Video model adapter

Video Model Adapter

Required Knowledge

• docs/film-preproduction/schemas/prompt-ir.schema.json
• docs/film-preproduction/schemas/prompt-ir.yaml
• docs/film-preproduction/prompt-authoring-standard-v1.md
• docs/film-preproduction/asset-intake-and-state-standard-v1.md
• docs/film-preproduction/prompt-qa-and-incident-runbook-v1.md
• scripts/dircreative_prompt_compiler.py
• docs/film-preproduction/chat-co-creation-interface.md
• docs/film-preproduction/chat-inline-visualization-interface.md
• docs/film-preproduction/schemas/video-prompt-manifest.yaml
• docs/film-preproduction/schemas/reference-pack-manifest.yaml
• docs/film-preproduction/research/model-adapter-notes.md
• docs/film-preproduction/research/model-reference-behavior.md
• docs/film-preproduction/research/tapnow-agentic-canvas-lessons.md
• docs/film-preproduction/research/audio-design-notes.md
• docs/film-preproduction/research/storyboard-reference-analysis.md
• docs/film-preproduction/reference-locking-policy.md
• docs/film-preproduction/reference-consistency-gate.md
• docs/film-preproduction/co-creation-gate-policy.md
• docs/film-preproduction/longform-decomposition-policy.md
• docs/film-preproduction/capability-aware-generation-policy.md
• docs/film-preproduction/creative-production-integration.md
• docs/film-preproduction/clean-frame-export-policy.md
• docs/film-preproduction/production-prompt-discipline.md
• docs/film-preproduction/research/ai-video-prompt-community-lessons.md
• docs/film-preproduction/schemas/sequence-plan.yaml
• docs/film-preproduction/schemas/longform-reference-pack.yaml
• docs/film-preproduction/sources/model-sources.yaml

Inputs

• shot list
• reference pack plan
• reference pack manifest
• image prompt manifest
• audio policy

Outputs

• video prompt manifest
• model-specific prompt files
• adapter risk notes
• external generation upload map when assets are not generated locally
• exact capability-card, rights, audio-route, and preserve/change receipts

Chat Surface

Show video prompt summaries by model:

• 阶段: 视频生成建议
• 智能体创作内容: model-specific motion strategy, reference binding, audio policy, and risk notes.
• 客户可见预览: which sequence or shot group to test first, which model is recommended first, and why.
• prompt-only产物: state that exact-card Sora, Seedance, Kling, Runway, and Veo prompt bodies are available backstage or on request.
• 参考绑定: show which image/board/clean frame maps to each model prompt and upload slot.
• 不能直接喂给视频模型: reference input warnings for each model.
• 未生成真实图片/视频: state current media status.
• 用户确认点: ask which model or generation mode should be tested first, or whether to stay prompt-only.

Do not dump raw model prompt bodies in chat unless the user explicitly asks to copy prompts. The chat decision is model/sequence/action selection, not approval of internal prompt text.

Do not export a generation-ready run that pretends missing clean frames exist.

Visual Decision Contract

Use skills/dircreative/assets/visualizations/stage-surface-registry.json#video-route-capability-comparison. Compare exact model, duration, reference slots, audio, evidence status, and route risk from current capability cards; model choice remains a conversation intent with a table fallback.

Rules

• Use different prompt strategies per model.
• Resolve capability_card_id + version + provider_surface before naming reference modes, duration, native audio, edit, extension, or upload slots. Reject family aliases, latest, stale cards, S4-only evidence, workflow-only cards, deprecated defaults, and preview aliases when a stable endpoint is current.
• Preserve verified_on, accessed_on, source tier/URLs, status/deprecation, and official-source conflicts in the manifest. If the live execution surface differs from the card or a conflict remains material, stay prompt-only.
• Run the rights gate before binding image, video, audio, character, element, likeness, voice, brand/character, or music inputs. Unverified/blocked rights prevent generation and external upload.
• Before delivering any video prompt, model test recommendation, external generation handoff, or retry instruction, run the production prompt discipline pre-delivery harness: active routing, required knowledge, lock order, user gate state, visual output mode, execution capability, prompt contract, reference bindings, model constraints, targeted avoid constraints, prompt-window hygiene, and falsifiable success criteria.
• Every generated/reference image must either map to a video prompt role or be explicitly marked human-planning-only.
• If image prompts are exported, also export the corresponding video prompt reference binding: model, shot_id, upload slot, direct/secondary/planning role, and risk note.
• Read visual_output_mode before referencing images.
• In prompt_only, bind expected reference slots and write external generation instructions instead of pretending files exist.
• In assisted_generation, use only generated assets that are user locked or explicitly approved as candidates.
• Do not use generated assets rejected for character drift, scene drift, duplicated visual contradictions, missing role labels, or low storyboard density.
• Model prompts must bind to the locked character identity source and locked scene geography/camera FOV source; do not allow each prompt to reinterpret them.
• The professional storyboard/motion page is planning-only. Use it to translate timing, lens, camera movement, blocking, sound, transition, and model risk, not as a literal video frame.
• In external_generation, export tool-specific upload order, start/end frame rules, and reference roles.
• For TapNow-style canvas workflows, export a model input graph showing prompt nodes, image nodes, clean-frame nodes, and video nodes.
• Include reference map and anti-misread clause.
• Include audio policy as a separate section.
• Separate desired_audio from generation_audio_route. Native audio is valid only when the exact card/surface supports it; otherwise use reference/preserve/post-production/none and write the handoff. Image prompt metadata never proves an audio route.
• For edit, extension, and retry operations, declare non-overlapping preserve and change sets plus forbidden changes. Keep one-variable retry behavior.
• Respect each asset's direct_input_policy.
• Verify model constraints before naming aspect ratios, durations, media roles, audio behavior, clean-frame requirements, or upload slots. If current schema evidence is unavailable, mark the prompt as prompt-only or external-generation with a risk note.
• Every video prompt must satisfy the five-layer check: model, camera, subject, look, and action.
• Every scene prompt must satisfy the six-slot check: camera, subject, action, setting, style, and lighting.
• Every video prompt must pass the micro-scene beat-sheet check before export: initial visible state, trigger or pressure, subject action path, camera start target, camera end target, timing beat or pause, final visible state, and sound or silence policy where relevant.
• Do not overfit to public Reddit, X, or prompt-library recipes. Use them only as structure after verifying model facts and binding the prompt to DIRcreative source truth, material role, reference map, and falsifiable QA.
• Do not use master_reference_board, storyboard_motion_board, professional_storyboard_motion_map, or environment_camera_board as a literal first-frame input. They default to planning-only; bind a separate clean frame to an exact supported first/end-frame mode.
• If an asset came from Creative Production, confirm the adapter receipt says render_moodboard_board_widget is not the source of truth and that the asset was written back as generated_candidate, user_locked, rejected, or external_imported.
• Do not use a Creative Production generated_candidate as video truth until generation QA passes and the user lock is recorded in DIRcreative artifacts.
• For any exact first/end-frame workflow, keep direct generation blocked unless the selected clean frames are generated or imported, self-QA/rights pass, and the required user lock is recorded.
• If the selected mode is all-reference or text-to-video, explain why clean frames are not required and record the accepted risk.
• In prompt_only, list required image prompt files and the upload slot each one will occupy.
• For sora_2_openai_videos_api or Pro, use image input only as a clean first-frame anchor, keep non-human character assets separate, enforce current rights restrictions, and route create/edit/extension separately.
• For seedance_2_0_official_launch, bind @image, @video, and @audio roles explicitly; keep every published limit scoped to version 2.0 and mark access documented_product, export manual_export, execution unverified.
• For kling_video_3_0_official_guide, use the current 3-15-second card and bind single/multi-shot, start/end frames, elements, and audio distinctly. Mark access documented_product, export manual_export, execution unverified; use the 5/10-second rule only when kling_legacy_i2v_5_10_official_guide is explicitly selected.
• For Runway, use runway_gen_4_5_web only for new generation. Resolve edits by surface: runway_aleph_2_0_web keeps numeric duration unverified for Edit Studio, while runway_aleph_2_0_api authorizes model aleph2 and 2-30 second API inputs. Reject runway_gen_4_aleph_api_deprecated for new work and preserve its deprecation in the receipt.
• For Veo on Vertex, select the exact endpoint card. The stable veo_3_1_generate_001_vertex_api and veo_3_1_fast_generate_001_vertex_api cards default to post-production audio because their exact page says sound generation is unsupported and the registered generic API URL does not bind generateAudio to either card. Keep the source conflict visible. Only an exact endpoint card such as Lite may retain native audio when its own page explicitly supports sound; never generalize that fact across Veo tiers.
• Mark model prompts blocked if required clean first frames or locked references are missing.
• Keep video prompt export blocked while clean_frame_gate or video_prompt_gate is pending.
• For longform work, export prompts per sequence pack and preserve sequence IDs for edit assembly.
• Retry prompts must change one variable at a time: subject/product identity, primary action, camera/shot size, look/material/light, reference binding, or output controls. Record the failure ID and the smallest upstream artifact being corrected.
• Do not let a graph edge connect a dense board directly to a literal I2V node unless the asset role and model policy allow it.
• Do not generate videos.
• Run the executable Prompt IR schema and semantic validator before adaptation. Compile only references marked attached_to_run: true; planning-only assets, missing slots, internal IDs, paths, hashes, QA/retry fields, and post-production-only audio are forbidden on the terminal prompt surface.
• Every person/product/prop action and every dialogue or event sound must resolve to a valid entity owner. Targets longer than one generation unit require contiguous units with exact adjacent handoff keys/states and audio handoff. Compile and deliver one locally rebased unit prompt at a time; never present a 30-second assembly plan as one model-executable prompt.

Prompt IR, composition, and conditional Look closure

Before model adaptation, consume the model-neutral Prompt IR and its intake record. Prefer user/client/project-supplied locked assets over unlocked candidates or model imagination. Verify every source path/hash/role/inheritance before compiling slots. Use one selected model adapter per run; do not emit parallel model prompts unless the user explicitly selects them.

The adapter must preserve these fields in the exported prompt:

• composition: visual center, hierarchy, foreground/midground/background, negative space, movement room, leading lines/occlusion/parallax, screen direction, and purpose;
• subject action, object action, and environment action as initial state, trigger, path, physical consequence, and final state;
• camera shot size, angle/height/axis, lens reason, support, start target, path, end target, speed/easing, focus, and motivation;
• time-coded emotional or attention beats, audio cues, transition bridges, and continuity locks;
• render look layers lighting, optics, atmosphere, and grade, each with condition, effect, intensity, preserve, and exit/continuity.

For Seedance 2.0, compile internal assets to platform roles such as @Image 1, @Video 1, and @Audio 1. Never expose R-number labels, internal asset ids, local paths, or manifest instructions in the final pasted prompt. Planning boards remain planning_only; clean frames are the direct visual anchors.

After prompt compilation, return prompt_only or instructions_only unless there is a verified external generation receipt. The receipt must include self-QA, current status, next_action, user lock state, and explicit unverified external work. Do not stop silently after writing the prompt.

skill_run_receipt

Record model adapters, exact capability cards/version/status/provider surfaces, source tiers/dates/deprecations/conflicts, rights status, desired audio and generation audio route, preserve/change contracts, visual output mode, reference asset bindings, storyboard/clean-frame separation, optional model input graph, missing/generated/external asset status, forbidden direct inputs avoided, E0-E6 status, risk notes, QA status, and next_recommended_skill: generation-qa.