How to use Seedance 2 for directed AI video

Use text-to-video when subject identity and composition can be invented. Use a first-frame workflow when an approved image must anchor the opening. Use reference-guided generation when the connected controls accept the required source type and the movement, character, or audio reference contributes information that cannot be described reliably in text alone.

Do not upload every available asset. A reference image may define appearance, a reference video may define movement, and a reference audio file may define timing, but overlapping or conflicting examples can weaken the brief. Name each role in the prompt and remove any file that does not change the expected review decision.

Describe what the subject does with a start state, action, pace, and end state. Then describe the camera independently: static, push in, orbit, pan, handheld follow, or another single readable move. Finally add environmental motion such as wind, reflections, traffic, particles, or background figures only when it supports the primary action.

This separation reduces ambiguity. A prompt that says dynamic cinematic motion does not tell the model whether the subject moves, the camera moves, or both. Observable directions such as the dancer turns once while the camera tracks left create a clearer target and make the completed clip easier to score.

Select the aspect ratio for the intended channel and describe a composition that fits it. Choose a duration long enough to complete the action without padding the prompt with a second scene. Resolution, reference duration, native audio, output format, and other connected options can affect the current quote, so configure them before requesting it.

Enable audio only when the clip benefits from synchronized ambience, speech-like timing, impact, or another audible event. Include concise sound direction in the prompt and review the result with headphones. If the workflow can return a final frame, use it when that still will support continuity into the next shot or downstream image work.

• Choose the destination frame first so movement and negative space fit the actual canvas.
• Give one action enough time to complete instead of using duration to request more scenes.
• Treat audio and final-frame output as production requirements, not default decoration.

Inspect identity, object count, motion continuity, camera path, speed, edge behavior, and audio timing. If a reference was used, compare the opening and key visual details with the source. A successful result can still need editorial review; provider completion means the media was delivered, not that every creative or legal requirement has been approved.

CreateForge persists eligible output to the private Library after the task and storage path settle. Use the stored media for review, record the specific issue and its timestamp, and change the smallest relevant part of the prompt or controls. That approach produces more information than starting over with a completely different brief.

Multimodal input is useful only when the sources do not compete. Before upload, assign each image, video, or audio file a role such as subject identity, opening composition, movement reference, color language, or timing. State those roles in the prompt and rank the properties that must remain stable. A source bundle that silently mixes different people, lenses, lighting, and motion can create contradictions that no amount of descriptive style language will resolve.

After generation, review alignment against the role list. A result may preserve identity while missing the requested camera path, or match motion while drifting from the opening frame. Mark the exact failure and decide whether to revise a source, simplify the prompt, change duration, or switch the mode. This diagnostic step protects useful progress and prevents the common pattern of uploading more references to solve a conflict that was caused by too many references in the first place.

CreateForge reference inputs are temporary generation materials, not a long-term asset library. Keep lawful originals in the team's controlled source system and use the CreateForge Library for eligible successful outputs. Do not depend on a temporary upload reference as the only copy of a customer image, audio track, or production plate.

If a later revision requires the same source, confirm that the file and permission are still current before uploading it again. This check prevents an expired link, outdated logo, revoked consent, or obsolete edit from silently entering a new task. It also keeps the prompt and source roles aligned with the version being reviewed.