{ "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": "8db8820f82cb14799e52343a5581502e0f62239237e4883205ca3462b1672c15", "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": 24.044001692906022, "raw_duration_seconds": 10.125, "full_decode": "PASS", "width": 960, "height": 544, "fps": 24, "raw_frames": 243, "preview_sha256": "6963ac526cea6107a6b0708b03111776fcd7b521fde7c307fd24391b0907a3fe", "raw_sha256": "8ea96bc1e98de5d9b25fd6e21657768e036551f92c793130107ca0e7c2824c08", "preview_frames": 240, "preview_duration_seconds": 10.0, "raw_file": "https://github.com/lishunyang12/fasth3-lossy-ablation/releases/download/avatar-quality-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": 263.8453291330952, "raw_duration_seconds": 10.125, "full_decode": "PASS", "width": 960, "height": 544, "fps": 24, "raw_frames": 243, "preview_sha256": "bc0c3cbaeab6e21384eabfaf8d95e008474eca0d5b37fcfda741201fa5c042d3", "raw_sha256": "9f08fcd24681f7d8e71fbab3fcd0fcb84772aa38fd79fb58dde01b8972e98cff", "preview_frames": 240, "preview_duration_seconds": 10.0, "raw_file": "https://github.com/lishunyang12/fasth3-lossy-ablation/releases/download/avatar-quality-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": 6.992767889983952, "raw_duration_seconds": 10.125, "full_decode": "PASS", "width": 960, "height": 544, "fps": 24, "raw_frames": 243, "preview_sha256": "501fe72c82289ff4ea4d8ba494d49475678fb2d3e3e020c0d25feedfaaa86b10", "raw_sha256": "5854351ab922bf477201024e3f090240bff1fc7eec23bf5e903670559d6185d8", "preview_frames": 240, "preview_duration_seconds": 10.0, "raw_file": "https://github.com/lishunyang12/fasth3-lossy-ablation/releases/download/avatar-quality-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.", "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 今天陪你讀書也聊聊你的故事", "words": [ { "word": "你好", "start": 4.520000000000001, "end": 5.000000000000001, "probability": 0.4759526252746582 }, { "word": " 我是", "start": 5.000000000000001, "end": 5.48, "probability": 0.6711855530738831 }, { "word": "芝", "start": 5.48, "end": 5.84, "probability": 0.7062749713659286 }, { "word": "薇", "start": 5.84, "end": 5.94, "probability": 0.8936715126037598 }, { "word": " 今天", "start": 6.9, "end": 7.12, "probability": 0.992720365524292 }, { "word": "陪", "start": 7.12, "end": 7.3, "probability": 0.992363840341568 }, { "word": "你", "start": 7.3, "end": 7.5, "probability": 0.9828341007232666 }, { "word": "讀", "start": 7.5, "end": 7.6, "probability": 0.8002204596996307 }, { "word": "書", "start": 7.6, "end": 8.02, "probability": 0.9620335102081299 }, { "word": "也", "start": 8.02, "end": 8.52, "probability": 0.518186628818512 }, { "word": "聊", "start": 8.52, "end": 8.7, "probability": 0.99346524477005 }, { "word": "聊", "start": 8.7, "end": 8.86, "probability": 0.9723032116889954 }, { "word": "你的", "start": 8.86, "end": 9.12, "probability": 0.9830270409584045 }, { "word": "故", "start": 9.12, "end": 9.4, "probability": 0.9992664456367493 }, { "word": "事", "start": 9.4, "end": 9.6, "probability": 0.9996438026428223 } ], "asr_start": 4.520000000000001, "asr_end": 9.6, "speech_timing_description": "ASR 估计 4.52–9.60 秒（跨度 5.08 秒）", "visual_observation": "抽帧可见从读书转向镜头，4.5 秒时已张口；人物、衣服和桌面物件保持可辨认。说话时微笑和口部张幅较大。", "audio_signal": { "peak": 1.307417869567871, "samples_at_or_above_full_scale": 1713, "first_five_seconds_rms": 0.13370922207832336, "final_five_seconds_rms": 0.33520373702049255, "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": 4.520000000000001, "end": 5.000000000000001, "probability": 0.44559162855148315 }, { "word": " 我是", "start": 5.000000000000001, "end": 5.48, "probability": 0.7207739949226379 }, { "word": "芝", "start": 5.48, "end": 5.84, "probability": 0.7476098537445068 }, { "word": "薇", "start": 5.84, "end": 5.94, "probability": 0.8984614014625549 }, { "word": " 今天", "start": 6.9, "end": 7.12, "probability": 0.9930223822593689 }, { "word": "陪", "start": 7.12, "end": 7.3, "probability": 0.9927383959293365 }, { "word": "你", "start": 7.3, "end": 7.5, "probability": 0.9851257801055908 }, { "word": "读", "start": 7.5, "end": 7.62, "probability": 0.8166076242923737 }, { "word": "书", "start": 7.62, "end": 8.0, "probability": 0.9992983341217041 }, { "word": "也", "start": 8.0, "end": 8.52, "probability": 0.6091889142990112 }, { "word": "聊", "start": 8.52, "end": 8.7, "probability": 0.9844983220100403 }, { "word": "聊", "start": 8.7, "end": 8.86, "probability": 0.9824956655502319 }, { "word": "你的", "start": 8.86, "end": 9.12, "probability": 0.9911148548126221 }, { "word": "故", "start": 9.12, "end": 9.4, "probability": 0.9988242983818054 }, { "word": "事", "start": 9.4, "end": 9.56, "probability": 0.9993960857391357 } ], "asr_start": 4.520000000000001, "asr_end": 9.56, "speech_timing_description": "ASR 估计 4.52–9.56 秒（跨度 5.04 秒）", "visual_observation": "抽查帧中的转头时序、构图与持书动作和原精度四步接近，未看到明显的整体质量跃升；局部细节仍需看原始视频。", "audio_signal": { "peak": 1.3405354022979736, "samples_at_or_above_full_scale": 1511, "first_five_seconds_rms": 0.13235589861869812, "final_five_seconds_rms": 0.3318886160850525, "finite": true } }, { "id": "deployed-4step", "label": "上线四步配置", "recognized": "你好 我是芝薇 今天陪你讀書也聊聊你的故事", "words": [ { "word": "你好", "start": 4.52, "end": 5.0, "probability": 0.5006723999977112 }, { "word": " 我是", "start": 5.0, "end": 5.48, "probability": 0.6925023794174194 }, { "word": "芝", "start": 5.48, "end": 5.84, "probability": 0.7152585983276367 }, { "word": "薇", "start": 5.84, "end": 5.96, "probability": 0.9045699238777161 }, { "word": " 今天", "start": 6.2, "end": 7.12, "probability": 0.9938246011734009 }, { "word": "陪", "start": 7.12, "end": 7.3, "probability": 0.9933652281761169 }, { "word": "你", "start": 7.3, "end": 7.5, "probability": 0.9855194091796875 }, { "word": "讀", "start": 7.5, "end": 7.6, "probability": 0.7456988096237183 }, { "word": "書", "start": 7.6, "end": 8.12, "probability": 0.9652389287948608 }, { "word": "也", "start": 8.12, "end": 8.54, "probability": 0.5430667996406555 }, { "word": "聊", "start": 8.54, "end": 8.72, "probability": 0.994694173336029 }, { "word": "聊", "start": 8.72, "end": 8.88, "probability": 0.9724268317222595 }, { "word": "你的", "start": 8.88, "end": 9.12, "probability": 0.983274519443512 }, { "word": "故", "start": 9.12, "end": 9.4, "probability": 0.9991660118103027 }, { "word": "事", "start": 9.4, "end": 9.64, "probability": 0.9996378421783447 } ], "asr_start": 4.52, "asr_end": 9.64, "speech_timing_description": "ASR 估计 4.52–9.64 秒（跨度 5.12 秒）", "visual_observation": "抽查帧中保持了书斋构图、持书动作和对镜头说话；面部轮廓、口部动作与发丝局部细节和原精度组有细微差别。", "audio_signal": { "peak": 0.9693443775177002, "samples_at_or_above_full_scale": 0, "first_five_seconds_rms": 0.0929427519440651, "final_five_seconds_rms": 0.22446703910827637, "finite": true } } ], "visual_review_scope": "Observed six source frames at 0.5, 2.5, 4.5, 5.5, 7.5 and 9.5 seconds and full-resolution 7.5-second frames. No frame-by-frame lip-sync or perceptual ranking is claimed.", "name_transcription_note": "The requested name is 知微; Whisper transcribes the homophone 芝薇. This alone does not establish a spoken pronunciation error." }, "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" } }, "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." ] }