How to make an AI building growth animation from two images

An empty lot, a concrete core, a stack of exposed floors, then glass climbing toward the roof. That sequence is what makes a building growth animation read as construction. Asking only for a skyscraper to "grow" leaves the model too much room to stretch a finished building upward.

Here is our local test, with the prompt below.

Open the Building Growth Animation prompt to copy the video direction and download the matching first and last images. The full workflow and local settings are below.

Example images and experimental video, made outside Studio. The clip runs for eight seconds. Its 1080p delivery version is an upscale of the 864 × 480 local generation. Unmute the player for music.

The idea came from 夏夏的AI日记's building growth animation on Rednote. Prompt inspiration only. We made our own city and tower images, wrote our own prompts and generated the video locally. No images or footage from that post appear here.

Start with the completed building

The last frame defines what the model has to finish: the tower's position, stepped outline, facade and roof. Use a view that includes the entire building and enough ground to see where it starts. A clear silhouette is easier to follow than a tower hidden behind neighboring buildings.

Our final frame is an original image made with Image Gen. You can instead start with a render or photograph you have permission to use.

Zoom in

Keep this image as the last frame. Make the first frame by editing this same picture, rather than generating another city independently. The two endpoints need to agree about the camera and surroundings.

Make the empty-lot first frame

Remove the subject building and its attached podium. Keep the roads, nearby buildings, trees, skyline and light in place. This is the image-editing prompt we used:

Edit this image into the BEFORE frame of a construction timelapse.
Remove ONLY the central tall glass skyscraper and its low attached
podium. In its footprint show a flat empty brown gravel construction
plot at street level. Reconstruct the small distant background
buildings and sky previously occluded by tower.

Strictly preserve original pixel alignment of all existing surrounding
buildings, roads, trees, foreground, sun, sky, mountain skyline and
lighting. Same image dimensions, exact fixed camera and composition.

Absolutely no tall tower, no podium, no cranes, no debris, no writing.
Do not alter surrounding city or regenerate composition.

Zoom in

"Preserve" is an instruction, not a guarantee. Compare the images before sending them to the video model. Check road corners, foreground roof edges, neighboring windows and the skyline. If those move between the endpoints, the animation has to reconcile them too.

The area behind the removed building is invented. It is not a recovered photograph of what was actually there.

Copy the construction animation prompt

Set the empty lot as the first frame and the completed tower as the last frame. Use a model and workflow that accept both endpoints. This is the exact English prompt sent to our local H3 workflow:

A photorealistic architectural construction timelapse, fixed locked
camera. Start with the empty plot in the first image and progressively
construct exactly the skyscraper in the last image. Construction
begins immediately.

The central concrete core rises first, followed closely by steel
columns, beams and horizontal floor slabs assembling one floor at a
time from ground to roof. For most of the shot the upper floors show
exposed structural framing while completed glass curtain wall panels
install floor by floor below. Temporary tower cranes and scaffolding
accompany the active top floors, then disappear as construction
finishes. The stepped crown and bronze vertical fins complete last.

Finish with exactly the completed tower in the final image, hold the
finished view briefly. Clear bottom-to-top physical assembly, visible
intermediate skeletal construction stages.

Preserve all surrounding city buildings, roads, trees, mountains,
sunlight and sky, with the camera absolutely stationary. No camera
movement, no cuts.

Do not lift a completed tower out of the ground, stretch an existing
building, fade in a transparent building, melt glass, or distort the
tower. No text or logos.

The useful part is the order of operations. The core leads. Floors follow. Glass trails the active structure. The model has visible intermediate states to aim for, rather than only an empty start and a finished end.

Keep the camera locked for the first attempt. An orbit adds another problem: inventing sides of the building and surroundings that neither endpoint shows.

The local settings behind this clip

This was a ComfyUI experiment using the installed MiniMax H3 weights, not a generation through Astravisualize's video service.

Setting	This test
GPU	NVIDIA GeForce RTX 4080 SUPER, 16 GB VRAM
Model	MiniMax H3 FL2VA, pruned INT8
LoRA	H3 FL2V turbo, eight-step version
Conditioning	First and last images
Sampler / schedule	res_multistep / simple
Sampling steps	8
Seed	2026102501
Generated dimensions	864 × 480
Frames / playback	192 frames at 24 fps, 8 seconds
Logged execution time	About 4 minutes 49 seconds
Delivery version	1920 × 1080, Lanczos upscale with light sharpening

For reproducing the local recipe, these are the exact installed filenames:

Model: minimax_h3_fl2va_pruned_int8_convrot.safetensors
LoRA: minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors

We generated one local video for this example. That makes it a worked example, not evidence of a reliable success rate. The execution time belongs to this machine and run; it is not a service speed estimate. No paid video API request was made, but local hardware and electricity still have a cost.

Upscaling makes the file suitable for a larger player. It does not recover structural detail missing from the generated frames. The edit also replaces the model's audio with an excerpt of Alex Morgan's "Corporate Business Background," a track previously selected for our films, with short fades at the ends.

What worked, and what still looks generated

The core appears before the completed facade. Exposed floor plates and cranes make the middle of the clip legible as a construction site. Glass then fills the lower structure while the top remains unfinished. The surrounding skyline stays broadly stable in the inspected frames.

The finish is less convincing. The upper structure and crown resolve quickly near the end, and temporary equipment disappears as the building completes. This is a plausible visual shorthand, not a checked construction sequence. We have not verified crane mechanics, structural connections or buildability.

Watch the middle before judging the last frame. The last frame was supplied to the model. A good landing alone says little about whether the path there looks convincing.

For another attempt, change one thing at a time. If the whole tower stretches, describe exposed floors and a leading core more clearly. If the city drifts, repair the endpoint images first. If the roof snaps into place, give completion more time or try an intermediate unfinished-building guide in a workflow that supports one. Those are troubleshooting directions, not additional results tested here.

Where this effect is useful

A short concept film, a presentation opener or an architectural social post can use this kind of compressed construction story. It should be labeled as an AI visualization. It should not stand in for a construction schedule, progress record or structural simulation.

For an interior film with a moving camera, our room-to-room transition breakdown covers a different problem: keeping direction and geometry consistent while hiding the join. To prepare still design images from your own project, start in Studio. The local H3 recipe above is separate from that product workflow.