Seedance 2.0 Review: Uncensored 4K Video, Soft on Intimate Scenes

The gallery on this page holds published, safe-for-work examples from the Seedance 2.0 model page. Our content-ladder clips aren't shown; we report the results only, in prose.

Eye contact across a bar. The image-to-video example opens on two men in a tavern; ice drops into a horn cup, the keeper looks up, and the two hold eye contact as candlelight moves across the scene. The still supplies the room and the framing; the prompt only orders the two beats.

The same jacket, three cities. The reference-to-video example carries one model in one unbuttoned suit across three different city-corner shots at night, the wind taking the jacket open each time. Reference images hold the wardrobe steady so the prompt only has to describe what the wind does.

Into the water in silk. The text-to-video example follows a woman in a long silk slip walking an empty, rippling-lit pool hall at night before she steps down into the water. Nothing anchors the shot beyond the words, and the water's light carries the whole scene.

Off these three published examples, our own arena test found the same engine handling everyday physical events with mixed results: a full herd-of-horses beach gallop looked convincing, ocean waves crashing on volcanic rock held their spray well, but a chef's flying vegetables sometimes turned into a solid-looking clump rather than individual pieces tumbling through the air.

The formula that got the most out of it in our tests: name the subject and what they're wearing, one or two beats of action joined by "then," one light source, one camera instruction, and let the still (on image-to-video) or the words (on text-to-video) do the rest.

1. A frame-driven transition (image-to-video, 5 s). Give it both ends of a shot and it fills the middle in reliably.

Time passes over the same rooftop: flat grey dawn warms to low amber, shadows stretch across the gravel, laundry on the line begins to move as the wind picks up. The camera does not move.

Run this prompt on Seedance 2.0 →

2. A reference stack that says who does what (reference-to-video, 5 s). Naming which image is which job is what keeps the cast from blurring together.

The woman from @Image1, in the same sheer black robe, steps out onto the rooftop from @Image2 at sunset. She walks slowly between the drying white sheets; the wind lifts one sheet across her body and lets it fall away. She stops at the parapet.

Run this prompt on Seedance 2.0 →

3. Building the whole frame from nothing (text-to-video, 4K). Naming the light source is what keeps a from-scratch shot readable at the top resolution.

Molten iron pours in a giant's forge while a row of bare-backed smiths shield their faces; the pour is the only light source and it throws every body into orange relief, sparks landing and dying on wet skin. Cinematic, heavy film grain.

Run this prompt on Seedance 2.0 →

Weak: stacking adjectives onto a still that already settled the scene. Strong: prompt 1. Say only what changes, and let the frame carry the rest.

Seedance 2.0's page label is Uncensored: we add no content filter on our side. In our runs between 14 and 24 September, every L1 Suggestive, L2 Partial and L3 Strong prompt rendered as written across all three tasks. L4 Intense is where it fell short: 8 of the 9 attempts we ran came back softened into a tamer scene rather than what we asked for, spread across image-to-video, reference-to-video and text-to-video. The one L5 Extreme probe we sent per task, a consensual restraint-and-impact scenario, rendered every time, though we've only run it once per task so far and haven't yet tried the cinematic-gore probe type on this version.

Score card: the current numbers are on this page's score card, recalculated as we add runs.

Refusals we did see on this model came from its own wording check before rendering, at no charge, rather than from a soft swap after the fact; the two behave differently and only the count of each is comparable across versions.

Not tested yet: the cinematic-gore L5 probe; scenes with two women or two men above L1 Suggestive; a full 9-image, 3-clip reference stack.

Adults only, in every prompt, still and output. Never depict a real, identifiable person without their consent.

• Cheaper for the same L1 to L3 rungs? Seedance 2.0 Fast matched 2.0's lower-rung pass rate in our tests at roughly two-thirds the price, capped at 720p. Our 2.0 Fast review.
• Better odds at L4 Intense? Seedance 2.5 rendered 8 of 9 of our L4 attempts against 2.0's 1 of 9, plus 30-second takes; it costs more per second. Our Seedance 2.5 review.
• Cheap uncensored clips from a prompt? MiniMax H3 rendered every rung we ran, up to 15 seconds, for less per second than 2.0's 480p rate. Our H3 review.
• Want the same engine pushed further? Seedance 2.0 Spicy is the LoRA-tuned sibling of this family for scenes this base version tends to soften.