{ "date": "2026-09-23", "task": "Ref2VA", "requested_seconds": 10, "speech_target_seconds": [ 5, 10 ], "dialogue": "你好，我是知微。今天陪你读书，也聊聊你的故事。", "selection": "First formal sample after one same-input warmup per model, fixed before visual review.", "reference_sha256": "95029c1803684ef269c161da628a52eaa9db74803d9863716cade4c43ca671ca", "prompt_sha256": "b82c82c2e0d6a066a02261977cba57866b68b5c60426009c5769b80daa3ce218", "arms": [ { "id": "lightx2v-original-4step", "label": "LightX2V 原精度四步", "configuration": "Turbo v0.1 · BF16 DiT · 原版 VAE", "seed": 230923, "denoiser_evaluations": 4, "request_seconds": 24.092093851882964, "raw_duration_seconds": 10.125, "full_decode": "PASS", "width": 960, "height": 544, "fps": 24, "raw_frames": 243, "preview_sha256": "43bc296730f702596cdd32e52e6717cdb0fef128130ef64fb6ffcf7d0909b4c7", "raw_sha256": "8a0172d034ddaf63addb662fce1e8a71d62e70024a1f03a5831f699cfbe06fd8", "preview_frames": 240, "preview_duration_seconds": 10.0, "raw_file": "https://github.com/lishunyang12/fasth3-lossy-ablation/releases/download/avatar-quality-reference2-20260923/lightx2v-original-4step-raw.mp4", "preview_file": "lightx2v-original-4step.mp4" }, { "id": "minimax-original-50step", "label": "MiniMax-H3 原版 50 步", "configuration": "无蒸馏 LoRA · BF16 DiT · 原版 VAE", "seed": 230923, "denoiser_evaluations": 50, "request_seconds": 263.847028760938, "raw_duration_seconds": 10.125, "full_decode": "PASS", "width": 960, "height": 544, "fps": 24, "raw_frames": 243, "preview_sha256": "cd5ec5df0f46b8b79e3bbed9b805e987fa036a5f05e97a647983a23a7719fc12", "raw_sha256": "29ad4d1ca2249834e0d4386fed41d799b59246cfc8dc650eea81652d49299a41", "preview_frames": 240, "preview_duration_seconds": 10.0, "raw_file": "https://github.com/lishunyang12/fasth3-lossy-ablation/releases/download/avatar-quality-reference2-20260923/minimax-original-50step-raw.mp4", "preview_file": "minimax-original-50step.mp4" }, { "id": "deployed-4step", "label": "上线四步配置", "configuration": "MXFP8 + Sage · NVFP4 VAE · 10 秒帧数适配", "seed": 230923, "denoiser_evaluations": 4, "request_seconds": 6.958027348853648, "raw_duration_seconds": 10.125, "full_decode": "PASS", "width": 960, "height": 544, "fps": 24, "raw_frames": 243, "preview_sha256": "25ca3d6eb74719e7c81f3c8c76c7e4c48c39ed63c9f57b701e564527c7dfe4c0", "raw_sha256": "3455799896d9c0274b5dccbdf631f59232d8a10fb29f8c0238fa54dd9a3d97b0", "preview_frames": 240, "preview_duration_seconds": 10.0, "raw_file": "https://github.com/lishunyang12/fasth3-lossy-ablation/releases/download/avatar-quality-reference2-20260923/deployed-4step-raw.mp4", "preview_file": "deployed-4step.mp4" } ], "review": { "scope": "CPU Whisper small word timestamps and full decoded audio diagnostics; not a lip-sync or human listening score. Decoded AAC samples above full scale can be codec overshoot and do not alone prove waveform clipping. Visual observations inspect six source frames per arm (0.5, 2.5, 4.5, 5.5, 7.5 and 9.5 seconds); proper-name spellings are ASR output, not a verified pronunciation error.", "expected_dialogue": "你好，我是知微。今天陪你读书，也聊聊你的故事。", "target_speech_interval_seconds": [ 5, 10 ], "arms": [ { "id": "lightx2v-original-4step", "label": "LightX2V 原精度四步", "recognized": "你好 我是芝薇今天陪你读书也聊聊你的故事", "words": [ { "word": "你好", "start": 4.460000000000001, "end": 4.98, "probability": 0.468161940574646 }, { "word": " 我是", "start": 4.98, "end": 5.5, "probability": 0.7206525802612305 }, { "word": "芝", "start": 5.5, "end": 5.84, "probability": 0.7214449346065521 }, { "word": "薇", "start": 5.84, "end": 5.94, "probability": 0.8910984992980957 }, { "word": "今天", "start": 6.6000000000000005, "end": 7.12, "probability": 0.6985118389129639 }, { "word": "陪", "start": 7.12, "end": 7.3, "probability": 0.9931080043315887 }, { "word": "你", "start": 7.3, "end": 7.5, "probability": 0.9807112812995911 }, { "word": "读", "start": 7.5, "end": 7.62, "probability": 0.7816160023212433 }, { "word": "书", "start": 7.62, "end": 8.0, "probability": 0.9993934035301208 }, { "word": "也", "start": 8.0, "end": 8.52, "probability": 0.5443407893180847 }, { "word": "聊", "start": 8.52, "end": 8.7, "probability": 0.9916449189186096 }, { "word": "聊", "start": 8.7, "end": 8.86, "probability": 0.9798868894577026 }, { "word": "你的", "start": 8.86, "end": 9.12, "probability": 0.9922010898590088 }, { "word": "故", "start": 9.12, "end": 9.4, "probability": 0.9988248944282532 }, { "word": "事", "start": 9.4, "end": 9.6, "probability": 0.9994959831237793 } ], "asr_start": 4.460000000000001, "asr_end": 9.6, "speech_timing_description": "ASR 估计 4.46–9.60 秒（跨度 5.14 秒）", "visual_observation": "抽帧可见由低头读书转为面对镜头，4.5 秒时已张口。后半段口部动作和微笑较明显，书本、服饰与桌面茶具保持可辨认。", "audio_signal": { "peak": 1.3414987325668335, "samples_at_or_above_full_scale": 1020, "first_five_seconds_rms": 0.1280062198638916, "final_five_seconds_rms": 0.30663004517555237, "finite": true } }, { "id": "minimax-original-50step", "label": "MiniMax-H3 原版 50 步", "recognized": "你好 我是芝薇今天陪你读书也聊聊你的故事", "words": [ { "word": "你好", "start": 4.440000000000001, "end": 4.960000000000001, "probability": 0.5100014805793762 }, { "word": " 我是", "start": 4.960000000000001, "end": 5.48, "probability": 0.7007771134376526 }, { "word": "芝", "start": 5.48, "end": 5.84, "probability": 0.7508285343647003 }, { "word": "薇", "start": 5.84, "end": 5.92, "probability": 0.8969546854496002 }, { "word": "今天", "start": 6.6000000000000005, "end": 7.12, "probability": 0.7370511889457703 }, { "word": "陪", "start": 7.12, "end": 7.3, "probability": 0.9939956068992615 }, { "word": "你", "start": 7.3, "end": 7.5, "probability": 0.9851073622703552 }, { "word": "读", "start": 7.5, "end": 7.62, "probability": 0.8739420771598816 }, { "word": "书", "start": 7.62, "end": 8.0, "probability": 0.9994166195392609 }, { "word": "也", "start": 8.0, "end": 8.52, "probability": 0.5786786079406738 }, { "word": "聊", "start": 8.52, "end": 8.7, "probability": 0.9862059950828552 }, { "word": "聊", "start": 8.7, "end": 8.86, "probability": 0.9825882911682129 }, { "word": "你的", "start": 8.86, "end": 9.12, "probability": 0.9916279315948486 }, { "word": "故", "start": 9.12, "end": 9.38, "probability": 0.9988558292388916 }, { "word": "事", "start": 9.38, "end": 9.58, "probability": 0.9994003772735596 } ], "asr_start": 4.440000000000001, "asr_end": 9.58, "speech_timing_description": "ASR 估计 4.44–9.58 秒（跨度 5.14 秒）", "visual_observation": "抽帧可见先低头读书，再抬头面对镜头；4.5 秒时已张口，后半段保持正面交流姿态。书本和双手持续可见，窗格、茶具及衣袖纹理保持可辨认。", "audio_signal": { "peak": 1.3102874755859375, "samples_at_or_above_full_scale": 857, "first_five_seconds_rms": 0.12457426637411118, "final_five_seconds_rms": 0.2986317276954651, "finite": true } }, { "id": "deployed-4step", "label": "上线四步配置", "recognized": "您好,我是芝薇。今天陪你读书,也聊聊你的故事。", "words": [ { "word": "您", "start": 4.0200000000000005, "end": 4.5, "probability": 0.1568136364221573 }, { "word": "好,", "start": 4.5, "end": 4.72, "probability": 0.9907346963882446 }, { "word": "我是", "start": 5.34, "end": 5.5, "probability": 0.961224377155304 }, { "word": "芝", "start": 5.5, "end": 5.86, "probability": 0.6725572794675827 }, { "word": "薇。", "start": 5.86, "end": 5.92, "probability": 0.928257554769516 }, { "word": "今天", "start": 6.84, "end": 7.12, "probability": 0.9649995565414429 }, { "word": "陪", "start": 7.12, "end": 7.32, "probability": 0.9912581145763397 }, { "word": "你", "start": 7.32, "end": 7.5, "probability": 0.8851830363273621 }, { "word": "读", "start": 7.5, "end": 7.64, "probability": 0.8993741571903229 }, { "word": "书,", "start": 7.64, "end": 8.02, "probability": 0.9991107285022736 }, { "word": "也", "start": 8.42, "end": 8.56, "probability": 0.9915772676467896 }, { "word": "聊", "start": 8.56, "end": 8.7, "probability": 0.9841914176940918 }, { "word": "聊", "start": 8.7, "end": 8.86, "probability": 0.983761727809906 }, { "word": "你的", "start": 8.86, "end": 9.12, "probability": 0.9728408455848694 }, { "word": "故", "start": 9.12, "end": 9.42, "probability": 0.9982847571372986 }, { "word": "事。", "start": 9.42, "end": 9.6, "probability": 0.9994165897369385 } ], "asr_start": 4.0200000000000005, "asr_end": 9.6, "speech_timing_description": "ASR 估计 4.02–9.60 秒（跨度 5.58 秒）", "visual_observation": "抽帧可见先读书，再面对镜头开口；后半段有明显口部动作与微笑。书本、衣袖、桌面物件保持可辨认；抽帧不能单独证明连续口型同步。", "audio_signal": { "peak": 0.9385237097740173, "samples_at_or_above_full_scale": 0, "first_five_seconds_rms": 0.08741135150194168, "final_five_seconds_rms": 0.213079035282135, "finite": true } } ] }, "model_files": { "lightx2v_ref2va_turbo_adapter": { "filename": "minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors", "bytes": 1383677768, "sha256": "9e642fc8749c74f8da5e2382877ab5c7aa37b9a73b7fd0d6d457bd1b3cb1ae99" }, "base_model_index": { "filename": "model_index.json", "bytes": 707, "sha256": "b160b13700bf08aea7fce1bc3dd0d3a49ea15036298c50082ec107f87843b90c" }, "base_transformer_config": { "filename": "config.json", "bytes": 604, "sha256": "f619093a231fcfbcc3d035bec26c50ad864e7331a500d5c519f5045dc1e50458" }, "four_step_model_index": { "filename": "model_index.json", "bytes": 789, "sha256": "15477c22a5021179b36439c27e63672c5c5b29e5f4bc11bd50e011c91ae9e973" } }, "reference_provenance": { "source_url": "https://github.com/lishunyang12/vllm-omni-rankings/blob/main/videos/8a65f5fc738acc3e0a78d0a727444902.png", "download_url": "https://raw.githubusercontent.com/lishunyang12/vllm-omni-rankings/main/videos/8a65f5fc738acc3e0a78d0a727444902.png", "bytes": 2151225, "sha256": "95029c1803684ef269c161da628a52eaa9db74803d9863716cade4c43ca671ca" }, "reference_image_comparison": { "previous_size": [ 1672, 941 ], "new_size": [ 1672, 941 ], "identical_rgb": false, "mean_absolute_rgb_difference": 5.989999906780767, "max_rgb_difference": 147 }, "previous_comparison": "https://lishunyang12.github.io/fasth3-lossy-ablation/avatar-quality.html", "interpretation": [ "Original means original checkpoint and inference precision, executed through the local native vLLM-Omni port; it does not mean stock Diffusers execution.", "Original precision arms use TP8/SP1; the deployed configuration uses TP1/SP8. Timings are observations of these exact setups, not an isolated quantization speedup.", "Both original arms use BF16 DiT, dense BF16 cuDNN attention, FP32 full video and audio VAEs, and no approximate denoising cache.", "The deployed arm uses the captured production source, four-step LoRA, MXFP8 DiT, dense Sage attention and NVFP4 full video VAE; audio VAE remains FP32.", "The saved live VAE shape allowlist rejected a single 10-second request. An isolated source copy adds only 72 to its permitted latent lengths (14 windows, 243 frames). Weights, numerical paths and runtime audits remain unchanged; see qualification.json for the exact one-line adaptation and hashes.", "All three samples start from one clean reference, with no external audio reference or incoming continuation tail. This evaluates a single 10-second generation, not long-stream drift.", "Native 243-frame outputs last 10.125 seconds. Previews contain the first 240 frames, re-encoded identically without frame interpolation, enhancement or retiming.", "One prompt and seed do not establish general quality; a 50-step generated sample is not a ground-truth video." ] }