Kumar K Jha PRO
kumarkjha
AI & ML interests
None yet
Recent Activity
repliedto their post 1 day ago
I built a pipeline that compiles a screenplay into a limited-animation series, and made one with it: The Brightwater Family — Picture Day, three parts of about nine minutes, rendered on one Mac.
Why compile instead of prompt: before building anything I measured one open-weight generator (LongLive 2.0 on Wan 2.2 TI2V-5B) on a 22-minute test episode. Anchored characters never went missing, but 57% of candidate clips were flagged by a VLM judge for background change, camera drift, or an impossible dog pose (162 clips, 24 segments). So the pipeline stopped generating backgrounds. A room is one drawing, used every time it appears.
The script becomes a typed source of truth, compiled into a per-frame exposure sheet and rendered by a deterministic compositor from a locked library (FLUX.2 [klein] 4B stills from my own designs, Kokoro voices, Rhubarb lip timing). Every frame passes eight hard checks; an AI judge reviews every shot, including "does the frame show what the line says." Remaining judge flags are published: 27 / 15 / 39 across 40,408 frames.
Write-up, with numbers and limitations: https://atveanimation.com/brightwater/how-its-made
Playlist: https://www.youtube.com/playlist?list=PLZOx3V-lNQb0
90-second making-of: https://youtu.be/VwsTME0l0vg
Limited animation by design. Code, drawings, and the bible are private; the method is public. Made with AI tools; all characters, voices, and places are fictional, no real person's likeness or voice. Music composed for the series. posted an update 1 day ago
Second story through the same pipeline, same bible, same rubric: The Brightwater Family — Halloween, three limited-animation parts of about 7½ minutes, rendered on one Mac.
The test was whether the first series' drawings, rules and checks carry over. An AI session with no access to the pipeline wrote all three scripts from a format and an asset list; four revisions went back to it; no line and no frame was edited by hand. Manual corrections: 0 / 0 / 0. Compile rejections: 102, all format or compiler gaps, none writing errors. Reused from series one: 178 of 211 pose drawings, 54 of 65 character looks, 33 of 42 plates. New: four places, eleven costume looks, fangs and chocolate mouth sets, six props. Scripts to three parts passed by review: 26 hours. Compositor render: 8–12 minutes per part; the judge looking at every shot is most of the remaining hour.
Every frame passes the same eight hard checks (0 failures); judge flags remaining after review 7 / 4 / 10; "line is shown" 3 / 1 / 6 of ~172 lines per part; scripted beats performed 13 of 23 (Picture Day: 25 of 86, with the caveat that the format scripts fewer beats). Both series' counts, and the overlap between held beats and flagged lines that a commenter here asked for: https://atveanimation.com/brightwater/how-its-made
Playlist: https://www.youtube.com/playlist?list=PLYPuSb4GK878
Limited animation by design; the voices are synthesized and unchanged from series one. Code, drawings and the bible are private; the method is public. Made with AI tools; all characters, voices and places are fictional; the costumes are generic; music composed for the series. repliedto their post 3 days ago
I built a pipeline that compiles a screenplay into a limited-animation series, and made one with it: The Brightwater Family — Picture Day, three parts of about nine minutes, rendered on one Mac.
Why compile instead of prompt: before building anything I measured one open-weight generator (LongLive 2.0 on Wan 2.2 TI2V-5B) on a 22-minute test episode. Anchored characters never went missing, but 57% of candidate clips were flagged by a VLM judge for background change, camera drift, or an impossible dog pose (162 clips, 24 segments). So the pipeline stopped generating backgrounds. A room is one drawing, used every time it appears.
The script becomes a typed source of truth, compiled into a per-frame exposure sheet and rendered by a deterministic compositor from a locked library (FLUX.2 [klein] 4B stills from my own designs, Kokoro voices, Rhubarb lip timing). Every frame passes eight hard checks; an AI judge reviews every shot, including "does the frame show what the line says." Remaining judge flags are published: 27 / 15 / 39 across 40,408 frames.
Write-up, with numbers and limitations: https://atveanimation.com/brightwater/how-its-made
Playlist: https://www.youtube.com/playlist?list=PLZOx3V-lNQb0
90-second making-of: https://youtu.be/VwsTME0l0vg
Limited animation by design. Code, drawings, and the bible are private; the method is public. Made with AI tools; all characters, voices, and places are fictional, no real person's likeness or voice. Music composed for the series.Organizations
None yet