Big focus on speed and usability this sprint. The media loading was painfully slow because every image round-tripped through the Google Drive API, which was killing the editing flow. We also finally tackled the render pipeline bottleneck that had been bugging us since v0.5.

• Photo caching layer — Server-side LRU disk cache for Drive-hosted images (5 GB cap, with a 20% disk-free safety valve). First load still hits Drive, but every subsequent request is served straight from disk. Cuts image load from ~800ms to under 10ms.
• Parallel render pipeline — Music generation and AI video clips now fire off simultaneously instead of waiting in line. On a 4-clip job with Veo, this shaves off minutes.
• Concurrent Veo submissions — Multiple Google Veo clips are submitted to the cloud API in parallel. Each poll loop runs independently. Wan2.1 jobs still serialize on the GPU side (hardware constraint), but Veo is all cloud so there’s no reason to queue them.
• Voiceover duration fix — Scripts were consistently coming up short. Added a 20% buffer to the word-count target so the voice actually fills the video. The LLM prompt and client-side estimator both account for the extended target now.
• Audio track auto-sync — Voice and music bars now default to the full video duration and resize automatically when you add or remove clips. Once you manually drag-resize a track, it stays pinned to your setting.
• Audio offset & trim in renderer — The voice and music track positions you set in the timeline are now actually sent to FFmpeg. You can shift the voice start or trim the music shorter and it will render exactly as shown.
• Timeline scrollbar overhaul — The old 4-pixel scrollbar was nearly invisible. Now it’s 14px with a gold accent thumb and a hover state. Much easier to grab and scrub through longer timelines.
• Veo error reporting — Veo errors now surface properly in the stats panel and clear themselves on the next successful generation. No more stale “offline” badges after a transient failure.

Massive milestone — Google Veo 3 is now the default video engine. This was a week-long effort that involved rewriting the entire clip generation pipeline to support both cloud and local models. Debugging the Wan2.1 memory-mapped file crash on Windows took two late nights before we pinpointed it as an OS-level paging file limit.

• Google Veo — Cloud-based video generation via Google’s Veo 3 API. Fast, high-quality, no local GPU needed.
• Wan2.1 stability fix — Resolved Windows memory-mapped file crash (OS error 1455) that prevented the 14B model from loading.
• GPU memory management — SDXL and MusicGen are now offloaded to CPU during video generation, freeing ~13GB VRAM.
• GPU serialization — Only one video generation runs at a time, preventing VRAM contention crashes.

The feature everyone was waiting for. Getting Wan2.1’s 14-billion parameter model running on a single consumer GPU required serious optimization work. The Ken Burns fallback was added after we realized not every clip would survive the AI pipeline on the first try.

• Wan2.1 Image-to-Video — Generate cinematic motion from static images with text prompts.
• Text-to-Video mode — SDXL generates a starting frame, then Wan2.1 animates it.
• Per-clip AI controls — Scene description, duration, and motion mode per clip in the timeline.
• Retry logic — 8-attempt retry for model loading, 3-attempt for API errors, 12-min polling timeout.
• Ken Burns fallback — If AI generation fails, clips gracefully fall back to pan/zoom motion.

Went from “upload photos and hit generate” to a proper NLE-style editor. The drag-and-drop timeline was built from scratch. Getting the ElevenLabs voice pipeline and MusicGen audio to play nicely with FFmpeg’s filter_complex was the hardest part — the mix chain was rewritten three times before the timing was right.

• Timeline editor — Drag-and-drop clip reordering, per-clip duration and transition settings.
• AI background music — MusicGen generates mood-matched music (9 presets + custom prompts).
• ElevenLabs voiceover — Premium neural voices with preview playback.
• Three templates — The Epicure (16:9), The Rush (1:1), The Reel (9:16).

Where it all started. The first prototype was built over a weekend — a simple photo uploader that rendered Ken Burns videos with FFmpeg. We knew right away this had legs.

• One House Video Studio — Upload photos, configure Ken Burns motion, render restaurant video ads.
• FFmpeg rendering — H.264 output with crossfade transitions and text overlays.
• User accounts — Login, registration, job queue, and gallery.