Give every reference one clear job
Start by separating identity from performance. Portrait A identifies the first person. Portrait B identifies the second. The source video establishes how the scene moves. When a tool supports explicit reference assignment, connect those roles there as well as describing them in text. A sentence cannot replace an input field that the chosen mode does not have.
Decide whether the goal is to preserve a particular source shot or merely create a similar orange-studio scene. Those are different briefs. Asking for a new camera move, a different sequence of gestures and the exact original timing at once introduces competing goals. Pick the property that matters most before describing the rest.
For the fixed site workflow, your practical controls are the photographs, their positions, the source template, frame and processing tier. There is no hidden text field that becomes available by choosing Pro. If you want to direct a new location or custom choreography, use a tool whose visible controls support that project.
Make left and right refer to the viewer's frame
Use an unambiguous assignment such as the first reference remains on the viewer's left and the second remains on the viewer's right. A person's own left is the opposite direction when they face the camera. Explicitly describing the viewer's frame removes a small but avoidable ambiguity from the instruction.
Keep the assignment consistent throughout the brief. If the opening calls one reference Person A and a later sentence calls the same person the right performer, verify that those labels agree. Long prompts often accumulate contradictions during revision because an earlier line survives after the image order changes.
Position continuity matters beyond the first frame. Inspect crossings, turns and moments when a hand covers a face. If identities exchange during motion, the issue is not necessarily the opening composition. The model may be losing track of the reference association over time, which a sharper reference or simpler source can sometimes make easier to evaluate.
An original Hotel Lobby AI prompt brief
For a tool that accepts a source video and two image references, an initial brief could read: Use the supplied video as the performance reference. Replace the left performer with the person in reference A and the right performer with the person in reference B. Keep each identity associated with that side throughout the shot. Follow the source gestures, camera position and timing. Preserve the two-person composition and the central microphone. Do not introduce additional performers or scene cuts.
This example specifies a visual task. It does not grant the model a capability to preserve audio, upload copyrighted material or accept two identities when the chosen mode supports only one. Set the tool's audio options separately where they exist, and check what its documentation says about the output soundtrack.
If the tool instead creates motion from still images, remove the instruction to follow a nonexistent source video. Describe the intended interaction and frame as an original scene. Calling that output a recreation of the exact performance would be misleading unless the resulting movement actually matches it.
Use the brief as a starting point rather than adding every desirable adjective. The most valuable words establish reference roles, continuity and what should remain unchanged. Terms such as perfect, flawless or award-winning do not specify how to resolve a face swap or an occluded hand.
Preserve the action you want before adding action you do not need
The source clip already contains a sequence of events. If you want that sequence, avoid adding instructions for both people to dance, jump or move toward the camera unless those actions are present. A model receiving both preservation and replacement instructions has to reconcile them, and the result may satisfy neither consistently.
Choose an excerpt where the performers remain distinguishable. If one blocks the other for most of the shot, your prompt cannot reveal the hidden face in the original. You can request a generated interpretation, but that is different from preserving visible evidence. Selecting a clearer source is often the more direct change.
The same principle applies to framing. A portrait crop that cuts off a performer cannot be repaired by repeatedly writing keep both people visible while also demanding the crop remain exact. Decide whether reframing is permitted. On this site, choose the intended portrait or landscape template version before submission.
Revise the instruction that matches the visible error
When both people acquire similar facial features, first confirm that the uploaded references are different, clear portraits and that the selected mode accepts both. A stronger separation sentence is only useful after the inputs are correctly associated. Rewriting the background color will not address an identity problem.
When clothing changes unexpectedly, decide whether clothing is supposed to follow the portrait or the source performer. That choice can be unclear when a close-up portrait supplies no outfit information. Do not demand an exact full-body costume from a reference that shows only a face; provide an appropriate reference if the tool supports it.
When gestures look wrong, compare the result against the source at the same moment. If the requested action conflicts with the source, simplify the brief. If the source itself contains an abrupt cut, choose a continuous excerpt before blaming the identity instruction. Record the change so you know what the next generation is meant to test.
Consider a result where both people start correctly but the right face changes when a hand crosses it. A targeted next test would keep the source, frame and left reference fixed while replacing the right portrait with a clearer one. If you instead add a new camera move and several wardrobe instructions, the next result tells you much less about the original identity problem.
For a different case, imagine the faces are stable but both people perform at once when the intended source alternates between them. Review whether the brief introduced synchronized action. Remove that contradiction before adding more identity language. The point of revision is to narrow the task the model has to reconcile, not to make the instruction longer.
A new attempt costs time and may cost credits. Change one meaningful variable at a time, such as the second portrait or the source excerpt. Changing the prompt, both images, the frame and the model together prevents you from learning which change helped.
Match the brief to the tool's actual controls
Higgsfield documents source-video transformation through Genjutsu, with modes for broader motion transfer or selected changes. Viggle distinguishes a single-character Mix from workflows involving several character images. These are different input arrangements, so a Hotel Lobby AI prompt copied between them may need more than a wording change.
Check duration, reference slots and available controls in the active generator. Do not assume an example from a tutorial uses the same mode or account access you have today. Also check whether the interface offers presets: writing a long prompt can be unnecessary when the required transformation is already described by a supported preset.
A negative instruction is not an independent repair system. Saying no extra people describes your preference, but it does not guarantee that a difficult occlusion will be resolved. Evaluate the output, especially during the moments most likely to confuse identities, before describing any prompt as reliable.
Apply the reasoning without typing a prompt here
For this site's generator, prepare two clear portraits, verify the left and right slots, and inspect the available template preview. Those actions provide the references that the fixed internal prompt uses. A template without configured source media remains unavailable regardless of how strong a written brief might be.
After generation, review the same properties the brief describes: separate identities, stable positions, a coherent shared scene and appropriate motion. Use the main creation guide for submission and download steps. Use a general video workflow if the project requires instructions beyond the current two-person Hotel Lobby format.