Submission polish: README, thumbnail, blog notes, song fit

Reframe Tiny Titan -> Tiny Mode (award is the judges' call; describe the feature)

Revert intro-trim + duration/cap changes — they broke lip-sync/song alignment

Video: drop subtitles, trim long intro/outro, sing more of the lyric

Prep example re-render: new-API render script + HTML-trace handling

Tiny Mode: detect calibration-anchor echo and regenerate hotter

Song box plays the video's audio (your voice, not AI); tighter lyric fit

Subtitles: energy-based vocal-span timing (fix blaze-forward); turbo step cap

Tiny Titan: full <=4B path (gpt-oss-free) + Understudy quality + README

Subtitle timing, drill rap, honest Tiny Mode label

Subtitles: derive from the same trimmed lyrics ACE-Step sang

Song fidelity + scroll cue on Write My Lyrics

Songs: trim long drafts so the whole lyric is sung; tighten structure

Fix résumé sound first-go + steer lyric gender from Singing Voice

Add theater-moment sound effects (quiet, baked, Off-Grid)

Full-width Libretto stage: velvet-desk bg, bulb-frame glow, clean Producer's-trace sheet

Act III stages: curtain/parchment bg art + Pinyon-script idle text + song loader; single-icon labels; fix headshot bottom border (overflow clip); bold dark accordion label

Re-encode replaced ACT ribbons; voice card = centred mic + full-text Record + consent (no stray X, no truncation); headshot source icons in gold-outlined row below boundary; hide fullscreen on loaded photo; hide mic when recording waveform shows

Photoreal vintage mic + camera dropzone icon from dropped assets; fix stray line by re-slicing plaques at true frame edges (no neighbor bleed)

Fix outer plaques: re-slice using detected gold-frame boxes so all 4 borders + dark fill are intact (SE/DS were edge-cropped)

Match reference exactly: ornaments sliced from reference art, baked into src/ornaments.py (base64 data-URIs, git-safe, Off-Grid)

Add 'Try an Example' cached showcase + fix Tiny Mode via prebuilt wheel

Wire Producer agent loop, Tiny Mode (Understudy GGUF), and trace publishing

Revert to pre-Antigravity layout: remove splash/lazy-load/viewport-lock, restore clean natural-flow theatrical dashboard

Optimize startup: lazy load models, background preloader with Broadway-themed splash screen, and fader UI fixes

Carousel cast-to-select (glow only on cast + themed popup); gold ticket-stub Write button; dashed ink-stamp Tiny Mode; themed centered voice label + emoji chips; fader number-box removed + darker typewriter text + bigger brass fader; typewriter libretto font + typing sound

Per-genre poster glow color (--case-accent) + recast pop animation; move count badge clear of title

Act I + carousel polish: dark-mat binder (single outline), top title bars w/ visible subtitles, SVG clapperboard dropper, vintage mic emblem, centered Record, hide mic-picker + clear-X, equal-height cards; redesigned 9 Broadway-style posters (signature color fields + bold motifs); centered carousel arrows + stronger gold case glow + pick caption

Broadway playbill UI: marquee, 3-act widescreen dashboard, poster carousel (gr.State+SVG), mixing-desk fader, gold dossier button, typewriter JD memo, velvet proscenium curtains, in-browser WebAudio score

Wire seed-VC own-voice into video chain; harden roast (token budget + channel-leak guard)

Two-stage roast: moondream sees, gpt-oss wits; vendor seed-VC source for own-voice

Swap roast model MiniCPM-V-2 -> moondream2 (modern, transformers-4.57 compatible)

Load MiniCPM-V-2 in fp32 to avoid vision-tower dtype mismatch

Pass context=None to MiniCPM-V-2 chat (required positional arg)

Restore removed numpy aliases (np.float et al) in compat shim for SadTalker

Whisper subtitle alignment in Space video chain; hidden debug output for API-visible video errors

Day-2 completion push: video chain + MiniCPM-V roast wired into the Space

Wiring prep: vendor SadTalker source (MIT, 812K, no binaries), packages.txt (ffmpeg+fonts), torchvision compat shim

Fix branch order: custom vibes were falling into AI-decide path

3 new genre presets (R&B/EDM/Arena Rock) + free-text custom vibe

Front-load + strengthen voice tag; 12 diffusion steps for adherence

Write wav via soundfile (torchaudio.save now requires torchcodec)

Save returned waveform tensor ourselves — this ACE-Step build skips disk write

Resolve audio path defensively: ACE-Step may return dir instead of file

Pin gradio 6.2.0 (sdk) + transformers 4.57: ACE-Step's own tested combination

Plan B: re-init ACE-Step inside GPU context per render; startup is download-only

Cast bf16 on real CPU tensors before managed cuda handoff (NaN latents)

Hoist _move above startup init block (NameError)

ACE-Step to cuda at startup via ZeroGPU-managed path; drop per-call shuttle

Full traceback in music errors; seed-VC chain test job (demucs -> seed-vc -> remix)