Before you start
The questions below separate the social-video intent from game-building searches and explain why this Seedance workflow is paid. Read them before uploading so the source photos and expected output match the tool.
What is a hotel lobby generator?
Here it means a browser-based AI video tool that turns two separate photos into a short lobby performance. It is built for a social clip, rather than a 3D building, map, or downloadable room asset.
Is this the same as a hotel lobby generator for Minecraft, Roblox, or DnD?
No. Those searches usually mean a game scene, map, role-playing location, or downloadable asset. This generator creates a performance video from two photos, so the output and the workflow serve a different intent.
How many photos do I need?
The duet flow needs two separate photos, one for each performer. Use one clear subject per image. A future solo variation can use one photo, but a group photo is not a substitute for two controlled inputs.
How long does a generation take?
The task runs asynchronously through the video model. The page shows queued, processing, and ready states while it checks the task, so the interface does not promise a fixed wait time that the provider cannot guarantee.
Why did my result look wrong?
Start with unobstructed faces, separate source photos, similar framing, and a locked camera prompt. If the two people move at the same time or switch sides, fix one instruction at a time instead of rewriting the whole prompt.
Can I use the result on TikTok or Reels?
Yes. Review the 720p preview first, keep the original export, and add captions in the platform editor so the text stays readable when the platform crops the frame.
Two Photos In. One Lobby Performance Out.
The generator is designed around a simple input contract: one clear photo per performer. The reference performance, music direction, camera, timing, and Seedance settings stay on the server, so there is no prompt box or scene picker to configure.
Upload two separate photos
Choose a front-facing image for each person. A visible face, shoulders, or full body gives the model more information than a tiny crop. Avoid group photos, busy backgrounds, face coverings, and strong motion blur. Similar lighting is helpful, but exact matching is unnecessary because the reference performance supplies a consistent stage.
Let the built-in reference guide the render
The private reference clip is about fifteen seconds long and is sent to Seedance together with the two uploaded images. Its movement, camera framing, rhythm, and musical cues guide the result. The default output is 720p; the service keeps Person A on the left and Person B on the right.
Submit and follow the task
After both files upload, the app creates an asynchronous task. The status moves from queued to processing and then to ready when the KIE callback returns a private output. A failed task returns a clear failure state so you can retry through the paid checkout flow, and the interface reports the failure instead of displaying a made-up placeholder video.
Preview and download
Watch the preview before publishing. Check that both faces remain recognizable, the mouth movement follows the performance, and the two performers stay in their assigned positions. Download the finished MP4 during its retention window and keep an original copy before adding captions or making a crop.
Seedance Quality and Paid Generation
The first release uses a paid generation model because the target look depends on high-fidelity faces, stable left-right placement, and natural movement around the microphone. Seedance is expensive to run compared with a lightweight preview model, but it is better suited to a finished social clip that you intend to download and share.
What one paid render includes
One paid request includes the two-photo input, the private reference performance, Seedance model processing, task status updates, and a private result link after the callback succeeds. The exact price should be visible before checkout, along with the expected 720p format and retention window.
Why the cost is higher
High-fidelity image-to-video needs more compute to preserve faces, body proportions, lighting, and timing across multiple frames. Paying per render keeps the product transparent: you purchase the generation you want instead of subsidizing unbounded retries that would reduce output quality or create a long queue.
Best export settings for social platforms
Keep the original 720p MP4, preview it with sound, and add captions after the video is generated. TikTok, Reels, and Shorts may crop interface areas differently, so keep important faces and the center microphone away from the extreme edges. Upload the original export once rather than repeatedly compressing it in several editors.
Common Errors and How to Fix Them
Most weak renders come from a small number of repeatable mistakes. Fix the input or one prompt instruction at a time so the next result teaches you something. Replacing every sentence at once makes it impossible to tell whether the problem came from the photos, the staging, or the motion direction.
Faces blend or the wrong person takes a side
Use separate photos and name the subjects consistently. State “Person A on the left” and “Person B on the right” in the prompt, then repeat the same order in any revision. Do not upload a group crop and expect the model to infer which face belongs to which side.
Both performers move together
Describe the exchange as a sequence: A performs while B reacts, then B performs while A reacts. Add a locked camera and restrained gestures. A short action list is easier to follow than a paragraph full of unrelated adjectives.
The frame crops out feet or the microphone
Use a full-body or wider source photo when possible and explicitly request full-body framing with feet visible. Keep the microphone at center frame and avoid camera pans, zooms, or cuts. If the result is still too tight, change only the camera slot on the next attempt.
The output feels like a generic AI clip
Keep the scene-specific anchors: warm orange booth, one hanging microphone, static centered camera, and an alternating performance. Then add a real caption in the social platform editor instead of asking the video model to draw text into the frame.