这次我最喜欢最后两秒的回眸，金色头饰和银缎花纹很吃光。 把 @LiuLoky50572 的《古风花蝶》拆成 5 张参考图，交给本地 H3，生成了这段 12 … · @superfang119 · TwKit

@superfang119

这次我最喜欢最后两秒的回眸，金色头饰和银缎花纹很吃光。 把 @LiuLoky50572 的《古风花蝶》拆成 5 张参考图，交给本地 H3，生成了这段 12 秒视频。只跑一版，没有重抽。 抽帧看，蝴蝶、长笛、花间抬手到回眸的顺序都出来了；不足也明显：开场到约 1 秒才看清人物，4.25 秒的前景花朵遮得太满。最后的回眸要到片尾才落定。长笛指法和音画同步未验证，生成音轨尚未主观试听。 如果重剪，我会留下回眸镜头，把开头的虚焦缩短一点。你会选蝴蝶还是回眸做封面？ 下面附完整实际提示词和参数。原作者未公开完整提示词，这份是我按画面重新写的；原片无音轨，本次另加 H3 生成音轨。 Local ComfyUI / H3 Ref2VA, one generated sample, no reroll. Five image references extracted from the source at 1.5 / 3.5 / 5.5 / 9 / 11.5 seconds: butterflies, face close-up, flute, garden, looking back. These are image references, not a first-frame lock; the source video was not a model input. The source clip has no audio stream; this recreation adds newly generated H3 audio. 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, CFG 1; lowres_scale 0.6, transition_step 6, rho 0, w_min 0.5 / w_max 1. Upscaler: h3_upscaler_lms_v0.1.safetensors. Attention: comfy kitchen attention + sol-attn (tau 1.3, start 0.2 / end 1); sigma shift video 12 / audio 3. Negative conditioning: original ConditioningZeroOut. Canvas: 9:16 (Portrait Widescreen), 1.5MP, multiple 32 -> 928x1664, 24fps. Requested 12s, native 294 frames / 12.25s. Delivery uses the first 288 frames and first 12s of generated audio, H.264 CRF19 / AAC stereo 32000Hz, exactly 12.000s; no speed change or interpolation. Video VAE: minimax_h3_video_vae_int8_convrot.safetensors Audio VAE: minimax_h3_audio_vae_fp32.safetensors Seed: 211144472532063. The full prompt below is copied from the actual submitted workflow. Source / visual reference: @LiuLoky50572 https://x.com/LiuLoky50572/status/2111444725320630539 AI-generated H3 recreation; this is a newly written prompt, not the source author's published prompt. Full actual prompt: subject_definitions: <Subject 1> is the adult East Asian woman referenced across <Picture 1> through <Picture 5>: long dark hair arranged into ornate updos with loose face-framing strands, warm brown almond-shaped eyes, delicate red lips, a small red forehead ornament, dangling gold earrings and a calm attentive expression. Keep a recognizable adult face across the different looks. <Subject 2> is the flowing embroidered wardrobe: silver-blue translucent sleeves in <Picture 1> and <Picture 2>, lustrous pale pink sleeves over a blue floral bodice in <Picture 3>, a blue-and-ivory flowered robe in <Picture 4>, and the silver floral high-collared robe with red edging and large gold jeweled hair ornaments in <Picture 5>. These are deliberate wardrobe and hair-arrangement changes between shots, not separate people or a continuous physical costume transformation. <Subject 3> is the luminous floral setting: white and pale pink blossoms, occasional blue foreground flowers, dark wooden architecture, cool blue recesses and strong warm rim light, as seen in the five pictures. Soft petals and foreground bloom occlusions connect the compositions. <Picture 1> is the first portrait composition reference, with the woman's hands raised among blossoms and warm orange-red luminous butterflies near her face. <Picture 2> is the first facial close-up reference, preserving the red forehead ornament, loose hair strands, earrings and softly lit skin. <Picture 3> is the medium portrait reference with a transverse metal flute held at the lips, pale pink sleeves, ornate updo and a flower-filled pavilion. <Picture 4> is the garden composition reference, with the woman framed through pale blossoms, one arm lifted toward a branch and a flowered blue robe. <Picture 5> is the final over-the-shoulder portrait reference, with silver floral fabric, red collar trim, large gold jeweled hair ornaments and a curved fan-like accessory behind the updo. summary: [reference generation] Generate a 12-second vertical floral portrait film using five image references. A luminous butterfly portrait leads into a face close-up, a flute portrait, a downward-looking close-up, a blossom-framed garden pose and a final look back over the shoulder. Preserve the soft cinematic beauty imagery, rich embroidered fabric, adult facial continuity and shallow depth of field. Use petals, foreground flowers and brief light blooms for motivated transitions. Generate a new soundtrack; no video or audio reference is supplied. retention_analysis: <Subject 1> (all shots): fully_preserved - maintain the referenced adult facial character, dark hair, red forehead ornament and elegant gaze while allowing the intentional hairstyles of each look. <Subject 2> (all shots): fully_preserved - follow the wardrobe assigned to each shot, preserving glossy embroidery, translucent fabric and the final red-edged collar. Costume changes occur across transitions only. <Subject 3> (all shots): fully_preserved - retain abundant pale blossoms, warm sparkling backlight, blue shadow accents and layered photographic foreground depth; the garden and pavilion are distinct views. <Picture 1> ([Shot 1] composition): weak_reference - preserve the raised hands, butterfly placement, floral depth and luminous color relationship; do not require an identical first frame. <Picture 2> ([Shot 2] composition): weak_reference - preserve the facial close-up and red ornament, without a pixel-identical crop. <Picture 3> ([Shot 3] composition): weak_reference - preserve the flute, two-hand placement, pale pink sleeves and pavilion framing. <Picture 4> ([Shot 5] composition): weak_reference - preserve the raised arm, framing branches and blue floral robe. <Picture 5> ([Shot 6] composition): weak_reference - preserve the backward glance, silver-red wardrobe and gold ornament silhouette. detailed_description: Use richly detailed cinematic ancient-inspired fashion portrait imagery: reflective silk, embroidered flowers, soft realistic skin, bright warm edge light, cool blue shadows, creamy bokeh and shallow depth of field. The 12-second film is vertical 9:16. Changes of composition are intentional edits with brief flower or light occlusion, not uncontrolled morphing. No captions, logos, extra people or modern objects. [Shot 1] Begin on an abstract soft field of cream, amber and blue bokeh. Over the first 00:00.500, rack focus through foreground white blossoms to reveal <Subject 1> in the silver-blue look of <Subject 2>, following <Picture 1>. She appears in a waist-up three-quarter portrait, face turned slightly toward screen right. Her hands rise gently, palms relaxed, with two warm orange-red luminous butterflies near her fingertips and the sides of her face. Their wings beat slowly while the hands remain separate and legible. The camera makes a small forward drift; sheer sleeves and loose hair move softly. White blossoms and blue flowers create several depth layers, with backlight sparkling on embroidery. Keep the butterfly glow localized rather than washing out her eyes. [Shot 2] At 00:03.000, a pale foreground petal crosses the lens, leading into the close facial composition of <Picture 2>. Show the red forehead ornament, warm brown eyes, soft parted red lips, dangling earrings and loose dark hair strands. She tracks a nearby butterfly with a subtle eye movement and a slight turn of the chin, never an exaggerated smile. Hold a small camera push toward her eyes. A soft warm flare passes at the edge of frame near the end, without replacing her face. [Shot 3] At 00:04.250, transition through a bright blossom to a medium portrait based on <Picture 3>. The woman now wears the pale pink sleeves and blue floral bodice, with the referenced flowered updo. A slender transverse metal flute extends horizontally toward screen left from her lips, supported by both hands. Keep the flute as one solid instrument and preserve clear hand separation. She holds a poised playing posture with only a small, controlled finger adjustment; the camera drifts laterally a little through the flowers. Gold earrings sway subtly. Newly generated gentle flute notes accompany this visible playing pose; there is no speech. [Shot 4] At 00:06.500, cut behind a drifting petal into a close three-quarter view of the same face. Her chin lowers slightly, eyes soften downward and a small highlight catches beside the eye. Preserve the pink-blue look from Shot 3, the red forehead ornament, skin texture and fine hair strands; do not change her identity. A large out-of-focus white blossom traverses the lower foreground, providing a natural transition. The camera is nearly still. [Shot 5] At 00:08.000, reveal the garden pose from <Picture 4>. Pale pink and white flowering branches frame the woman on several planes. She wears the blue-and-ivory floral robe, with one arm raised lightly toward an overhead branch, and looks calmly toward the camera. A restrained sideways camera drift reveals her through a gap in the foreground flowers. Keep the raised arm attached and the other hand near the body; blossoms may obscure parts of the sleeves but not permanently cover the face. Warm sunlight outlines hair and embroidered cloth; petals drift with gentle air movement. [Shot 6] At 00:10.000, use a soft bloom wipe to reach the final over-the-shoulder view based on <Picture 5>. The woman wears the high-collared silver floral robe with red edging and the elaborate gold jeweled updo. With her body turned away, she rotates her head a small amount to look back toward the lens. The curved fan-like accessory stays behind her hair, distinct from the head and shoulder. Gold ornaments and earrings make a delicate small swing and settle. Hold the backward glance through 00:12.000 with a subtle camera approach, warm bokeh behind her and no additional costume change. overall_soundscape: Generate a new quiet floral-garden ambience: delicate air movement, faint fabric movement and restrained jewelry sounds. In Shot 3, newly generated soft transverse-flute notes relate to the visible instrument, then recede across the next transition. No copied source audio, dialogue, singing, applause or abrupt impact sounds. non_diegetic_music: Add a newly generated gentle instrumental bed of sparse guzheng plucks and soft sustained strings, slow and airy, leaving space for the brief visible flute performance. Let the harmony brighten slightly during the blossom reveal, then resolve softly under the final backward glance. No vocals, recognizable borrowed melody or heavy rhythmic beat.

曝光 77 · 评论 0 · 点赞 1 · 书签 1