A frame from the 720p clip in this test: a fictional cliffside suite at dawn, generated from a paragraph of text.
Luxury hotels sell something that is hard to film: a feeling of calm. The best resort films are slow. The camera glides rather than moves, the light is soft, and the details carry the story: the weight of a linen sheet, the veining in a marble basin, the stillness of a pool at the end of the day. Producing that footage is expensive: a director, a camera team, a stylist, models, several days of dawn and dusk light, and rooms closed while it happens.
AI video models now claim to produce "cinematic" footage from a written description. For a hotel brand team, the interesting question is not whether those clips are impressive in general, but whether they meet a specific standard: does this look like us, and is it good enough to show a guest? We put that question to Wan 3.0, a video model from Alibaba's Wan family, and judged the results the way a luxury brand manager would.
For the test we invented a resort. Casa Alvera is a fictional 22-suite hotel on a Mediterranean cliff, with a brand book that asks for "quiet, tactile, unhurried" imagery: natural linen, pale stone, water, candlelight, no crowds, nothing loud. Casa Alvera does not exist and no real hotel was involved. The prompts, timings, costs and every flaw described below are real, and all of the clips can be seen in the walkthrough video at the end.
We wanted to know what a brand team could learn cheaply, so we used Wan 3.0's lowest settings for most tests: 480p resolution and five-second clips. That is far below what a hotel would publish, but it is enough to judge composition, movement and mood, which is what a team needs at the planning stage. We allowed one higher-quality clip at 720p, eight seconds long, to see how much detail improves. We planned nine tests in advance and generated each scene once.
Wan 3.0 is used through an API, a service that software connects to. For the technically curious, this is one request:
import time, requests
BASE = "https://api.apimart.ai/v1"
H = {"Authorization": "Bearer YOUR_KEY"}
job = {
"model": "wan3.0-video",
"prompt": SHOT_DESCRIPTION,
"resolution": "480P", # 480P / 720P / 1080P
"duration": 5, # 2 to 30 seconds
"size": "16:9",
"audio": True, # sound is generated with the picture
"seed": 11,
# for image-to-video: "image_urls": [STILL_URL], "generation_type": "frame"
}
task_id = requests.post(f"{BASE}/videos/generations", headers=H, json=job).json()["data"][0]["task_id"]
while (d := requests.get(f"{BASE}/tasks/{task_id}", headers=H).json()["data"])["status"] not in ("completed", "failed"):
time.sleep(6)
print(d["cost"], d["result"])
We submitted all ten clips at once. The 480p clips took between 2 minutes 38 seconds and 3 minutes 49 seconds each and cost 14.8 cents apiece. The 720p clip took just under four minutes and cost 47 cents. The whole test, including one still image, cost $1.81.
The signature shot of almost every resort film is a slow move through a room towards a view. It is also a good stress test, because AI video often wobbles, warps or speeds up when the camera moves.
Luxury cliffside resort suite at dawn. A very slow, perfectly steady dolly
forward past a low bed dressed in natural washed linen toward open terrace doors;
sheer white linen curtains breathe softly in a sea breeze; pale travertine floor,
a marble side table with olive branches; the sea beyond in soft pink light.
Quiet, minimal, cinematic, shallow depth of field, no people. Ambient sound of
distant waves only, no music.
(480P, 5 s, seed 11)
Test 1: four moments from the five-second dolly. The move is steady, with no jumps.
This was a strong start. The camera moves forward and slightly to the side past the bed, the sheer curtains lift and settle, and the pink dawn sky sits beyond a glass balustrade. The bedding has the rumpled, heavy look of washed linen rather than the shiny plastic finish that gives away a lot of computer-generated interiors. To check the steadiness objectively, we measured how much the picture changes from one frame to the next. The change was small and even, with no sudden jumps, which is what a smooth, slow move should look like.
Would Casa Alvera publish it? Not as is. At 480p it is soft, and the room is a beautiful but generic "luxury suite" rather than any particular hotel's architecture. But as a reference for a director, "this is the pace and the light we want", it is excellent, and it took less than three minutes.
Luxury brands rely on close-ups that make you want to touch something. We asked for a macro shot with a slow change of focus from linen to water drops on marble.
Extreme close-up macro sequence feel in one shot: a slow rack focus across folds
of natural washed linen on a bed, then onto water droplets beading on a polished
white Calacatta marble vanity, soft morning side light, very slow camera slide.
Tactile, premium, editorial. No people, no text. Quiet room tone.
(480P, 5 s, seed 12)
Test 2: the first and last frames. The linen weave and the beaded water both read clearly even at 480p.
This was the most convincing clip of the test. The linen shows individual slubs in the weave, and as the camera slides, a white marble counter with grey veining and a brass tap comes into view, covered in beaded water drops that catch the light. The change of focus from fabric to stone happens smoothly. A brand team could drop this into a social teaser between real shots and few viewers would notice. Close-ups of materials contain no faces, hands or text, the three things AI video most often gets wrong.
Pools and the sea appear in nearly every resort film, and water is notoriously difficult to generate convincingly. We asked for a slow sideways tracking shot along an infinity pool.
Test 3: the pool at the start and end of the clip (top) barely changes. The hammam (bottom) from Test 5 handled steam and light better.
The water itself was lovely: gentle ripples on the surface, light patterns moving on the pale stone floor of the pool, and a clean infinity edge meeting a flat blue sea. But the camera did not track. The first and last frames are almost identical, and our frame-to-frame measurement for this clip was the lowest of all ten, about a third of the dolly in Test 1. We asked for movement and got a locked-off shot with rippling water. A camera instruction is a request, not a command; if the move matters, plan to generate it more than once.
Food is central to luxury travel marketing, and the classic shot is the chef's final touch. It is also a hard test, because it involves hands, a small tool and a precise action.
Fine-dining kitchen pass, close-up: a chef's hands use long tweezers to place
three tiny edible flowers and a few drops of green herb oil onto a scallop crudo
arranged on a matte dark stone plate. Precise, slow, deliberate movements, soft
top light, shallow depth of field. Quiet kitchen ambience.
(480P, 5 s, seed 14)
Test 4: the chef places the flowers with tweezers, then adds oil from a dropper. We asked for three flowers and got four.
The hands are the best we have seen in a budget AI clip: natural fingers and knuckles, and the tweezers held the way a chef holds them. The scallop slices are fanned neatly on a dark slate plate with a rough edge, and the soft top light is exactly right for food. In the last second, the chef switches to a small glass dropper to add drops of green oil, a sensible interpretation of the prompt.
The flaws are the kind that a head chef would spot immediately. We asked for three flowers and the finished plate has four, two yellow and two orange. And the dish is the model's idea of a scallop crudo, not the restaurant's recipe; for a resort that markets a named chef, generated food is a mood board at best.
Low light is where cheap cameras and cheap AI models both tend to fall apart, with noise, flicker or muddy shadows. We generated two atmospheric interiors and exteriors: a private dinner on a cliff terrace by candlelight, and a marble hammam with steam and shafts of light.
Test 5: the night terrace at the start and end of the slow push-in.
The night terrace is clean and atmospheric: a linen-covered table for two, tall candles, glass lanterns along a stone wall, a starry sky and a dark sea with a lit headland. The candle flames flicker gently. Our frame check flagged one abrupt jump in an otherwise steady clip, small enough that most viewers would not notice it at full speed. The bigger problem is aesthetic. There are dozens of lanterns, more than any real terrace would light for two guests, and the scene tips from "intimate" towards "stock photo of romance". A luxury brand would ask for fewer candles and more darkness.
The hammam was the more surprising success. Star-shaped skylights in a domed ceiling, steam drifting through the light, and a thin stream of water running from a brass tap into a marble basin. The camera tilts up slowly as asked. It feels calm and expensive, and like the texture shots, it contains nothing that can easily go wrong.
Most hotels already own a library of approved photography. One of the most practical uses for a model like this is to add a few seconds of movement to a still, for a website header or a social post. We gave Wan 3.0 a golden-hour suite image as the first frame and asked for a slow push towards the terrace doors with the curtains moving.
Test 6: the source image (left), the first frame (middle) and the last frame (right). The camera swung sideways instead of moving towards the doors.
The start matched the source closely. A standard image similarity score between our still and the first frame was 0.87 out of 1, so the bed, the linen throw, the marble side table and the sunset all carried over. The light on the sea shimmers, and the curtains move. But instead of moving towards the terrace, the camera swung to the side, and a curtain that was at the edge of the original frame swept into the shot. By the last frame the similarity to the source had fallen to 0.35. The image had not been distorted; it was simply a different view of the same room.
For a brand team this is an important lesson. Animating approved photography is the safest way to stay on brand, because the room, the materials and the colors are real. But the camera move needs to be checked against the brief, and anything new that comes into the frame, such as a curtain or a piece of furniture the photographer never shot, has to be approved like any other image.
All ten clips in this test, from the text-only scenes to this animated still, were made with the same Wan video API request, changing only the description, the optional starting image and a few settings. That makes it practical to build a whole shot list in an afternoon, before anyone books a flight.
Resort films usually follow a guest through a day. With no casting and no real person, can the model keep the "same" guest across two separate shots? We used an identical description of the woman, "in her thirties with dark hair in a low bun, wearing a flowing white linen dress and gold hoop earrings", in two different scenes: walking along the terrace at golden hour, and pouring orange juice at breakfast.
Test 7: same description, two shots. The hair and earrings match; the dress does not.
Each shot on its own is good. The walk is slow and graceful, barefoot on warm stone, and the breakfast scene handles a difficult action: juice pours from the carafe and the glass fills. But they are not quite the same person. The hair, the hoops and the general look match, while the dress changes from a strappy sundress in the first shot to a wrap dress with sleeves in the second, and the faces are similar rather than identical. The breakfast table also gained a plate of croissants that we did not ask for, which a food stylist might like but a brand book might not.
There is also an ethical point that matters more for luxury brands than for most. Guests buy an experience partly on trust. A video that shows a person who does not exist, enjoying a breakfast that the resort may not serve in that way, should be treated as concept art. If it is ever published, it should be clearly identified as an illustration, and most platforms now ask advertisers to label realistic synthetic people.
Wan 3.0 generates sound along with the picture, and sound is part of the luxury experience. We asked for specific ambience in each clip, such as "distant waves only, no music" or "only the sound of trickling water, no music, no voices".
Test 8: loudness and movement for every clip. The quietest and loudest soundtracks differed by almost 30 LUFS.
We ran every soundtrack through speech recognition to check for unwanted voices. None of the ten contained recognizable speech. Interestingly, the transcription software "heard" phrases like "Thank you for watching" in eight of the ten clips, in English and in other languages. That is a known quirk of the software, which invents video sign-offs when it hears ambient noise, and not something in the clips themselves.
The bigger issue was level. Measured with the broadcast loudness standard (LUFS), the soundtracks ranged from -43.4 for the texture close-up, which is almost silent, to -13.7 for the animated still, which is about the level of a social media video. That is a difference of nearly 30 LUFS. Played back to back, some clips would seem silent and others loud. For any published film, a sound designer would replace or rebalance all of this anyway, and in luxury work that is usually the case: the soundtrack is composed or licensed. The generated sound is a useful sketch of the mood, not a final mix.
Finally, we spent 47 cents on the dolly prompt from Test 1 at 720p and eight seconds, with the same seed, to see how much quality improves at the next step up.
Test 9: the 480p clip (left) and the 720p clip (right). Same prompt and seed, different composition, clearly more detail.
The first thing to know is that the same seed did not produce the same shot. The 720p version is a new composition: the bed is angled differently, and two wooden loungers have appeared on the terrace. So a team cannot draft a scene at 480p and then "upgrade" it; the higher resolution is a new roll. The second thing is that the extra detail matters a lot for this kind of brand. In the 720p clip you can see the weave of the linen throw and the pitted surface of the travertine floor, and the curtains have fine, believable folds as they billow. At 480p those details are suggested rather than shown. For a luxury brand, whose whole visual language is texture, 720p is the minimum for anything a guest sees, and a real campaign would expect 4K from a camera.
Judged against Casa Alvera's imaginary brand book, the clips failed in predictable places. They show a beautiful, generic resort rather than a specific one: the architecture, the view, the furniture and the food are invented. They include details that nobody approved, an extra flower, a plate of croissants, a crowd of lanterns. They do not reliably follow camera directions. People change between shots. And even the best of them stop at 720p in our test, below what a premium brand publishes.
Above all, luxury marketing promises a real place. A guest who books a suite because of a video expects to find that suite. AI footage that misrepresents the property is a reputational risk, and in some markets an advertising compliance one.
The same clips are genuinely useful earlier in the process. For mood boards, five-second clips at 15 cents each communicate pace, light and texture far better than still references. For pre-visualization, a director and a brand team can agree on a shot list, camera moves and the balance of interiors, food and landscape before the crew is booked, which saves expensive shooting days. For social teasers, material close-ups such as linen, marble, water and steam were close to publishable, especially when mixed with real footage. And for animating approved stills, the property stays real while gaining movement, provided someone checks the camera move.
The video below shows all ten clips with commentary on what met the brief and what did not.
Walkthrough: ten Wan 3.0 clips judged against a luxury resort brief (about 3 minutes).
Shot
Settings
Generation time
Cost
Suite dolly (Test 1)
480p, 5 s
2:38
$0.148
Linen and marble macro (Test 2)
2:58
Infinity pool (Test 3)
3:37
Plating (Test 4)
Night terrace and hammam (Test 5)
480p, 2 × 5 s
2:50–3:10
$0.296
Still image plus animation (Test 6)
$0.151
Guest in two scenes (Test 7)
2:38–3:49
Suite dolly, 720p (Test 9)
720p, 8 s
3:56
$0.473
Total
10 clips, 53 s
$1.81
Ten clips and 53 seconds of footage cost $1.81. It does not replace a production budget, but it does make it cheap to test ideas before committing one.
Wan 3.0 did not produce a luxury resort film, and we did not expect it to. What it produced, for $1.81 and about four minutes of waiting per clip, was a convincing set of references: a steady dawn dolly, tactile linen and marble close-ups, a serene hammam, a candlelit terrace and a chef's hands that look like hands. Its failures, unfollowed camera moves, invented details, changing guests and modest resolution, are the same failures that make real photography and a real property irreplaceable for anything a guest will see. Used where it is strong, it lets a brand team arrive at the shoot already knowing what the film should feel like. If you want to experiment with the same model, or compare it with other video and image models, you can find them on apimart.ai.
