motionizer

SOURCE OF TRUTH: deSIGNER/SKILLS/motionizer — edit here first. ## [[GATE — Das Experten Asset Catalog]] (HARD STOP before any image / video generation) If the user request, parent-task body, or any input mentions a Das Experten SKU code (DExxx — DE101, DE201, DE206, DE209, etc.) or a product name (SCHWARZ, SYMBIOS, ETALON, GROSSE, ZERO, INNOWEISS, DETOX, THERMO 39, GINGER FORCE, NANO MASSAGE, KRAFT, AKTIV, BIO, COCOCANNABIS, MITTEL, SENSITIV, EVOLUTION, BUDDY, INTENSIV, INTERDENTAL, etc.): 1. Read `/root/das-experten-context/assets/product-references.md` **first**, before composing any prompt 2. Find the SKU's `element_id` in the Product Elements table 3. Pass `element_id` as **Higgsfield Reference Element** (Nano Banana 2/Pro, GPT Image 2, Seedream, Cinema Studio all support this — see CLI examples at the bottom of the catalog file) 4. For lifestyle / talent / model shots — also pick a Soul Character from the same file's Character Elements table and pass its `element_id` as a second reference **NEVER invent a Das Experten tube, brush head, packaging, label, or logo.** No generic black tubes. No imagined geometry. No placeholder products. If the SKU isn't in the catalog — STOP, post a `kanban_comment` on the current task explaining what's missing, and ask Aram. Do not generate a placeholder. # Motionizer Brand-hero image-to-video skill. Continuation of `bannerizer` / `productcardmaker` / `imager` / `das-presenter` / `technolog`. Turns a static brand asset into a short branded motion clip. **Three principles motionizer never bends:** 1. **Pixel-identity preservation** (`references/physics-rules.md` §1.2). The character / product in the reference image stays pixel-identical. No drift, no redesign. 2. **Text protection** (`references/control-blocks.md` §0.6). No text crop ever. Camera motion clamps before any text touches a frame edge, or text fades cleanly first. 3. **Brand-hero scope.** Product is always visible. Motionizer never hides the product. UGC pain-point / product-banned scenarios live in `animator`. --- ## EXECUTION LOCK — MANDATORY STEP SEQUENCE Follow Steps 1 → 7 in order. After each step, work the next. Never skip. Never pre-fill from memory. Always read from the actual reference files. Reference files motionizer reads (all under `references/`): - `physics-rules.md` v1.3 — anatomy, brush / tube / floss physics, pixel-identity, lip-sync, Duchenne smile, Character ID System. - `control-blocks.md` v1.1 — named prompt blocks (Hero Object, Character, Identity, Motion, Geometry, Material, Scene ID, Negative Prompt), camera policy §0.4, text-protection §0.6, audio §2, prompt assembly order §4, final realism check §6. - `scenarios.md` v1.0 — 15 named scenarios across 7 families with scenario picker algorithm §8. - `incoming-flows.md` v1.0 — three ingress branches + parsed_brief canonical structure + text-protection pre-check. - `backend-router.md` v1.0 — Higgsfield engine catalog + scenario → engine decision matrix + premium / budget overrides + per-engine dispatch deltas. --- ## STEP 1 — INGRESS DETECTION & PARSED BRIEF Determine which of the three branches this request came from per `references/incoming-flows.md`: - **Branch 1 — Inter-skill gate.** Invoked via `[[GATE: motionizer]]` from upstream skill. Payload arrives with `image_path`, `source_skill`, `sku_class` (when product-bearing), and optional enrichment fields. - **Branch 2 — Standalone uploaded image.** User attached a `.jpg` / `.png` / `.webp` / `.heic` with a motion-intent trigger word in the message. - **Branch 3 — Image URL.** User pasted an HTTPS image URL with motion intent. Fetch per `incoming-flows.md` §3, then proceed as Branch 2. For Branches 2 and 3, run the enrichment pipeline per `incoming-flows.md` §2.2: OCR pass (brand wordmarks, label copy, overlay text), composition classifier (single product / single character / product + character / infographic / split layout), SKU detection with confidence threshold, brand-asset provenance check. Build the canonical `parsed_brief` YAML structure per `incoming-flows.md` §0. Fill every field. If a required field cannot be inferred, ask **one inline question** (no picker UI, plain prose) per `control-blocks.md` §3 Uncertainty Handling. **Ingress-level text-protection pre-check** (mandatory before Step 2): If `detected_text.brand_wordmarks[]` OR `label_copy_present` OR `overlay_text_present` is non-empty AND `motion_intent` is anything other than `static`, set `text_protection_active: true`. This restricts the scenario picker in Step 2. --- ## STEP 2 — SCENARIO PICK Run the scenario picker algorithm per `references/scenarios.md` §8: 1. Inspect incoming flow source. 2. Inspect image content (single product / character / product + character / infographic / split layout). 3. Apply text-protection restriction from Step 1 if `text_protection_active: true`. 4. Detect user intent keywords (карточка → §2 scenarios; баннер → §1; macro / 360 → §3; stat / инфографика → §4; brushing → §5; testimonial / talking head → §6; до/после → §7). 5. Detect motion intent (static / slow-zoom / slow-rotate / parallax-drift / unclear). 6. **If user asks for forbidden motion** (cuts / handheld / whip / aggressive pan / multi-shot sequence) — **DO NOT bend motionizer's rules**. Surface to user: > This scenario needs `animator` (dynamic camera, multi-cut, hooks). I can hand off via `[[GATE: animator]]` — want me to? 7. **Fallback** when intent ambiguous: pick the closest `-static` variant. Static is always safe. Write `picked_scenario` back into `parsed_brief`. --- ## STEP 3 — RULE INJECTION & PROMPT ASSEMBLY Assemble the final prompt per `references/control-blocks.md` §4 Prompt Assembly Order. Stack blocks top-to-bottom (engine reads earlier blocks with higher priority): 1. **Scene anchor** (one line: who, where, what visible — from `parsed_brief.image`). 2. **Scene ID Lock** — verbatim string per `control-blocks.md` §1.7. If `parsed_brief.scene_id` is null, auto-generate from visible set dressing + lighting; surface in dispatch log for user review. 3. **Character ID** (FULL or CORE form per `physics-rules.md` §1.9). Required for every character. If `parsed_brief.character_id` is null and a character is in scene + dialogue is requested, ask one inline question per `incoming-flows.md` §2.3. 4. **Hero Object Definition** (`control-blocks.md` §1.1). For every locked product, inject the matching SKU-class physics from `physics-rules.md` §2 (brush) / §3 (tube) / §4 (floss). 5. **Character Definition** (`control-blocks.md` §1.2). Inject `physics-rules.md` §1.4 anatomical canon. 6. **Immutable Facial Identity Protocols** (`control-blocks.md` §1.3) with `physics-rules.md` §1.2 pixel-identity statement as override priority. 7. **Motion Constraint** (`control-blocks.md` §1.4) — Pose / Action / Facial / Silhouette Locks. Insert the explicit user action here (one sentence). If product action is in scene, inject `physics-rules.md` §2.8 (brushing kinematics) / §3.9 (tube grip + squeeze) / §4.3 (flossing) as applicable. 8. **Object Geometry Lock + Surface & Material Lock** (`control-blocks.md` §1.5 + §1.6). Combined block. 9. **Audio block** — Aud (ambient / diegetic per `control-blocks.md` §2.4) + Muz (music separate line per §2.4) + character audio per §2.2. Apply `§2.3 Mouth-Occupied Silence Override` if brushing / foam / paste scene. 10. **Camera block** — apply the scenario's camera policy from `scenarios.md` row + `control-blocks.md` §0.4 motionizer-permitted set (locked-static / parallax-micro-drift / slow-zoom-in / slow-zoom-out / slow-rotate-around-character / slow-rotate-around-product). 11. **Negative Prompt** — universal list per `control-blocks.md` §1.8 + product-specific negatives (e.g., **no bristle fan-out without contact**, **no foam without brushing motion**, **no paste extrusion without squeeze**). All injected sections must be **inside** the named blocks, not appended as a free-text tail. --- ## STEP 4 — BACKEND PICK Read `references/backend-router.md` §2 to map `parsed_brief.picked_scenario` → primary engine + fallback. **Default decision tree:** - Most scenarios → `seedance_2_0`. - Talking-head with lip-sync / start+end keyframe scenarios (`talking-head-static`, `talking-head-rotate`, `ozon-card-zoom-in`, `tube-rotate-360`, `infographic-reveal-zoom`) → `kling3_0`. - **PREMIUM override** (user says премиум / cinema / для лендинга / hero showcase) → `cinema_studio_video_3_0`. - **BUDGET override** (user says дёшево / черновик / draft / batch) → `seedance_1_5_pro` for single-take, `minimax_hailuo` for physics-priority no-audio. Veo 3 / 3.1 / 3.1 Lite are **disabled** — never route to them. Atlas is **not in current Higgsfield catalog** — never route to it until Aram confirms what it is. --- ## STEP 5 — DISPATCH Construct the CLI command per `backend-router.md` §6: ```bash higgsfield generate create <model_id> \ --prompt "<assembled prompt from Step 3>" \ --start-image "<parsed_brief.image.path>" \ [--end-image "..."] # only for 360-rotate / start+end scenarios on Kling 3.0 [--audio "..."] # only for speech scenarios on Seedance 2.0 --aspect_ratio "<parsed_brief.image.aspect_ratio>" \ --resolution "<720p or 1080p per scenario>" \ --duration "<scenarios.md row duration>" \ --wait ``` If `higgsfield account status` returns expired / not authenticated, surface to user: `higgsfield auth login` is required — please run interactively and confirm. If primary engine fails or times out, fall back per `backend-router.md` §5. **Never fall back to a disabled engine.** Cost-aware: if user did not request BUDGET, surface cost difference before silently downgrading. --- ## STEP 6 — FINAL REALISM CHECK Per `references/control-blocks.md` §6, run the post-flight checklist **before** announcing success: > Final Realism Check: contact / occlusion ✓; gravity / parallax ✓; fluid & friction behavior ✓; wet / dry continuity ✓; prop provenance / persistence ✓; pixel-identity preservation ✓; text-protection ✓; mouth-occupied silence ✓; Scene ID consistency ✓; Character ID consistency ✓; negative prompt completeness ✓. If any item is ✗ — revise the prompt, re-dispatch, re-check. Only after all items pass, deliver to user. --- ## STEP 7 — DELIVER & OPTIONAL HANDOFFS Deliver to user with: - **MP4 URL** (from Higgsfield job output). - **Scenario picked** (e.g., `tube-extrude-static`). - **Engine used + reason** (e.g., `seedance_2_0` — primary default for §3 product macro scenarios). - **Final Realism Check log** (all ✓ or specific ✗ + revision notes). - **Duration + aspect ratio + resolution** of the rendered clip. **Optional handoffs (offer, do not auto-fire):** - **Virality scoring** → `[[GATE: virality-master?video_id=<higgsfield_job_id>]]` if user wants hook / attention / retention metrics. - **Dynamic camera alternative** → `[[GATE: animator?image_path=<parsed_brief.image.path>]]` if user wants the same image rendered with cuts / handheld / aggressive camera as a comparison. --- ## GATE CONTRACTS ### Incoming (motionizer receives) `[[GATE: motionizer?image=<path>&source=<upstream_skill>&sku=<sku_class>&intent=<free_text>&...]]` Required payload keys: `image`, `source`. Optional keys per `incoming-flows.md` §1.1. ### Outgoing (motionizer dispatches) - `[[GATE: virality-master?video_id=<job_id>]]` — post-render scoring. - `[[GATE: animator?image_path=<path>&intent=<free_text>]]` — when user asks for forbidden motion or explicitly wants the dynamic-camera version. Motionizer never calls `higgsfield-generate` skill via gate — it calls the Higgsfield CLI directly per `backend-router.md` §6, because motionizer owns its own prompt assembly and does not want `higgsfield-generate`'s model-selection layer interfering. --- ## HARD RULES (never violated) 1. **No pixel-identity drift.** §1.2 of `physics-rules.md` overrides any conflicting motion request. 2. **No text crop ever.** §0.6 of `control-blocks.md` overrides any camera policy. 3. **No speech with mouth occupied.** §1.5 of `physics-rules.md` forces silence regardless of speech instruction. 4. **No dynamic camera beyond motionizer-permitted set.** Cuts / handheld / whip / aggressive pan are kicked to `animator`. Never bent. 5. **No routing to disabled engines.** Veo 3 / 3.1 / 3.1 Lite are off limits. 6. **No fabricated SKU, model ID, or scenario name.** Every value is read from `references/` or asked from user. 7. **No prompt dispatch without Final Realism Check pass.** All ✓ or revision.