MiniMax_H3_Singularity_Prompt_Writing_Specification_Enhanced_EN.md
19.3 KB · 358 lines · markdown Raw
1 # MiniMax H3 Singularity Video Prompt Writing Specification
2
3 Full-Reference Image-to-Video · Enhanced Guide for Action, Camera, VFX, Lighting, Sound, and Continuity
4 Purpose: This document consolidates the Full-Reference prompt structure and the recurring high-value patterns found in the provided MiniMax training-prompt corpus. It is an engineering-oriented writing guide, not a claim about undocumented internal MiniMax rules.
5
6 ## 1\. Core Design Principles
7
8 A strong H3 prompt should explicitly control the following layers:
9
10 * Reference — what visual information is inherited from each reference image.
11 * Composition — framing, shot size, subject placement, foreground/background relationships.
12 * Action — continuous, physically readable action rather than isolated action labels.
13 * Camera — position, movement type, direction, speed, amplitude, and what the camera follows.
14 * Physics / VFX — particles, impact, recoil, debris, cloth/hair response, smoke, light interaction, and other consequences.
15 * Lighting / Materials — light source, direction, exposure, reflections, surface response, atmosphere, and depth.
16 * Audio — ambient sound, synchronized effects, dialogue, and music.
17 * Continuity — identity, spatial direction, object state, damage state, and temporal progression.
18
19 ```text
20 Shot ≈ current composition + subject state + action chain + camera movement + environmental/physical feedback + lighting change + sound/dialogue
21 Action chain ≈ initial state → trigger → primary action → displacement/momentum → contact or reaction → final state
22 Camera chain ≈ camera position → movement direction → speed → amplitude → subject followed → focus/depth-of-field result
23 ```
24
25 ## 2\. Standard Full-Reference Prompt Structure
26
27 ```text
28 id="x..."
29 subject\_definitions:
30 ...
31 summary:
32 ...
33 retention\_analysis:
34 ...
35 detailed\_description:
36 ...
37 overall\_soundscape:
38 ...
39 non\_diegetic\_music:
40 ...
41 ```
42
43 The exact field names above should remain stable when the workflow or application expects this structure.
44
45 ## 3\. subject\_definitions
46
47 Use subject definitions to establish persistent identity, appearance, environment, props, and other reference-grounded elements.
48
49 * Define characters with stable visual identity: face, hair, clothing, body characteristics, and distinctive features.
50 * Define environments with architecture, terrain, major objects, atmosphere, and visual style.
51 * Define important props or VFX elements when they need stable visual identity.
52 * Do not place long time-axis actions here. Put evolving actions in detailed\_description.
53 * If a reference image only defines a character, scene, style, or appearance, connect it to the relevant Subject instead of treating it as an independent video keyframe.
54
55 ## 4\. Reference Images: Critical Distinction
56
57 A reference image is not automatically a video character or a first frame. Distinguish between visual inheritance and temporal anchoring.
58
59 * Use <Subject N> for the actual entity or scene content that should appear in the generated video.
60 * Use <Picture N> as a standalone reference when it genuinely functions as a first frame, keyframe, last frame, or composition anchor.
61 * If a picture only supplies appearance, identity, clothing, environment, or style, place that relationship inside the Subject definition.
62 * State clearly when reference-grounded content becomes active if the reference is intended to affect only part of the sequence.
63
64 ## 5\. summary
65
66 Summarize the main visual action and its progression. A useful summary describes what happens, how it develops, and the resulting state—not merely the plot premise.
67
68 ## 6\. retention\_analysis
69
70 Use retention relationships consistently when the format requires them:
71
72 * fully\_preserved — the referenced element remains substantially unchanged.
73 * partially\_preserved — only selected properties are retained.
74 * attribute\_transfer — selected visual attributes are transferred to another generated element.
75 * weak\_reference — the reference provides loose visual guidance.
76 * newly\_generated — the element is created without meaningful reference inheritance.
77 * For audio: fully\_copy, partially\_copy, reference, weak\_reference.
78
79 ## 7\. detailed\_description — The Core Field
80
81 Organize the description shot-by-shot in playback order. Each shot should answer seven practical questions: what is visible, where the subject is, what state it starts in, what it continuously does, how the camera moves, what visual/physical feedback occurs, and what is heard.
82
83 ```text
84 \[Shot 1] \[shot size / camera position] establishes \[subject position + composition].
85 \[Subject] begins in \[initial state], then \[trigger/action initiation].
86 As \[primary action] continues, \[body movement / displacement / prop movement].
87 The camera \[movement] at \[speed/amplitude], keeping \[subject] in \[composition relationship].
88 At the moment of \[contact / impact / emotional change], \[VFX / physical response / lighting change] occurs.
89 The shot ends with \[final pose / final state], while \[sound/dialogue] continues or resolves.
90 ```
91
92 For later shots, use explicit timestamps such as:
93
94 ```text
95 \[Shot 2] At 00:03.000, ...
96 ```
97
98 Shot 1 has no timestamp in the recommended structure; later shots should use timestamps when temporal precision matters.
99
100 ## 8\. Action Writing: From Labels to Processes
101
102 Avoid isolated verbs such as “walks,” “attacks,” “turns,” or “explodes” when the action needs to be visually reliable. Expand them into a causal sequence.
103
104 ```text
105 Preparation → Trigger → Acceleration → Primary action → Contact → Reaction → Recovery → Final state
106 ```
107
108 **Weak:**
109
110 ```text
111 The warrior attacks the enemy.
112 ```
113
114 **Stronger:**
115
116 ```text
117 The warrior lowers his center of gravity, shifts one foot forward, draws the sword back, then accelerates into a forward slash. The blade cuts across the frame with visible momentum; the opponent recoils from the impact, while loose fabric and dust react to the movement. The warrior completes the follow-through and settles into a guarded stance.
118 ```
119
120 For distant characters, explicitly preserve continuous movement:
121 The two distant characters continue walking throughout the shot at a steady pace; their steps remain continuous even though their small scale makes the movement visually subtle.
122
123 ## 9\. Camera / Cinematography
124
125 Avoid vague phrases such as “dynamic camera” or “cinematic camera.” Specify the camera behavior.
126
127 * Tracking shot — follows a moving subject along a defined path.
128 * Push-in — camera moves toward the subject, increasing visual emphasis.
129 * Pull-back — camera retreats to reveal space or reduce emphasis.
130 * Pan — camera rotates horizontally.
131 * Tilt — camera rotates vertically.
132 * Orbit / arc — camera moves around the subject while maintaining a spatial relationship.
133 * Swoop — camera travels through space with a pronounced curved or diving trajectory.
134 * Whip-pan — rapid directional rotation, useful for transitions or sudden changes.
135 * Dive / plunge — camera moves sharply downward or forward from an elevated viewpoint.
136 * Barrel roll — camera rotates around the lens axis while moving through space.
137 * Handheld — controlled small-scale camera shake suggesting physical presence or instability.
138 * Static / locked-off — camera remains fixed while the action unfolds.
139 A robust camera specification contains five elements:
140 camera position / shot size + movement + direction + speed / amplitude + subject to follow
141 Bind the camera to the action. For example, a side-tracking camera can maintain a character in profile while matching the character's running speed; a push-in can accelerate toward the instant of impact.
142
143 ## 10\. VFX and Physical Feedback
144
145 Do not write only “add cool effects.” Describe the trigger, visual form, motion, and environmental consequence.
146
147 ```text
148 VFX ≈ trigger condition + form + direction/motion + interaction with environment/subjects
149 ```
150
151 * Energy / magic — describe emission point, shape, direction, intensity, travel path, and impact.
152 * Explosion — specify the ignition/impact point, expanding fire/smoke, debris trajectory, pressure wave, and environmental response.
153 * Sparks — specify source, density, direction, brightness, and decay.
154 * Smoke / dust — specify origin, expansion, drift, turbulence, and interaction with light.
155 * Fragments — specify what breaks, fragment direction, scale, and momentum.
156 * Liquid / blood — specify the impact source, spray direction, droplets, and subsequent settling.
157 * Rifts / portals — specify formation, edge behavior, internal motion, emission, and disappearance.
158 * Lightning — specify origin, branching, flash timing, illumination, and contact effect.
159 Always consider physical feedback: hair and clothing react to force, weapons recoil, dust rises from footsteps, ice cracks under pressure, nearby surfaces receive light from explosions, and particles inherit momentum from impacts.
160
161 ## 11\. Lighting, Materials, and Visual Texture
162
163 Replace generic quality language with concrete visual controls.
164
165 * Light source and direction.
166 * Warm/cool or otherwise specified light characteristics when important.
167 * Exposure and contrast changes.
168 * Reflections and wet-surface response.
169 * Volumetric light, haze, smoke, or atmospheric depth.
170 * Depth of field and focus transitions.
171 * Motion blur only when justified by rapid motion.
172 * Material response: skin, metal, fabric, glass, water, ice, stone, etc.
173 Words such as “cinematic,” “epic,” “high quality,” and “dynamic” can support a description, but should never replace observable visual instructions.
174
175 ## 12\. Multi-Shot Timeline and Continuity
176
177 Write shots in playback order and preserve causal continuity.
178
179 * Maintain consistent screen direction unless a deliberate reversal is specified.
180 * Track weapon position, body pose, object ownership, and prop state between shots.
181 * Carry damage, dirt, blood, smoke, broken objects, and other persistent states forward when appropriate.
182 * Maintain consistent character identity and clothing.
183 * Preserve the intended reference-image anchor when a picture is used as a keyframe.
184 * Synchronize sound effects with visible impacts and actions.
185 * Use timestamps for later shots when precise timing matters.
186
187 ## 13\. Character Acting and Emotion
188
189 Convert abstract emotions into observable micro-actions.
190
191 * Eyes: gaze direction, blinking, widening, narrowing, tracking.
192 * Face: eyebrow movement, lip tension, jaw movement, smile/frown changes.
193 * Breathing: chest or shoulder movement, visible recovery after exertion.
194 * Posture: shoulders, spine, head angle, center of gravity.
195 * Hands: grip tension, finger movement, hesitation, release.
196 * Attention: explicitly state what the character looks at or reacts to.
197 Instead of only saying “she looks nervous,” describe the behavior: her gaze shifts toward the doorway, her lips tighten, her breathing becomes shallow, and her fingers repeatedly adjust their grip.
198
199 ## 14\. Dialogue, Soundscape, and Music
200
201 Keep speaker identities stable across shots.
202
203 ```text
204 (S1) speaks: "..."
205 (S2) replies: "..."
206 ```
207
208 * Use stable speaker IDs such as (S1), (S2), etc.
209 * Synchronize dialogue with visible mouth movement when the format requires it.
210 * Place important diegetic effects in the overall\_soundscape.
211 * Use non\_diegetic\_music for background score rather than mixing it with physical scene sounds.
212 * Synchronize footsteps, impacts, weapon clashes, explosions, wind, water, and other effects with visible events.
213
214 ## 15\. Reusable High-Value Templates
215
216 ### 15.1 Distant Continuous Walking
217
218 ```text
219 \[Shot 1] A wide establishing shot shows <Subject 1> and <Subject 2> far in the background.
220 Both subjects continue walking steadily along the path throughout the entire shot, maintaining consistent direction and spacing.
221 The camera performs a slow lateral tracking movement, keeping the distant figures within the wide composition.
222 Their clothing and hair move subtly with the surrounding breeze, while the environment remains visually dominant.
223 ```
224
225 ### 15.2 Wuxia / Xianxia Combat
226
227 ```text
228 \[Shot 1] A medium-wide side angle frames <Subject 1> and <Subject 2> in opposing positions.
229 <Subject 1> shifts into a lower stance, draws the weapon backward, then launches forward with accelerating momentum.
230 The camera tracks laterally with the attack, then arcs around the point of contact.
231 At impact, sparks and dust burst outward, the opponent recoils, clothing and hair react to the force, and loose debris is displaced across the ground.
232 The attacker completes the follow-through and settles into a controlled stance.
233 ```
234
235 ### 15.3 Explosion / Disaster
236
237 ```text
238 \[Shot 1] The camera begins in a stable wide shot as <Subject 1> stands near the center of the environment.
239 A sudden trigger occurs behind the subject, followed by a rapidly expanding explosion.
240 The camera pulls backward while panning to keep the subject and blast inside the frame.
241 Fire, smoke, debris, dust, and fragments expand outward; nearby surfaces are illuminated by the flash and respond to the pressure wave.
242 The shot ends as the initial blast subsides into drifting smoke and falling debris.
243 ```
244
245 ### 15.4 Magic / Energy Effect
246
247 ```text
248 \[Shot 1] <Subject 1> raises one hand and gathers a concentrated energy field around the palm.
249 The energy intensifies from a small glow into a dense rotating mass, with particles spiraling inward before being released forward.
250 The camera pushes toward the hand, then tracks the energy projectile as it travels through the environment.
251 At impact, a shockwave expands outward, nearby dust and loose objects are displaced, and the surrounding surfaces receive a brief burst of colored illumination.
252 ```
253
254 ## 16\. Common Failure Modes
255
256 * Using only generic descriptors such as cinematic, dynamic, epic, or high-definition.
257 * Treating every reference image as an automatic first frame.
258 * Using one-word actions instead of continuous action chains.
259 * Writing “dynamic camera” without movement type, direction, speed, or subject relationship.
260 * Writing “cool VFX” without trigger, form, motion, or physical consequence.
261 * Applying heavy motion blur to every movement.
262 * Packing too many simultaneous actions into one shot.
263 * Redefining the same Subject inconsistently across shots.
264 * Failing to synchronize sound with visible events.
265 * Describing emotions abstractly without observable acting.
266 * Failing to state continuous walking when distant characters are intended to keep walking.
267
268 ## 17\. Final Quality Checklist
269
270 * Is every important reference image clearly assigned a purpose?
271 * Are Subject identities stable?
272 * Does every major action have a beginning, progression, reaction, and ending state?
273 * Is camera movement concrete and connected to the action?
274 * Are VFX tied to physical triggers and consequences?
275 * Are lighting and material changes observable rather than generic?
276 * Are sounds synchronized with visible events?
277 * Are speaker IDs stable?
278 * Is continuity preserved between shots?
279 * Are distant/background actions explicitly described when they must continue?
280 * Have unnecessary generic adjectives been replaced with concrete visual information?
281
282 ## 18\. Copyable Master Template
283
284 ```text
285 id="x..."
286
287 subject\_definitions:
288 <Subject 1>: \[identity / appearance / clothing / distinctive features / reference relationship]
289 <Subject 2>: \[identity / appearance / clothing / distinctive features / reference relationship]
290 <Environment>: \[location / architecture / terrain / atmosphere / lighting]
291 <Prop or VFX>: \[appearance / material / persistent characteristics]
292
293 summary:
294 \[Main action progression and final state]
295
296 retention\_analysis:
297 \[Reference → Subject relationships and preservation level]
298
299 detailed\_description:
300 \[Shot 1] \[shot size / camera position] establishes \[composition].
301 \[Subject] begins in \[initial state], then \[trigger].
302 As \[primary action] continues, \[body / prop / environmental movement].
303 The camera \[movement] toward / along / around \[direction], at \[speed / amplitude], following \[subject].
304 At \[impact / contact / emotional beat], \[VFX / physics / lighting response].
305 The shot ends with \[final state].
306
307 \[Shot 2] At 00:03.000, \[new composition and continuing state].
308 \[Action progression + camera + physical response + sound].
309 \[Continue as needed.]
310
311 overall\_soundscape:
312 \[Ambient environment + synchronized diegetic effects + dialogue]
313
314 non\_diegetic\_music:
315 \[Music style / intensity / progression, if applicable]
316 ```
317
318 ## 19\. Single-Shot High-Quality Example
319
320 ```text
321 \[Shot 1] A low-angle medium-wide tracking shot frames the warrior in the foreground, with the opponent several meters ahead on the same axis.
322 The warrior begins in a compressed defensive stance, shoulders lowered and sword held close to the body. He shifts his rear foot forward, rotates his hips, draws the blade backward, and suddenly accelerates into a forward charge.
323 The camera tracks backward at matching speed, maintaining the warrior in the lower-center of the frame while preserving the opponent in the background.
324 As the warrior closes the distance, the blade swings upward and transitions into a diagonal slash. At the instant of contact, bright sparks burst from the collision point, dust lifts from the ground, and both characters' clothing and hair react to the force.
325 The opponent is pushed backward several steps, briefly loses balance, then regains footing. The warrior completes the follow-through and lowers the blade into a guarded stance.
326 The camera performs a small arc around the contact point and settles into a wider composition as the dust begins to disperse.
327 Footsteps, cloth movement, the metallic impact, and drifting environmental ambience remain synchronized with the visible action.
328 ```
329
330 The example is intentionally concrete: it specifies starting state, action progression, camera behavior, physical response, final state, and synchronized sound instead of relying on generic quality adjectives.
331
332 ## 20\. Training-Corpus-Derived High-Value Patterns
333
334 The following patterns are especially useful because they repeatedly improve action readability and visual specificity in the provided training material:
335
336 * Use explicit camera verbs: track, push in, pull back, pan, tilt, orbit, arc, swoop, dive, whip-pan, barrel roll.
337 * Use action verbs that imply visible mechanics: shift weight, brace, draw back, accelerate, lunge, swing through, recoil, stagger, recover, settle.
338 * Describe environmental response: dust, sparks, fragments, smoke, cloth, hair, reflections, debris, pressure waves.
339 * Connect cause and effect: trigger → movement → impact → reaction.
340 * Use temporal transitions: initially, then, as the action continues, at the moment of impact, afterward, finally.
341 * Keep the camera and action synchronized rather than describing them as unrelated lists.
342 * Use concrete lighting and material behavior instead of generic quality claims.
343 These are corpus-derived writing patterns. They should be treated as practical prompt-engineering guidance rather than undocumented official MiniMax implementation rules.
344
345 ## 21\. Recommended Prompt-Building Workflow
346
347 1. Define the reference relationship first: what is inherited and what is newly generated.
348 2. Lock the persistent Subject and Environment descriptions.
349 3. Write the composition and starting state of each shot.
350 4. Expand the main action into a continuous causal chain.
351 5. Add camera movement that physically follows or emphasizes the action.
352 6. Add VFX and environmental reactions at the exact trigger points.
353 7. Add lighting/material changes that result from the event.
354 8. Add synchronized sound and dialogue.
355 9. Check continuity across the timeline.
356 10. Remove vague filler and replace it with observable details.
357
358