对了，需要先提取首帧，这样h3出视频稳一些。

小muse的形象还挺可爱的。

完整资源➡️：genvizu.com/gallery/continuo…

ComfyUI H3 提示语： subject_definitions: <Subject 1>: The same young adult woman from the supplied references: light skin, long loose dark brown-to-black hair with wispy bangs, soft rounded features and dark eyes; wearing a fitted ribbed lavender short-sleeve scoop-neck top and pale cream ruffled shorts. Her performance moves from open playful warmth to bashful embarrassment and a gentle final smile. Source(s): <Picture 1>, <Picture 2>. summary: A vertically framed candid friend-camera moment: a young woman on a rug realizes she has been noticed, grows embarrassed, hides behind a fluffy patterned pillow, then peeks out and ends with a shy sideways look back, closely following the supplied nine-panel pose sequence. Hard constraints: Begin on the <Picture 1> composition: the same woman seated on the cream shag rug in the warm daylight living room, close to the high wide-angle camera with her hand reaching toward it. | Strictly retain the woman’s reference identity, long dark hair with wispy bangs, lavender fitted top, pale ruffled shorts, and cozy neutral living-room environment. | Use <Picture 2> as the required pose and emotional progression reference: playful openness shifts into being caught, shy embarrassment, hiding with the patterned fluffy pillow, then a cautious peek and relaxed sideways look back. | Exactly one continuous unbroken shot; no cuts, transitions, or montage panels. | Maintain a candid friend-filmed, intimate observer feeling rather than a polished staged performance. | No narration, subtitles, or spoken dialogue. Generation controls (authoritative) — Prompt-text rule: every section label and descriptive sentence is silent production direction, never speech content. Do not speak, recite, chant, sing, or read this prompt aloud. Background music: enabled. Generate a context-appropriate non-diegetic score unless the user explicitly requests silence or no music. Dialogue permission is enabled, but this plan contains no dialogue lines. Generate no character speech, conversation, chanting, singing, or other vocals. Dialogue language: automatic. Preserve each explicit <d> language tag; otherwise use the language required by the scene and user intent. Narration: disabled. Generate no narrator, voice-over, spoken description, or audible reading of prompt text, section labels, instructions, asset tags, or metadata. Subtitles: disabled. Do not render subtitles or dialogue captions. Creative direction (authoritative within explicit user intent and reference fidelity) Priority: Apply these creative directives clearly wherever they do not conflict with explicit user requirements, identity retention, or fixed first/last-frame composition. Format and direction: Build the clip around precise tactile interaction, close material detail, slow readable hand motion, and tightly synchronized diegetic micro-sounds. Avoid chaotic motion and keep any score subordinate to taps, folds, pours, brushes, clicks, or texture sounds requested by the scene. Keep the camera emotionally close and responsive to small gestures while avoiding intrusive or unmotivated movement. Camera execution: Follow the action continuously in a single unbroken take, using only motivated pans, small translations, and reframing needed to preserve spatial clarity. Look: Treat connected visual references as authoritative for identity, wardrobe, objects, environment, composition, lighting, color, and medium; do not impose an unrelated restyle. Pacing and composition: Use moderate, readable motion with natural acceleration and enough settling time for each important beat. Use exactly one continuous shot with no hidden or explicit cuts; express all action beats through blocking, performance, camera motion, reframing, and synchronized sound. An explicit user-requested shot or cut count remains authoritative over this density preference. Compose for a vertical frame: protect faces, hands, and essential action inside the central mobile-safe area; favor depth movement and avoid losing subjects at narrow side edges. retention_analysis: <Picture 1> -> first_frame: Use <Picture 1> as the authoritative first-frame reference for the woman's identity, lavender ribbed scoop-neck top, cream ruffled shorts, long dark hair with wispy bangs, cream shag rug, warm daylight living room, close high wide-angle friend-camera perspective, and her initial smile with a hand reaching toward the lens.. <Picture 2> -> composition: Use <Picture 2> as the authoritative visual storyboard and appearance reference for the same woman, the cozy sofa/rug setting, and the ordered bashful performance beats: open-hand reach, face-covering embarrassment, hiding behind and hugging the fluffy cream pillow with irregular dark patches, peeking out, turning inward, and ending reclined sideways with a small look back toward camera.. detailed_description: [Shot 1] At 00:00.000-00:15.083, At the start, <Subject 1> sits centered on the thick cream rug in the bright cozy living room, smiling gently and reaching one hand toward the close camera exactly in the spirit of <Picture 1>. She then notices the friend filming her and her expression changes to shy embarrassment. She withdraws her reaching hand, raises both hands to cover her lower face, and curls slightly inward. She brings the fluffy cream pillow with irregular dark patches up to hide behind it, hugs it closely, briefly buries her face, and lets her shoulders relax. After a pause, she peeks cautiously over the pillow toward the camera, turns inward with the pillow held against her, then reclines sideways and glances back with a small, bashful smile, matching the final emotional beat of <Picture 2>. Her performance is natural, quiet, and readable, with each gesture allowed to settle. Camera: One continuous handheld friend-camera take beginning in the close, slightly elevated wide-angle composition of <Picture 1>. The camera holds steady for the opening reach, then makes a gentle motivated backward and rightward drift as she pulls away, preserving her face and hands in the central vertical safe area. It performs small responsive pans and reframes to keep the pillow, her hidden face, and her later peek visible, then settles into a close three-quarter view for the final sideways look back. Natural handheld micro-movement only; no zoom, cut, or abrupt reframing. Motion: 0.000-2.800: hold the initial forward-reaching pose, then her hand lowers as she realizes she is being watched. 2.800-5.300: she grows visibly self-conscious and covers her face with both hands. 5.300-9.300: she lifts the patterned fluffy pillow, hides behind it, and hugs it close with a small inward curl. 9.300-11.800: she peeks over and around the pillow toward the camera. 11.800-15.083333: she turns inward, then reclines sideways while holding the pillow and lands on a quiet shy look back. Synchronized audio: Soft room ambience from a quiet apartment.; Light rug fibers and clothing rustle as she shifts toward the camera.; A small hand movement through the air near the lens.; Soft pillow lift, hug, and cushion-compression sounds during the hiding gesture.; Subtle fabric rustle as she peeks out and reclines sideways. overall_soundscape: Quiet sunlit apartment room tone with close textile rustle from the rug, clothing, and fluffy pillow; gentle hand movement and soft cushion compression are synchronized to the bashful gestures. No speech or narration. No narration, voice-over, spoken description, or audible prompt reading. non_diegetic_music: A very soft, warm, unobtrusive ambient instrumental bed enters under the room tone and remains subordinate to the tactile diegetic sounds; no vocals, lyrics, narration, or prominent beat.