Text-to-video or image-to-video? Pick the input mode once

Text-to-video or image-to-video? Start with whether you already have an approved still. No still → text-only (text-to-video) to explore framing. Yes → first-frame (image-to-video) to lock looks, then add motion. Need a known start/end or cross-shot identity → first+last frames or multimodal reference. Picking the wrong input mode wastes more retries than picking the wrong model name.

One table to choose

Your starting point	Prefer	What the prompt should do	Common failure
Text idea only	Text-only (t2v)	Subject, action, scene, one camera move, duration	Style words with no verb
Locked product / character still	First frame (i2v)	How it moves and how the camera moves	Re-describing looks that fight the still
Known open and close frames	First + last	Transition and rhythm in between	Huge aspect mismatch; motion maxed out
Same identity across shots	Multimodal reference (r2v)	Clear refs; change one variable per retry	Messy refs; crowded foregrounds

Rule of thumb: explore with text; match a signed look with image or reference.

Text-to-video: when and how

Best for narrative hooks, mood shots, pre-wardrobe concepts, and storyboard drafts.

Write the five parts: subject → action → scene → camera → duration use. One primary camera move in v1; keep duration aligned with the parameter—see camera moves and duration.

Image-to-video: when and how

Best for product openers, post-wardrobe character motion, signed packaging stills, and poster animation.

Spend words on motion (“slow orbit, steam rising, gentle push-in”), not on re-stating color and shape. Keep the subject sharp and the background quiet; identity craft lives in reference-frame consistency.

First+last and multimodal: do not confuse with first frame

• First+last pins both ends of the timeline—ideal for known transitions. When ratios differ, many pipelines crop to the first frame.
• Multimodal reference answers “who / which style,” not a single time origin. Name the primary ref in the first sentence.
• Do not mix conflicting first-frame contracts with a pile of unrelated refs in one job—pick one input contract.

A low-retry sequence

1. Decide: approved still? cross-shot identity?
2. Choose the mode, then write the prompt (the mode tells you what to delete).
3. Prove a short duration before raising camera drama or resolution.
4. For faces and packs, prefer still → image-to-video over endless text re-rolls.
5. Before paid media, skim the commercial checklist.

Takeaway

Text-to-video invents; image-to-video animates a locked look; first+last and multimodal reference handle transitions and identity. Choose the input mode first, then pick a family by job in Seedance vs Wan vs Kling—and you will retry less.