Crafting Video Synthesis Prompts

1. One role per source. Every reference input gets exactly one job (identity, motion, audio). Never let two sections imply the same source controls two things — this is the #1 cause of model confusion.
2. Separate identity from performance explicitly. State plainly: "Reference Image = WHO. Reference Video = HOW." This single framing eliminates most identity-drift failures.
3. Lock, don't suggest. Use directive imperative language ("Maintain exact X", "Do NOT reinterpret Y") rather than descriptive language ("she should look like..."). Video models respond better to hard locks than soft guidance.
4. Enumerate anatomy/motion granularly. Don't say "preserve her face" — list jawline, cheekbones, eye spacing, nose, lips, chin, forehead separately. Don't say "match the motion" — list head/shoulder/arm/hand/torso/hip/leg movement separately. Granular lists reduce the model's room for creative reinterpretation.
5. Pair every lock with its failure modes. After each "maintain X" instruction, list the specific drift/artifact to avoid (e.g., "ZERO jitter, snapping, teleporting, pose drift"). Naming the failure mode is more effective than only stating the desired outcome.
6. End with explicit priority ordering. When instructions inevitably compete (e.g., identity vs. motion video's own face), a numbered priority list tells the model which source wins.
7. Match realism register to the source footage. If reference video is smartphone-shot, explicitly forbid cinema-camera depth-of-field/color grading creep, and vice versa.

Identity Lock — always from the image reference only:

Preserve her exact recognizable facial identity: facial structure, face shape, jawline,
cheek structure, eyes and eye spacing, eyebrows, nose, lips, mouth proportions, chin,
forehead, hairstyle, hair color, skin appearance, overall body shape.
Do NOT reinterpret, average, morph, replace, redesign, or randomly alter her appearance.
ZERO identity drift. ZERO face morphing. ZERO identity substitution from [motion source].

Motion Transfer — always from the video reference only:

Transfer 1:1: [list body parts/movement types relevant to the scene].
Do NOT creatively reinterpret the motion. Do NOT add or remove movements absent from
the reference. Preserve original trajectories and timing. Maintain realistic
biomechanics, joint articulation, weight distribution, inertia, momentum, balance.
ZERO jitter, snapping, teleporting, pose drift, unnatural joint deformation.

Camera Lock:

Match camera position, height, distance, framing, angle, movement type (pan/tilt/
track/push/pull), and subject-to-camera distance from the reference video.
Do NOT introduce unnecessary cinematic movement, zoom, or reframing not present
in the reference.

Lip Sync Lock (only with audio):

Frame-accurate, phoneme-level mouth movement matched to source audio timing.
Match mouth opening/closing, lip compression/rounding, jaw movement, syllable timing.
ZERO delay between voice and lips. ZERO generic/looping talking animation.
Maintain [identity source] facial geometry during all articulation.

Consistency Locks — short, itemized, one block per category (body/hands/hair/clothing). Each block: (1) what to preserve, (2) what physics should govern movement, (3) named failure modes to avoid (e.g., extra fingers, fused fingers, clothing morphing, hairstyle changes).

Reference Separation — a compact summary restating each source's single role, placed near the end as a recap before the priority order:

REFERENCE IMAGE = identity/face/body
REFERENCE VIDEO = motion/pose/timing/camera/performance
SOURCE AUDIO = lip-sync timing

Priority Order — numbered list, most authoritative first:

1. Identity source (face/body)
2. Motion source (pose/timing/camera/performance)
3. Audio source (lip sync)
4. Temporal consistency
5. Physical realism
6. Capture-style realism (lens/camera stock)

Hard Constraints — closing negative checklist, terse bullet list, no elaboration:

NO identity drift. NO face morphing. NO body anatomy drift. NO motion reinterpretation.
NO random gestures. NO temporal flicker. NO unnecessary camera changes.
NO malformed hands. NO extra fingers.

• Write in imperative, declarative sentences. Avoid hedging language ("try to", "ideally").
• Group related constraints under clear section headers (use visual separators like ===) so the model and the reader can scan structure quickly.
• Repeat the identity-vs-motion framing at least twice (once early, once in the Reference Separation recap) — redundancy on the most failure-prone point is good.
• Scale granularity to what's visible in the reference video (e.g., only include finger-articulation detail if hands are visible/active).
• When physical dynamics (hair, clothing, soft-tissue) are relevant, always pair "transfer natural dynamics" with "do not exaggerate beyond what's in the reference" to avoid overshoot.
• Keep the Hard Constraints list short and scannable (bullet negatives, no sentences) — it functions as a final checksum, not new information.

• Don't let the motion-source video imply any identity authority. Never write a sentence that could be read as "look like the person in the video."
• Don't bury the priority order in the middle of the prompt. It must come near the end, after all sections, so it resolves conflicts between everything stated above it.
• Don't use vague umbrella terms ("keep her looking right," "move naturally") without enumerating the specific attributes — models need the itemized list to anchor on.
• Don't request physics/realism intensity beyond the reference. Always cap exaggeration explicitly ("no more than observed in reference") to prevent runaway secondary motion.
• Don't mix capture-style registers (e.g., asking for smartphone realism and shallow cinema depth-of-field simultaneously) without specifying which one wins.
• Don't omit the negative/failure-mode lists. Stating only the desired outcome without naming the common failure (drift, flicker, extra fingers, morphing) is measurably less effective for these models.