AI Whitebox to Render

One clip is required: the grey-box, white-model or greenscreen pass you exported from Blender, C4D, Unreal or whatever you model in. A single style reference still is optional. Under the hood the references are addressed by position in the prompt — the clip is [Video1] and the still is [Image1] — and the still only sets materials, palette, lighting quality and mood. Geometry and camera always come from the clip.

No. The clip is treated as fixed blocking: camera path, focal length, framing and timing stay identical frame for frame, and every object keeps the same position, scale and silhouette. The model is only allowed to add materials, textures, lighting and atmosphere — not to move, resize, remove or invent geometry. That is the whole reason for feeding it previz instead of a text prompt.

Yes. The green is replaced by the environment you describe rather than keyed out, so a subject shot against green ends up standing inside the Sci-Fi Ruins or Cozy Interior you picked. The same rule still applies to whatever is actually in frame: it keeps its position and silhouette, and only the surroundings and the lighting are new.

Sci-Fi Ruins gives weathered concrete and corroded metal, drifting dust in shafts of light, distant haze and a cold overcast key. Cozy Interior gives timber, textile and matte plaster, soft afternoon window light, and plants and props filling the empty volumes. Product Turntable gives a seamless backdrop, polished and brushed finishes, a soft key with a clean rim light and crisp contact shadows. Architectural Exterior gives real concrete, glass and metal, accurate sky reflections, landscaping at the base and natural daylight. Scene notes layer on top — name the materials, the time of day or the mood you want.

A reference video is always attached, and fal bills the input clip's seconds alongside the output, so the billed length is the clip plus the result. At 720p that is 29 credits per second on each side, at 480p 14. Five seconds at 720p is 290 credits, ten seconds is 580 and fifteen seconds is 870; five seconds at 480p is 140. Because the uploader trims your clip to the length you picked, the maths is exact rather than a surprise, and the cost is shown before you generate.

No, and it is worth being clear about that. The model renders rather than ray-traces: there is no physically accurate light transport, no guarantee that your exact shaders come back, and small details can drift between takes. Treat it as a look-development and pitch tool — block a shot, see four lighting directions in minutes, show a client the mood before you commit a render farm to it. It also rewards clean input: evenly lit grey-box footage with clear silhouettes reads far better than a noisy viewport capture with gizmos, grids or overlays visible, and where the blocking is ambiguous the model will interpret it, so simple readable geometry wins.