Genjutsu vs Seedance: generate from text or rebuild real footage

The core difference in one sentence

Seedance is built around generating footage: you describe a scene in text - or start from an image - and the model invents the shot. Genjutsu is built around recasting footage: you bring a clip you already shot, and its motion, camera and timing anchor everything while references rebuild who and what appears.

That single difference decides most comparisons. One tool answers "what if this shot existed?", the other answers "how do I change this shot I already have?".

Head to head

• Starting point - Seedance: a text prompt, optionally with a style or first-frame image. Genjutsu: a source clip plus 3 to 6 reference images per element.
• Motion - Seedance: invented by the model from your description. Genjutsu: kept from your footage, frame-accurate.
• Best for - Seedance: scenes nobody could shoot, concept shots and b-roll that does not exist yet. Genjutsu: footage that exists and needs a new cast, wardrobe, location or product.
• Consistency - Genjutsu: one reference set keeps a cast consistent across a whole series. Seedance: output varies with each prompt and seed, so continuity takes deliberate effort.
• Fidelity to a real performance - Genjutsu: the original take is the anchor, so choreography and timing survive. Seedance: a described approximation of the action, not a specific take.
• Length and pipeline - check both vendors for current clip limits, pricing and licensing, since they change frequently.

When Seedance is the better pick

• The shot does not exist and never could - imagined landscapes, impossible camera moves, concept films.
• You have no footage and no plan to shoot any; everything starts from an idea.
• You want a fast stack of b-roll for a montage where exact continuity does not matter.
• You are exploring looks before committing to a real shoot.

When Genjutsu is the better pick

• A clip you own is approved except for who, where or what is in it.
• The performance must survive untouched - a dance, a demo, a punchline delivered on camera.
• A product, label or logo has to stay accurate inside the shot.
• One reference set needs to keep a cast consistent across an entire series.
• The same ad needs local faces and streets for different markets, with the edit unchanged.

Can you use both?

Yes, and production teams increasingly do. Seedance-style generation is useful for previsualisation and for cutaways that were never filmed; Genjutsu recasts the takes that actually matter. In a music video, for example, you might generate impossible backdrops and recast the performer into them - the two workflows feed each other.

How to test both in an afternoon

• Pick the smallest real example of your use case - and start from your own footage if the job is footage-driven.
• Run the same creative brief through each tool: same references where possible, same single change.
• Score three things: identity stability across the clip, artifacts on hands and edges, and whether the original motion survived untouched.
• Decide per project, not per brand. The tool that wins a dream sequence is not the tool that wins a product recast.

A note on model names

Model generations move fast, and versions get renamed, merged or retired. Judge these tools by the workflow they support - generate from nothing, or rebuild something real - rather than by benchmark charts from last quarter. Check each vendor for current specs, pricing and terms on the day you decide.