial close-up. Tilt from the black choker and gold earring to her angular jaw and restrained smirk as loose white strands cross her cheek. Keep the same adult face and matte painted planes. She does not speak; the wind briefly thins around the held face. [Shot 5] At 00:05.500, cut to the katana hilt over her feathered shoulder. Her hand closes securely on the black diamond-wrapped grip and begins a controlled draw. A short metallic release and feather rustle punctuate the move; retain one hand, one grip, and one blade. [Shot 6] At 00:07.000, cut to her narrow golden eyes beneath loose white bangs. Hold the focused gaze; a small eye shift anticipates movement in the bamboo. Wind briefly thins around her held face. [Shot 7] At 00:08.500, cut to a low wide ambush on the bridge. Four dark masked attackers emerge ahead between bamboo trunks carrying blades. The camera rushes forward briefly along the bridge while keeping Shiro-Kane spatially separate from the approaching silhouettes. Foot impacts and blade air-cuts rise. [Shot 8] At 00:10.500, cut to a medium side view. <Subject 1> plants her feet and makes one fast horizontal sword draw. Her shoulder rotation leads the blade; her braid and white feather trim follow. A thin burnished-gold spark arc crosses dark bamboo with a sharp metallic clash. Show no blood or dismemberment. [Shot 9] At 00:12.000, cut to one masked attacker's face. A clean crack travels through the rigid mask; shocked eyes remain visible behind it as gold sparks pass the lens. Keep the face intact. Use one brief dry cracking accent. [Shot 10] At 00:13.200, cut to a low circling medium-wide shot. Shiro-Kane ducks one incoming strike and pivots beneath it, keeping both feet grounded during the turn. The camera makes a fast partial orbit while her braid, coat, and feather trim sweep in one broad delayed arc; sword whooshes cross the wind bed. End at exactly 00:15.000 while the pivot is still resolving, with no completed vault or landing; this moving state is the handoff to the next segment, not a freeze-frame. overall_soundscape: Deep bamboo wind and canyon air continue beneath light wet footsteps, one small splash, cloth and feather movement, and a single clear brass-bell chime around 00:03.000. From the ambush onward, foot impacts, blade air-cuts, one metallic clash, and a brief dry mask crack rise over the wind. There is no dialogue, narration, chanting, or lyric. non_diegetic_music: A restrained original instrumental tension bed uses low sustained strings and sparse deep percussion at a slow pulse under the approach. Add brief stronger accents during the ambush and sword draw, then keep the music unresolved at the 15-second handoff. No vocals, recognizable melody, or triumphal wall of music; keep the brass chime and physical sword sounds audible. B段完整实际提示词（成片15–30秒）： subject_definitions: <Subject 1> is Shiro-Kane, the adult swordswoman in <Picture 1>, with a platinum-white messy bun and thick braid tied with black ribbons, narrow golden eyes, an angular face, gold hoop earrings, a black choker, a black halter wrap top, an off-shoulder black coat with white feather-plume trim, pleated black hakama trousers, platform sandals, large brass waist bells, and one katana. Preserve her painted matte skin and elongated fashion-illustration proportions. <Subject 2> is the wet black-stone bridge, bamboo silhouettes, and pale mist canyon shown in <Picture 2>. It guides the environment, surface, and light rather than fixing every camera view. <Picture 3> is the full-body costume and final hero-composition reference extracted from the source clip. It plans the last shot's upright silhouette and low-angle finish, without demanding a rigid frame match or copying its signature or lettering. <Picture 4> is frame 359 of the delivered segment A, showing <Subject 1> in a low wide crouch with the sword extended behind her and her white braid lifted sideways. It is a loose state and continuity guide for the opening transition of this segment, not a hard keyframe. summary: [reference generation] The second 15 seconds of a 30-second vertical painterly action introduction begin from <Subject 1>'s low stance in <Picture 4>, then show her vault above the bamboo, rotate briefly inverted, descend in one gold-trailing strike, recover, and settle into the low-angle hero finish guided by <Picture 3>. <Picture 1> guides identity and costume, while <Picture 2> guides the canyon setting. The still images are references rather than exact frame locks; no source video or source audio is supplied to the model. retention_analysis: <Subject 1> (appears in [Shot 1] through [Shot 6]): fully_preserved - retain the same adult identity, white braid and black ribbons, golden eyes, earrings, black costume, white feather trim, brass waist bells, platform sandals, and single katana through the pivot, vault, strike, recovery, and hero pose. <Subject 2> (appears in [Shot 1] through [Shot 6]): partially_preserved - retain the wet dark bridge, bamboo silhouettes, pale canyon mist, and warm directional highlights while allowing the aerial camera views and combat aftermath. <Picture 3> ([Shot 6] final hero-composition guide): weak_reference - use its full-body upright silhouette and low-angle editorial finish as broad planning guidance without forcing a static image match. <Picture 4> ([Shot 1] opening state guide): weak_reference - carry the low wide stance, sword position, costume and character continuity forward broadly, without freezing or exactly matching its pixels. detailed_description: The target is a full-frame vertical 9:16 painterly 2.5D action film with matte digital-gouache texture, broad opaque shapes, hard-edged warm ivory-gold light from the upper left, cool slate fill, a black-white-gold palette, and atmospheric chalky fog. Keep the same adult swordswoman and canyon environment from the preceding segment. This is a moving continuation, not a character sheet or static reference panel; no glossy CGI, photoreal skin, captions, logos, signatures, split screens, freeze-frame sheet, or end title. [Shot 1] Begin in a low medium-wide view loosely guided by <Picture 4>. During the opening 00:00.000-00:00.200, Shiro-Kane shifts her weight forward from the same wide crouch, with her single sword still extended behind her. Her braid and white feather trim settle from the preceding turn. Preserve the visible body direction and costume; do not restart the dodge or add a new attacker touching her. Wind and a short blade whoosh carry the transition. [Shot 2] At 00:00.200, cut to a low rising angle. She pushes off near an attacker's shoulder and vaults above the dark bamboo line. Show a clear launch, airborne separation, and continuous body trajectory. The camera tilts up to follow her as the coat opens in the airflow; one controlled foot impact and a rising sword whoosh carry the movement. [Shot 3] At 00:02.000, cut to a tight airborne medium shot against bright misty sky. Slow the motion as she rotates briefly inverted, her gold earrings catching the light. The same katana stays securely in her hands, and her composed expression, braid, and cloth trail the rotation. Keep the body continuous and the blade in a single stable grip; the wind bed thins during the held inverted moment. [Shot 4] At 00:04.000, cut to an extreme-wide descending view over <Subject 2>. She drops toward the bridge with one two-handed vertical strike and a narrow gold trail. A stylized parted curtain of pale fog reveals the dark stone landing below. Show her visibly land by 00:06.500 with one low impact; no blood, gore, or body fragments. [Shot 5] At 00:06.500, cut to a medium aftermath view. Dark opponents retreat or fall out of focus behind her. Shiro-Kane regains an upright stance, flicks water from the blade once, and guides the sword back toward its scabbard. The camera dollies inward slowly while the feather trim settles and the black-white-gold costume geometry becomes clear again. Keep the sword as one intact prop and the action controlled, with no gore. [Shot 6] At 00:10.500, cut to a low-angle full-body hero composition broadly guided by <Picture 3>, not an exact frame match. Hold until 00:15.000: her katana is safely sheathed, her chin lifted, her golden eyes looking down toward the lens, and her brass bells resting at her waist as sparse white feathers drift around her. Finish with one clean sheath click, then let the wind and score fade into a quiet final hold. Do not turn the finish into a freeze-frame sheet or add an end title. overall_soundscape: Deep bamboo wind and canyon air continue through the aerial movement. Blade whooshes accompany the vault and descending strike, followed by one low stone landing impact, retreating footfalls, a single blade-water flick, soft fabric and feather movement, and one final scabbard click. No dialogue, narration, chanting, or lyrics. non_diegetic_music: Continue a restrained original instrumental tension bed with low sustained strings and sparse deep percussion. Keep a slow pulse under the opening pivot and vault, use a brief stronger accent at the landing, then thin the score during the aftermath and fade it completely before the final sheath click. No vocals, recognizable melody, or triumphal wall of music.

曝光 82 · 评论 2 · 点赞 4 · 书签 1