AI Timelapse / Transformation Films

Agricultural change happens slowly. A rice field transforms over months, not seconds. A raw timelapse can show this transformation, but it often carries visual instability: clouds jump, light changes abruptly, exposure shifts, field colour flickers, and missing stages interrupt the continuity.

For this project, the challenge was precise: keep the field real while making the transformation cinematic and legible.

The left and right Tikee captures were exported, stitched in Hugin, and batch-processed through a Python-assisted workflow. Once the crop and final dimensions were locked, the frame became the coordinate base for the entire project.

After this stage the image could not be casually recropped — doing so would break the relationship between the placed sun, moon and final frame.

The film was built from selected keyframes. The selection followed a diagonal path through the season: early dates paired with morning light, mid-season frames with stronger daylight, later stages with evening and dusk.

This let one visual day stand in for the full rice-growing season. The selection was not only technical — it was editorial.

The sun and moon were not left for the model to invent. They were placed manually and deterministically — adjusted in Photoshop for realism, including opacity shifts with cloud cover — and then fixed. The AI relighting stage was instructed to preserve their exact position and only harmonise the scene around them.

This kept light direction intentional and prevented the model from redesigning the sky. Drag to compare.

The relighting stage was designed to be narrow. The model had to preserve the field, horizon, poles, shrubs, crop structure, framing and sun position. It could improve atmosphere, haze, cloud depth and light integration — but it could not repaint the scene.

Broad prompt language caused drift: moved suns, repainted fields, oversized skies. The working prompts became technical and constrained. Drag to compare the constrained relight.

The selected frames were interpolated with LTX 2.3 inside ComfyUI. LTX was not used to invent a field from text — it interpolated motion and transformation between real photographic states. Rather than one long unstable film, the sequence was divided into seven overlapping windows of five keyframes each, with one repeated handoff frame shared between consecutive windows.

The hard part was not making a single short AI clip. It was making several generated clips behave as one continuous film. A segment could look convincing alone and still fail when joined to the next: clouds changing direction, sky texture shifting, haze and luminance jumping, field colour flickering, growth discontinuity.

The workflow solved this through locked keyframes, fixed light placement, restrained sky reconstruction, repeated overlap frames, multi-pass generation and manual assembly. The duplicated handoff frames were used during generation, then trimmed at the joins.

The final film compresses the rice-growing season into a 75-second visual arc — morning, growth, drying, ripening, harvest, dusk and a night coda. Roughly 70 seconds of daylight and dusk, with a 5-second night close, and a date counter to keep the agricultural timeline legible.

The result is neither a raw timelapse nor a fully generated video. It is a hybrid transformation film built from real capture and controlled AI reconstruction.