Seedance 2.5 Full Guide

00:00 That whole video came out of a single prompt run and the sound arrived in that same pass. I never opened an editor and I never cut a frame of it by hand. This is Cense 2.5 and this video is the working manual for it. We start at the very first line of a prompt and finish on a full multi-shot scene. So this is the whole model from start to finish

00:23 with nothing skipped. Before any of the fancy stuff, you need the right mental model for this. Seed dance 2.5 takes four separate kinds of input and each one owns a different job. Your text prompt owns intent, meaning what happens and in what order. Your image references own identity and style, so faces, wardrobe, locations, and color. Your video reference owns motion, and that is

00:46 the R2V lane. Your audio file owns timing because it decides when things happen on screen. The biggest mistake I keep seeing is people stuffing all four jobs into the prompt box. They describe the face in words and then they describe the motion in words, too. After that, they wonder why every run drifts somewhere new. If a detail can live in a reference, it never goes into the

01:10 prompt. I keep the prompt for events and camera work only. And my results got dramatically more stable the day I started doing that. Watch this clip first because the prompt behind it is the whole lesson. I asked for a chef flamating a pan while a cat knocks a wine glass off the counter behind him. A waiter then has to catch that glass before it hits the floor. That is three

01:33 subjects carrying three length actions which is the exact structure that used to break video models. On 2.5, the chain usually plays out in the order I wrote it. And on most of my runs, the catch reads like a genuine reaction. Here is one run where it didn't get there. The chef flamays. The cat goes for the glass and the clip simply runs out before the waiter ever arrives. That is the failure

01:55 mode you will actually hit. Not a broken frame, just a chain that stops one length short. It only lands consistently when you write the prompt as a chain of causes. So, rule one is to bind your actions with the words while and then. The model clearly parses both of them as time and instructions. A flat list of subjects standing in a room gives you a photograph with a little jitter on top.

02:19 A chain gives you an actual scene with a beginning and an end. Rule two costs people more renders than anything else in this video, and it is simple. Give every subject one verb. Three subjects doing one thing each is completely fine. And that chef scene is the proof of it. What breaks is stacking two actions onto the same subject. When I handed my

02:41 waiter a catch and then a stumble on top of it, my success rate collapsed across eight runs. Split that second action into its own clip instead because one extra generation is cheaper than fighting the model. Rule three is about specificity in the right places. Vague adjectives get you nothing, while concrete nouns get you everything. When I asked for a fancy kitchen, the model

03:04 handed me a different generic countertop on every single run. When I named a copper pot on a marble counter instead, those same two objects came back every time. Materials work the same way because naming the material, hence the model its physics for free. I asked for water spilled across a steel counter and it pulled and spread the way water actually does on metal. The reflection

03:27 came with it. So, the room sits inside that puddle and shifts as the surface settles. I wrote a country rap track. You [music] left your booze by the door and you name my truck, man. >> And I wanted a real clip for it. This part is paid for by Top Youu and I built the whole video inside their film studio. Here it is on screen. I never wrote a technical line. The camera and the lens come from a list and so do the

03:51 visual style and the camera move. First, I generated the singer himself right there as stills from a few different angles. That's what keeps him the same guy in every shot later. For the singing shots, I dropped in the audio from my Sunno track. His mouth lands on the beat. Two kinds of shots came out of it. Some of it is plain B-roll, boots in the dust and a neon gas station at 2:00 in

04:16 the morning. Then there are the performance shots where he sings straight down the lens. The edit happens in the same window. So, every shot lands on a timeline inside top view. Here's what came out. [music] You left your booze by the door and your name on my truck bench. Two [music] coffee cups, one cold like

04:38 the blue in your last text. I saw you tail ass. [music and singing] >> 90 seconds of country rap and the cowboy stuff reads modern instead of costume. If you want to build one yourself, my link sits in the top line of the description. The ultra annual plan is the one to look at. It runs Cedance 2.5 at 12 cents a second with 500 credits a

05:03 month. 60 days of unlimited Cedence comes with it as a bonus. That's top view.ai. Here is one prompt with three camera instructions chained into a single continuous shot. I asked for a slow dolly toward a lighthouse at dusk followed by a 90° orbit around it. The last instruction was a rising crane move that ends looking straight down at the

05:25 rocks. The model executes all three as one move, and the parallax during that orbit is genuinely correct. Background rocks slide against the tower exactly the way a real crane shot would render them. Use real cinematography vocabulary because the model was obviously trained on it. I reran that exact lighthouse shot in plain language, asking the camera to move closer and then go around

05:48 the building and rise at the end. What came back is genuinely nice footage and it is still not the shot I asked for. The move around the tower reads as a mushy sideways drift and the rise never lands on that top down look at the rocks. When I wrote the single word orbit instead I got a clean orbit on the first attempt. Plain words get you something watchable while the real terms

06:12 get you the move you actually pictured. The vocabulary is not decoration. It is how you keep control over what comes back. Order matters just as much as vocabulary. The model runs your moves in the sequence you type them. So, put the move your shot depends on first. Three moves per clip is the practical ceiling on 2.5. I stacked the fourth move into one prompt and the last two blended into

06:35 shaky wobble. One more habit that fixed a lot of drift for me. Attach every camera move to a subject instead of to empty space. Orbit around the lighthouse holds steady while orbit right wanders because the model has nothing to rotate around. Every clip you have watched so far was generated on AI master which is my own platform and cadence 2.5 is already running inside it. So here is

07:00 the whole trip on screen. I open the generation panel, pick cense 2.5 from the model list, drag my reference set into the tray and hit generate. The render comes back in the same window where I wrote the prompt at the same tier I would get anywhere else. That barista still came out of the same tab a few minutes earlier because I asked GPT image 2 for the frame with a motion room

07:24 already built in. I never exported a file and I never opened a second tab for any of it. That same window also holds Claude Chad GBT Grog and Gemini. So I am not paying for four separate subscriptions. The per token cost comes out lower than going to each of them directly. Image, video, and audio generation sit in that same place and

07:47 again cheaper than buying every tool on its own. My barista is saved there as a consistent character. So, she comes back looking identical next week without me rebuilding her reference set. You can also publish a character you built and earn from it on the platform and push finished content into the feed so it actually gets reach. AI content engine is the part I lean on hardest because

08:09 [music] AI agents in there write the scripts, the descriptions, and the thumbnails while the channel keeps running. Around 12,000 people are already working inside the platform, and the results page is stacked with their screenshots and real testimonials pulled from demo videos rather than polished marketing copy. There is an academy attached to all of it with over 200 lessons and roughly 30 hours of

08:34 material. And that is where I would start if reference-based generation is new to you. Getting in takes about 2 minutes. You land on the page, pick the annual plan where the discounts live, and you are generating the same day. There is a 7-day money back guarantee sitting on top of that. So, if the workflow does not match how you actually work, you take your money back and walk. The link is in the description. And now,

08:57 let's get back to the tests because the model is about to get pushed a lot harder than a single still image. Most real production starts from an approved frame. So image to video matters more than text to video. Your client signs off on a still and then you have to bring that exact still to life. This barista shot animates cleanly and the movement lands exactly where the still

09:20 left room for it. Her face never turns fully the camera in this one. And that is my source frame doing that, not the model. Watch how much of the result is already decided by the picture I fed in. The credit here belongs to that still as much as it does to seed dance, which is what the next two rules are really about. Compose your still with motion room already built in, meaning empty

09:43 space in the direction you expect movement. I fed the model a cluttered flat lay with no depth and the animation stayed almost frozen because there was nowhere for anything to go. The second rule will save you a pile of rerenders. If your character's face is smaller than roughly a tenth of the frame, identity still drifts on 2.5. I confirmed that across five runs with progressively

10:08 wider shots, so crop tighter whenever the face carries the scene or plan a cutaway right at the drift point. Your prompt on an image to video run should describe change only, not the contents of the picture. The model already sees the frame, so describing it again just competes with what it is looking at. Cance 2.5 gives you three output tiers on the same render panel. Native, 4K,

10:33 1,080p, and 720p, all from that same 30-cond pass. Native means the pixels get rendered at full resolution in that one pass instead of being upscaled afterward. And that is exactly why it costs more. A full 4K render runs noticeably slower than a 1,080p pass on my machine, and it burns more credits

10:57 per generation. So, my rule is simple. I draft everything at 720p or 1,080p while I'm still hunting for composition, and I only spend the 4K render once the shot is actually locked. This is the feature that changes what you can actually make. So, we are spending real time on it. R2V means reference to video and it lets you hand the model a clip of motion instead

11:22 of describing that motion in words. So yes, that is me on a green screen walking forward and settling into a nightstand. I recorded it in one take so the motion stayed natural instead of rehearsed. Then I fed that clip as an R2V reference with a prompt for a medieval knight in full plate armor. The knight walks my exact stride and drops into that same stance right on my beat.

11:46 The model even translated my weight shift into armor movement, so the puldrons settle exactly where my shoulders dropped. Motion is the hardest thing to prompt because language barely describes movement at all. R2V means you stop describing motion and start performing it instead. The night is my own body, but the stronger use of R2V is stealing motion from somebody who moves

12:10 better than you do. I pulled a dancer's clip off YouTube, a tight robot style routine built out of sharp isolations and hard stops, and fed that in as the motion reference with a set of my own face photos for identity. My face lands on his robot moves and every freeze holds for exactly as long as his did down to the little pop in the shoulder

12:33 before he snaps the other way. Identity does break in one spot though. Every photo I gave it was front on. So the frames where he turns his back to the lens had nothing to work from. The model invents a generic profile for those two or three frames. Build your face set from more than one angle and add a three/arter turn whenever your motion reference turns away from camera. Keep

12:56 your reference clip under 15 seconds because sync starts slipping on longer takes. Shoot on a locked off camera since camera movement in your reference fights the camera movement in your prompt. And keep the whole body inside frame because anything cropped out gets invented from scratch. One more thing that tripped me up early. Your prompt should describe the character and the

13:18 environment only and never the motion. If you describe motion in text while also feeding a motion reference, the two instructions compete and you get mush. Left on your screen is a handheld take of me walking down my own street, camera held out at arms length, walking, right? Is that same take with me turned into a Disneystyle cartoon character, soft rounded shapes and big warm eyes, while

13:42 the street behind me stays basically the street I filmed. My walk, my gestures, and my framing all carry over precisely. I fed my own clip in as the reference and wrote the prompt for the character look only. So, I described how I should look and never what I was doing. The model reads your footage for motion and layout, then repaints you on top of it. This is style transfer, not the motion

14:06 control we just did, and the difference decides how you plan the shot. R2V hands the model a performance and lets it invent the frame around that performance. A restyle pass hands it a finished frame and changes only how the frame looks. You keep the composition and the timing you actually shot, which is why this is the fastest way to reuse footage you already have on a drive

14:29 somewhere. My body and my clothes convert reliably, and the cartoon silhouette holds its shape from frame to frame. My face is where it falls apart. Every time I swung the camera toward myself at arms length, the cartoon version of me shifted slightly between shots, and a stylized face gives the model fewer real features to hold on to than a realistic one does. handheld shake survives the conversion too and on

14:52 a cartoon character that reads wrong because animation almost never carries real camera jitter. So those are the two things to inspect on any restyle pass. Your own face at close range and the amount of shake baked into the original take. Two habits fix most of that for me. Attach a face reference set even on a restyle because the model treats your face as one more surface to repaint

15:15 unless you pin it down. Then stabilize your original footage before you feed it in since the model copies your shake faithfully and you cannot pull it back out afterward. And keep the prompt purely about the art style because any action word in there starts fighting the motion already sitting in your footage. Cedense 2.5 accepts up to 50 multimodal references in one generation. That

15:38 covers images, video, audio, and 3D blockouts all in a single pass. So, I loaded the tray until something broke because that number is useless without a working ceiling. I started with eight references, meaning three character shots, two style frames, two background plates, and one audio track. The output respected all eight, and likeness stayed

16:01 locked. At 18, it still held together cleanly. The interesting failures started around 25, and you have to watch closely to catch them. The model never errors out and never crashes. it quietly starts averaging instead. So, two similar character references merged into one blended face that matched neither photo. My style frames stopped acting like distinct influences and collapsed

16:25 into a generic mood. So, my work in ceiling is 15 to 20 references and the count matters far less than the assignment. Give every reference exactly one job, meaning one face source, one style source, and one plate per location. Never feed two references that answer the same question. Two photos of the same actor from different shoots will get averaged into a stranger. So

16:49 pick the better photo and delete the other one. That single habit fixed more identity problems for me than [music] any prompt change did. The native ceiling is 30 seconds. So any longer scene means stitching segments together. That is exactly where models usually start drifting. So I built a deliberate stress case. >> [music] >> Two characters walk through three connected rooms and every room is its own 32nd generation. I fed the last

17:13 frame of segment one back in as the reference for segment two and did the same again for segment 3. On AI master, the preview window has a safe last frame button right under the player. So I never had to pause and screenshot a frame or scrub back through a timeline to [music] grab one. Both characters hold their faces and their wardrobe through the full 90 seconds down to the scar placement on the lid's cheek.

17:37 Lighting is the thing that still slipped a little. Room two rendered slightly warmer than room one. Even though both prompts asked for identical neutral daylight, a casual viewer would never catch it and a colorist would flag it immediately in a final grade. So, the chain workflow that actually held up comes down to three habits. Carry the final frame forward every time. Never a

18:01 middle frame and never your original reference image. Lock your lighting description into every segment word for word instead of assuming the model remembers. Then budget one grading pass across your scenes because you will need it even on a good run. Plan that pass into your timeline from the start and long scenes stop being a gamble. Go back through every clip you have watched

18:23 today and listen instead of looking. the knight's armor and the room tone under the barista shot all arrived in the same pass as the picture. I never opened an audio editor once in this entire video and nothing you heard was layered on afterward. Cance 2.5 writes dialogue, ambience, and effects together with the frames, which is why the sound sits

18:47 inside the scene instead of sitting on top of it. You steer that sound the same way you steer the picture with one plain line in the prompt. Name what you want to hear, the way you would name a prop. So ask for rain hitting metal or a crowd murmuring two rooms away. If a character speaks, write the actual line out and the model times the mouth to it. Leave that line out entirely and you still get

19:12 audio, just whatever ambiance the model thinks the scene needs. And that default is usually generic. The other half of this is the audio lane we set up at the very start of the guide. Attach a music track and it stops being decoration because the model reads it as the timing instruction for the whole clip. Cuts and camera moves start landing on the beat without you typing a single timestamp

19:35 anywhere. Trust that on short clips and check the sync yourself on anything long since the further a segment runs, the more room there is for drift. Native Audio is a very strong first assembly, not a finished mix. Remember that opening shot from the very start of this video? Here is the full reference stack behind it laid out side by side. There is a photo set of my own face shot from

19:59 three angles under flat light. There is a costume reference for the suit and a car reference for the exterior. The driver behind the wheel gets his own separate reference because he and I are two different identities and the model will [music] average them into one stranger if you let it. And there is the music track that every cut in that shot is time to. Now look at what the prompt is not doing there. It never describes

20:24 my face. It never describes the paint on the car and it never names a single beat in the track. All of that already lives in the references. So the prompt only carries events. meaning who moves where and what the camera does while they move. That split is this whole guide compressed into one shot and it is the reason the clip holds together instead of drifting apart halfway through.

20:48 Making a shot like that is genuinely easy now. And that is not a pipeline. There was no crew and no compositor behind it. Just one generation panel and an afternoon of setup. But it looks clean for one reason only, which is that every rule from this video was already built into it. Each identity got its own reference and the motion came from

21:10 footage instead of adjectives. The prompt itself stayed on events and camera exactly the way we set it up at the start. Work that way and the model gives you the shot you pictured on the first or second run. Every test so far handed the model references and told it exactly what to do. So, for the last one, I took all of that away. One prompt, 10 seconds, a few camera angles requested inside the shot and nothing

21:35 else attached. The prompt asks for a street violinist playing on a rainy corner while a small crowd gathers around her. I ran it on Seance 2.5, on 2.0, on Clang 3.0, and on Gemini Omni, and all four are on screen. This is not a benchmark. This is the part where you see how each model thinks when nobody is

21:57 holding his hand. Cedence 2.5 goes for coverage. It gives me a wide established angle, then a closer pass on the bow hand, then a reverse from behind the crowd. The cuts feel like somebody shot the scene with a plan already in their head. 2.0 stays in one place far longer, and it treats the angle change as a slow reframe rather than a hard cut. That

22:21 version is calmer and less directed, and on a mood piece, it is arguably the better read of the same sentence. Cling 3.0 goes theatrical with it. The rain falls heavier than I ask for. The neon bleeds across the lens, and the crowd is styled like a frame from a feature film. Gemini Omni runs the opposite direction and shoots the whole thing [music] almost like documentary footage. The

22:43 camera breathes like it is on someone's shoulder and the color in frame stays close to what a phone would actually capture. Same 20 words of prompt and four completely different films came back. I am not calling any of those results bad because none of them is. What this test actually shows you is each model's default taste and that default is what fills the gap every time

23:07 your prompt goes quiet. Every prompt has gaps in it and the model closes them with its own instincts. So pick the model whose instincts already point where your project is heading. Reach for a seedance when you need directed coverage and clean multi-angle structure and reach for a moodier model when atmosphere matters more than blocking. Choosing that way saves you more