What software?

The greenhouse was created specifically for this reel as an image-editing example. Two fictional adult gardeners and an empty foreground bowl gave the later requests obvious targets. The sequence creates the scene, adds tomatoes, then removes the people. The framing needed to read as one image being edited. I noticed the colour cast changing between generations and asked for a consistent look, favouring the people and tomatoes.

Further generated colour-matching attempts introduced more variation in the measured background patches. We kept the original three images instead, derived one RGB curve from the preferred reference look, and applied that same curve to every plate. The change was small and repeatable. It did not depend on another model preserving the background.

The butterfly started with my photograph. The local Blender work used curved photographic wing surfaces, a rounded body, legs, antennae and a modelled flower. Before committing to the next animation, I asked for four still proofs: clay, textured, a modest side angle, and registration against the original.

Those views made the construction and its limitations visible. This is a limited-angle reconstruction from one photograph. Hidden anatomy and depth were inferred. The registration proof is a 50% model overlay; the final construction sequence blends back onto the unchanged original photograph.

The subsequent local flight rendered, decoded and passed saved-scene checks. I still rejected it as too rough. We retained the Blender construction, but replaced the flyout with a bounded Replicate test. The first generated take had the wing colour and texture I wanted. Some movement is still imperfect; I chose to keep it. That judgement is part of the result.

I approved a maximum of US$6 and three attempts for the ten-second flight. We submitted one. The recorded estimate for that attempt was about US$1.80. That is the estimate for this video-model call alone, not the total cost of the reel, earlier image work, agent usage or the existing projects.

The runner checked the model’s current input schema before submitting. It used the hand-held print as the first image and solid amber as the requested last image. It also recorded an “unknown” submission state before making the paid request and blocked automatic resubmission until that state could be reconciled. This was a precaution in the implementation; the actual job succeeded.

The output still had a soft highlight in the orange. A short local dissolve to the exact colour, #fda405, supplied the clean join. We solved that small finishing problem locally and kept the selected take.

The request text now begins fading in as each image change starts. The bee alternatives hold for an extra two seconds. The overhead loom view holds for 2.5 seconds. The construction sequence was finally shortened from 11.5 to 10.5 seconds, retaining all ten refinement stages. The main game passage plays at 1.6×; the later Halloween browser capture compresses seven seconds into 4.5.

The condensed creation request, eight-prompt recap and orange closing statement still come before the site tour. Revision 09 preserved the earlier four-second opening; its 96 decoded frames matched the preceding export. The later polish pass deliberately changed that opening and its premise.

For revision 10, we first made a 33-second proof of the new title, readable requests, greenhouse edits, a bee-page move and a diagonal tour entrance. I then suggested a fast-forward glimpse of the saved trials. We compared five-second and 15-second montages before the same title; I selected five seconds. It samples 30 source clips in groups, including earlier attempts. Those files overlap, so the montage is a glimpse of development rather than a record of elapsed working time.

The primary requests are now 104 pixels at 1080p, short edit labels 88 pixels and supporting text 64 pixels. Captions fade away to expose the result. The greenhouse view settles before each edit; the camera drawing moves towards the aperture detail; page views approach a species and its photograph comparison. The tour keeps its original interaction timing, with a short diagonal slide-and-fade between projects. Captured page text and the original drawing labels retain their source sizes.

The finishing pass kept the shots and timing intact. We compared the same 12-second opening with two sound treatments. I selected a locally synthesized fast-forward flutter over the five-second montage; it stops at the title and exposes the original calm score. We also clarified the opening’s wording and scoped the photo credit to “Monarch butterfly photography · Adrian Hensler.” Opening sound record.