{ "intro": "Hollywood is still failing this film’s voice-continuity and lip-sync requirements. This page shows the unfinished review cut, local tool experiments, and the concrete plan for a corrected Revision 3.", "baseline_url": "/assets/hollywood/review-20261012/baseline.mp4", "baseline_detail": "Revision 2 is the latest complete export, not a quality-approved final. Lip synchronization is visibly poor, and both ending voices differ from their earlier performances. Those defects remain in this 59.833-second cut. Later model tests produced excerpts only; they have not been assembled into a corrected film. Images and voices are synthesized.", "cards": [ { "name": "Revision 2 Maria shot · MuseTalk", "role": "Existing film reference", "verdict": "UNFINISHED BASELINE — lip-sync and voice defects remain", "detail": "The existing four-second crop, including the final phone composite. All diagnostic candidates use the pre-composite source; preserve the receipt, hand and fries when promoting any face result.", "video_url": "/assets/hollywood/review-20261012/baseline-crop.mp4", "seconds": 4 }, { "name": "LTX 2.3 · supplied audio", "role": "Generate acting from image + dialogue", "verdict": "EXECUTION PROVEN — artistic comparison required", "detail": "Rendered 97 frames in 305 seconds at 512×384; peak MLX memory 15.1 GiB. Four sampled frames retain the general character and warm room. Upstream audio level fell about 3 dB; this comparison restores the original dialogue without stretching. This is a short test, not long-scene continuity proof.", "evidence": [ "Runtime pinned at 91e6f6c9bd621ff2ae31adfee643e113d67d6ae8", "Original decoded audio same-timeline correlation: 0.99932" ], "video_url": "/assets/hollywood/review-20261012/ltx-a2v.mp4", "seconds": 4 }, { "name": "LTX DubIt · current MLX integration", "role": "Existing-video dubbing", "verdict": "REJECT FOR THIS FILM — dialogue changed", "detail": "Two runs completed at 512×384 in about 207 seconds each, peak 18.0 GiB. Both generated speech that failed the requested-line transcription check, including an exact-text prompt retry. The displayed video has original dialogue restored for visual diagnosis; that replacement does not establish correct lip sync. Hold this binding pending integration diagnosis, not a global verdict on DubIt. Audit found that target audio starts from noise; reference speech is conditioning, not a timing lock. All 1,344 LoRA pairs match. A separate stage 2 freeze defect has a tested patch artifact, not installed; it does not fix the stage 1 dialogue failure.", "evidence": [ "https://huggingface.co/Lightricks/LTX-2.3-22b-IC-LoRA-DubIt", "Both complete runs and CPU transcript receipts retained with the local test receipts." ], "comparison_default": false, "video_url": "/assets/hollywood/review-20261012/dubit-promptlocked.mp4", "seconds": 4 }, { "name": "DubIt · generated audio diagnostic", "role": "Failed exact-dialogue test", "verdict": "REJECT — keep original voice stems", "detail": "Requested: Okay, can you take this photo? CPU transcription instead reported substantially different words in both runs. This player contains the raw generated audio from the exact-prompt retry, without replacement.", "audio_url": "/assets/hollywood/review-20261012/dubit-raw-audition.wav", "seconds": 4 }, { "name": "LivePortrait", "role": "Drive a still portrait with a performance video", "verdict": "KEEP — controlled portrait motion", "detail": "Rendered four seconds / 96 frames in 107.9 seconds. Sampled frames preserve Maria’s general identity, with plausible mouth, blink and head changes. Phone, hand and background are static. Requires a driving video; this does not prove audio-only lip synchronization.", "video_url": "/assets/hollywood/review-20261012/liveportrait.mp4", "seconds": 4 }, { "name": "Genie · voice auditions", "role": "Clean reference → old ending → Fish → Qwen", "verdict": "EXECUTION + WORDS PASS — identity needs listening", "detail": "All four generated lines passed normalized transcript comparison with zero word errors and no rail clipping. The audition sections are normalized to −20 LUFS / −2 dBTP with half-second gaps. Different durations mean the old timeline cannot simply be reused.", "audio_url": "/assets/hollywood/review-20261012/genie-audition.wav" }, { "name": "Maria · voice auditions", "role": "Clean reference → old ending → Fish → Qwen", "verdict": "EXECUTION + WORDS PASS — identity needs listening", "detail": "All four generated lines passed normalized transcript comparison with zero word errors and no rail clipping. The audition sections are normalized to −20 LUFS / −2 dBTP with half-second gaps. Different durations mean the old timeline cannot simply be reused.", "audio_url": "/assets/hollywood/review-20261012/maria-audition.wav" }, { "name": "Unreal Engine MetaHuman Animator", "role": "Existing dialogue → processed MetaHumanPerformance → reusable named facial animation curves on the installed face template skeleton", "verdict": "KEEP — native local facial-solve foundation", "detail": "5 seconds of existing Genie dialogue. 150 processed frames; 260 curves exported, 105 varying. Full editor run 44.74 seconds.", "limitations": [ "No persistent Maria or Genie 3D rig exists in the audited projects.", "No character render or artistic facial-synchronization quality was qualified.", "Exported native motion is 30 fps and must be sampled in the 24 fps film sequence." ], "evidence": [ "https://dev.epicgames.com/documentation/unreal-engine/unreal-engine-5-8-release-notes" ], "seconds": 5 }, { "name": "speech-swift Audio2Face3D MLX", "role": "Existing dialogue → dense native facial coefficient frames", "verdict": "KEEP AS RETARGETING CANDIDATE — native coefficient inference works", "detail": "5 seconds of existing Genie dialogue. 151 frames × 169 finite, varying coefficients; inference 6.30 seconds.", "limitations": [ "Native James coefficients are not automatically ARKit or MetaHuman control names.", "Matching geometry/rig or a verified retargeting projection is still required.", "No character render or lip-sync superiority was tested." ], "evidence": [ "https://github.com/soniqo/speech-swift" ], "seconds": 5 }, { "name": "Rhubarb + Blender", "role": "CPU speech recognition → 31 timed visemes → nine editable Blender shape keys and 120-frame rendered animation", "verdict": "KEEP FOR SIMPLE CARTOON MOUTH ANIMATION — proxy only", "detail": "5 seconds of existing Genie dialogue. 31 timed cues, nine editable shape keys and a 120-frame proxy render. The proxy is deliberately not the film cast.", "limitations": [ "Simple proxy head, no cast likeness or full-face/body performance.", "Viseme mouth poses need manual refinement and artistic direction." ], "evidence": [ "https://github.com/DanielSWolf/rhubarb-lip-sync" ], "video_url": "/assets/hollywood/review-20261012/proxy.mp4", "seconds": 5 }, { "name": "LTX 2.5 · current Metal adapter", "role": "General video generation", "verdict": "NOT QUALIFIED for this dialogue test", "detail": "An earlier short generation smoke passed. The current Hollywood adapter does not expose supplied-audio conditioning; its quality profiles disable audio generation. No claim that general 2.5 support proves this dialogue path." }, { "name": "CUDA-oriented and missing-runtime candidates", "role": "Further porting, not ready production adapters", "verdict": "HOLD — concrete runtime gaps", "detail": "Wan S2V/Animate, InfiniteTalk/MultiTalk, Hunyuan Avatar and Hallo have no qualified end-to-end Mac adapter here. EchoMimicV3 and FlashHead have MLX weight leads, but the declared runnable implementations were unavailable at audit. Weight availability alone does not qualify a pipeline.", "evidence": [ "https://github.com/antgroup/echomimic_v3", "https://huggingface.co/Avdpro/EchoMimicV3-MLX", "https://github.com/Wan-Video/Wan2.2", "https://huggingface.co/Wan-AI/Wan2.2-S2V-14B", "https://github.com/dandandan1983/Wan2.2_mps", "https://github.com/Wan-Video/Wan2.2", "https://huggingface.co/Wan-AI/Wan2.2-Animate-14B", "https://github.com/MeiGen-AI/InfiniteTalk", "https://huggingface.co/MeiGen-AI/InfiniteTalk", "https://github.com/MeiGen-AI/MultiTalk", "https://github.com/Tencent-Hunyuan/HunyuanVideo-Avatar", "https://github.com/fudan-generative-vision/hallo", "https://huggingface.co/Avdpro/FlashHead-Lite-MLX" ] }, { "name": "Other downloaded audio models", "role": "Music, separation and general speech roles", "verdict": "ROLE TRIAGE — not a film continuity fix", "detail": "The 18-package prior inventory is retained. General music, cleanup or new-speaker models do not by themselves preserve this cast. Fish and Qwen received the actual held-out dialogue tests; imports of other libraries are not presented as inference proofs." }, { "name": "LatentSync 1.6 · repaired MLX port", "role": "Face-region lip correction", "verdict": "EXECUTION PROVEN — no quality promotion", "seconds": 2, "video_url": "/assets/hollywood/review-20261012/latentsync.mp4", "detail": "48 frames at 24 fps rendered in 330.5 seconds. Sampled physical peak 24.4 GB versus 71.6 GB in the initial failed attempt. Fused attention and per-chunk CPU checkpoints fixed major runtime problems. Maria remains recognizable; no clear visual win over MuseTalk. This excerpt contains an initial pause, limiting speech evidence. A separate speech-bearing test is included below; neither is promoted over the current face pass.", "evidence": [ "https://github.com/sb1992/latentsync-mlx" ], "limitations": [ "Requested 12 steps produces 13 scheduler iterations per chunk.", "Kept isolated; production lip-sync default unchanged." ], "comparison_default": false }, { "name": "RIFE 4.25 · 48fps presentation", "role": "Finishing / frame interpolation", "verdict": "KEEP — interpolation, not lip-sync repair", "seconds": 1, "video_url": "/assets/hollywood/review-20261012/rife.mp4", "comparison_default": false, "detail": "Actual 24→47 interpolated frames passed decode, with no adjacent duplicates. Sampled triplets looked plausible. Presentation wrapper adds exactly one endpoint frame to make 48 frames/1 second and restores the original 34–35 second film audio. Raw output is preserved. Runtime receipt includes shared-GPU queue waiting; it is not inference latency.", "limitations": [ "One-second screening does not qualify long-video temporal behavior." ] }, { "name": "SeedVR2 · frame reconstruction", "role": "Restoration / upscaling", "verdict": "REJECT THIS CONFIGURATION — cast and color drift", "detail": "Actual local inference produced a valid reconstructed frame, but sharpened lines accompanied changes to Genie’s facial hair/features and stronger orange contrast. Pixel MAE 7.89 versus 1.76 for bicubic on this synthetic downsample/reconstruction test. The visual identity changes drive the rejection; pixel error alone is not a perceptual quality score.", "image_url": "/assets/hollywood/review-20261012/seedvr-comparison.jpg", "image_alt": "Left: real baseline. Center: bicubic from 256px. Right: SeedVR reconstruction with changed facial details and contrast.", "limitations": [ "One-frame test; no temporal quality qualification." ] }, { "name": "Existing MuseTalk face pass · seconds 2–4", "role": "Matched speech-window baseline, before final phone overlay", "verdict": "BASELINE — fixed speech and frames", "seconds": 2, "video_url": "/assets/hollywood/review-20261012/musetalk-speech.mp4", "comparison_default": false, "detail": "Frames 48–95 of the existing MuseTalk face pass, with the exact original 2–4 second audio window. Both sides use the same 640×480 region and 24 fps. This diagnostic baseline precedes the finished receipt/hand/fries overlay." }, { "name": "LatentSync · speech window 2–4", "role": "Matched speech-bearing lip-sync comparison", "verdict": "WORKING — no demonstrated quality win", "seconds": 2, "video_url": "/assets/hollywood/review-20261012/latentsync-speech.mp4", "comparison_default": false, "paired_baseline_name": "Existing MuseTalk face pass · seconds 2–4", "detail": "The corrected port rendered 48 frames in 320.7 seconds, using the same speech-heavy 2–4 second window as the paired MuseTalk face pass. Full decode passed. Frame inspection keeps Maria recognizable but does not establish a synchronization or appearance improvement. Keep experimental; do not replace the production default. Some lower-face softening and smile/teeth bias remain in sampled frames.", "limitations": [ "Selecting this on the right automatically selects its matched baseline.", "No normal-speed audiovisual acceptance or objective synchronization score claimed." ] } ], "deployment_status": "As tested on 12 October 2026: local generation is the default. Hosted recommendations and fallbacks require explicit opt-in. Voice references stay paired with their transcripts. The routing change passed 16 router tests and 2 capability-registry tests. Experimental tools below have not all been added to the production interface.", "limitations": [ "Tests ran locally on an Apple M3 Ultra Mac Studio with 96 GiB unified memory. This public page hosts copies of the outputs; it does not run generation.", "Short excerpts do not establish long-form consistency or artistic acceptance. No replacement full-film cut is claimed.", "Native Unreal facial animation works, but matching Maria/Genie rigs, retargeting and authored body acting remain unfinished." ], "test_date": "2026-10-12", "published_date": "2026-10-13", "hardware": { "chip": "Apple M3 Ultra", "unified_memory_gib": 96 } }