Three sessions. Not a finished product – but a toolkit that is beginning to take shape.

What exists so far

238 episodes are secured. He-Man and She-Ra, in German and English. That was the foundation – everything else builds upon it.

The pipeline is running

Every episode goes through a chain: Video – individual frames – classification. Filmation recycled massive amounts of footage in the 80s. The AI separates this: characters, backgrounds, effects, props – each category into its own folder. Anything that appears multiple times is filtered out.

In parallel, the audio side. Three stems are extracted from every video: speech, music, sound effects. Separated. Whisper transcribes the speech track – with timestamps and character assignment where possible.

What the toolkit currently includes

Frame Library: Classified individual images by character, background, effect, and prop.

Audio Stems: Speech, music, and SFX cleanly separated, per episode.

Voice Reference DB: The best voice samples for each character and language, summarized into 15-30 second reference tracks.

Character Reference DB: Up to 20 different reference images per character, sorted by shot type.

4K Upscaling: RealESRGAN runs on the RTX 4080 – the cels are upscaled to 4K.

What is still missing

She-Ra in English is not yet complete. The TTS synthesis – F5-TTS using the voice references – is prepared but not yet integrated. And the final assembly: Frames + Audio + Synthesis into an output format. This is coming.

No date. When there is something to show, there will be an update.