Experimental report · Updated September 18, 2026

FastH3Precision Ablations

原版模型 · 有损消融、精度回退与组合首步保护

7 PROMPTS   ×   10 VARIANTS   =   70 VIDEOS

新参考图:三模型画质对比第二轮

换用指定链接中的书斋参考图,保持上一轮提示词、台词、种子和三套推理配置。三段 10 秒音视频,目标包含约 5 秒对镜头说话。

查看新参考图结果 → · 上一轮结果

数字人画质:原精度四步 / 原版 50 步 / 上线版本

同一知微场景、参考图、提示词与种子,三段 10 秒音视频;前半安静阅读,后半对镜头说话。支持同步播放、原始文件下载与同帧细节查看。

查看独立画质对比页面 →

After DiT: VAE, transfer, and media output

An English engineering report covering full-VAE decoding, gather and blend, GPU output fusion, pinned D2H, CPU video encoding, early AAC, and measured overlap.

Open the post-DiT optimization report →

Five-second 768p benchmarks

Original FastH3 and optimized serving: precision schedules, PR evaluations, complete HTTP-to-MP4 latency, and original audiovisual outputs.

Open the English benchmark report →

Abstract

以 FastH3 原版推理流程为共同基线,分别考察 DiT MXFP8、Sage Attention、VAE MXFP8 与 VAE NVFP4 四项有损替换。七组文本提示在相同 seed 与采样设置下各生成十个版本,构成 70 条视频的对照集。保留 O 与 Q/K/V/O 回退实验。此前两列分别测试 MXFP8 与 Sage 的首步保护:第 1 次 DiT forward 使用原版,第 2–4 次使用对应有损方法。由于实际只有 4 次 forward,这里保护的是 25% 的调用次数,后 75% 有损;没有将其标为 10% / 90%,也没有改变采样步数。本次在每个 case 顶部新增 MXFP8 + Sage 组合版:第 1 步全部 DiT 投影使用 BF16 与原版 VSA,第 2–4 步同时启用 DiT MXFP8 和 Sage。VAE 保持原版精度。两组 VAE 消融的输入 latent 与音频均与原版逐字节一致。本页呈现原始输出、完整 prompt 和执行核验记录,供观察空间细节、时序一致性与视听表现;未据此给出质量排名。

1280 × 704OUTPUT RESOLUTION
24 fpsFRAME RATE
15.04 sVIDEO DURATION
4 stepsDiT FORWARDS

01. Experimental Protocol

INDIVIDUAL AND COMBINED SCHEDULES

十组采用相同官方完整权重与上游源码,统一 seed 1101。每个实验组只引入下表所列替换;Q/K/V 分开计算,采样、稀疏选择、gate 与通信流程保持一致。

Variant替换范围数值格式投影数
Original原版推理流程原版精度—
DiT MXFP850 个 DiT block 的 Q / K / V / O 与两层 FFMXFP8 A8W8300
Sage AttentionVSA fine sparse attention kernelQK INT8 / PV FP8—
VAE MXFP836 个视频 VAE decoder block 的 Q / K / V / O 与两层 FFMXFP8 A8W8216
VAE NVFP4同一组视频 VAE decoder 投影NVFP4 A4W4216
DiT MXFP8 · O BF16相对 DiT MXFP8,仅 attention O 恢复原始 BF16;Q/K/V 与 FFN 保持 MXFP8O BF16 / 其余 MXFP8250 MXFP8 + 50 BF16
DiT MXFP8 · QKVO BF16在 O BF16 基础上继续回退 Q/K/V;全部 attention 投影用原始 BF16,仅 FFN 保持 MXFP8Q/K/V/O BF16 / FFN MXFP8100 MXFP8 + 200 BF16
MXFP8 · 首步原版第 1 个去噪步使用全部原版 BF16 投影,第 2–4 步启用完整 DiT MXFP8;VSA 始终原版1 步 Original + 3 步 MXFP8每步 300 个投影
Sage · 首步原版第 1 个去噪步使用原版 Triton VSA,第 2–4 步启用 Sage;投影始终原版 BF161 步 Original + 3 步 Sage每步 50 次 attention
MXFP8 + Sage · 首步原版第 1 步全部 Q/K/V/O、FFN 使用 BF16,attention 使用原版 Triton VSA;第 2–4 步同时启用完整 DiT MXFP8 与 Sage,VAE 保持原版1 步 BF16/VSA + 3 步 MXFP8/Sage每步 300 个投影 + 50 次 attention

共同设置:8 GPU · SP + FSDP · TP = 1 · VSA tile = 64 · sparsity = 0.9 · 5 个 sigma 网格点 / 4 次 DiT。输入 362 帧,原版混流后各文件为 361 帧;保留相同输出处理。首步保护按去噪时间顺序划分,不是每一步的前几层。组合版首步输入与输出 velocity 与此前两个首步原版对照逐字节一致,逐步调用审计确认两个有损方法同时在第 2 步启用。各组首次请求包含编译,运行计时不作为纯量化性能基准。

02. PSNR / SSIM vs Original

FULL FRAME · SAME TIMELINE

每条视频与同组 Original 在相同时间戳逐帧比较:1280 × 704、24 fps、361 帧,解码后的 8-bit YUV420p,不缩放、不裁剪、不搜索最佳时间偏移。PSNR 使用全视频加权 MSE,SSIM 使用 FFmpeg All;分量与逐帧结果可下载。指标包含编码影响,表示对原版的相似度,不等于主观画质排名。

VariantPSNR 均值(dB)SSIM 均值
MXFP8 + Sage · 首步原版19.76210.669343
DiT MXFP816.32940.566774
Sage Attention18.27530.622704
VAE MXFP839.41120.962671
VAE NVFP437.18100.952888
DiT MXFP8 · O BF1616.58810.572477
DiT MXFP8 · QKVO BF1617.36700.596293
MXFP8 · 首步原版19.75430.668078
Sage · 首步原版22.54440.758959

表中是 7 个 case 分数的算术平均,不是 pooled PSNR。Original 自检:PSNR ∞、SSIM 1。

完整指标 JSON / 逐 case CSV / 逐帧 PSNR / SSIM 日志 / 指标定义

03. Qualitative Comparisons

ORIGINAL OUTPUTS · NO RE-ENCODING

每组对应同一 prompt。顶部大卡片为本次新增组合版,下方保留此前九个版本。点击“同步播放”加载十个版本,用时间轴定位;“放大对照”可任选两个版本并排查看。默认静音,音频菜单可切换单一来源。浏览器同步适合定性观察,不保证逐帧锁定。

SAMPLE 01 / 07 · SEED 1101

Moonlit Teahouse月下茶屋

01-moon-teahouse
0.00 / 15.04 s
OriginalReference
13.1 MB · 361 framesMP4 ↓
PSNR ∞ dB · SSIM 1.0000
DiT MXFP8A8W8
12.9 MB · 361 framesMP4 ↓
PSNR 16.49 dB · SSIM 0.5946
Sage AttentionQK8 / PV8
13.4 MB · 361 framesMP4 ↓
PSNR 18.36 dB · SSIM 0.6615
VAE MXFP8A8W8
13.1 MB · 361 framesMP4 ↓
PSNR 40.13 dB · SSIM 0.9643
VAE NVFP4A4W4
13.5 MB · 361 framesMP4 ↓
PSNR 38.04 dB · SSIM 0.9567
DiT MXFP8 · O BF16O BF16
13.0 MB · 361 framesMP4 ↓
PSNR 16.66 dB · SSIM 0.5982
DiT MXFP8 · QKVO BF16FFN MXFP8
12.9 MB · 361 framesMP4 ↓
PSNR 17.27 dB · SSIM 0.6222
MXFP8 · 首步原版1 Original + 3 MXFP8
13.2 MB · 361 framesMP4 ↓
PSNR 20.50 dB · SSIM 0.7186
Sage · 首步原版1 Original + 3 Sage
13.0 MB · 361 framesMP4 ↓
PSNR 22.08 dB · SSIM 0.7702
完整 Prompt / Full generation prompt
integrated_multimodal_description: [Shot 1] Cinematic live-action fantasy inside a small wooden teahouse suspended beneath a colossal moon, shown in a symmetrical wide shot through open paper doors. An elderly tea master in an indigo robe, with silver hair pinned high and a warm gravelly Mandarin voice (S1), pours glowing tea into two porcelain cups while a young female radio astronomer in a cream flight suit, with a clear fast Mandarin voice (S2), adjusts a brass receiver beside her. The camera pushes in with small amplitude at slow speed as lunar dust drifts upward instead of falling, and the tea master says: <d>[Chinese] 月亮不是沉默,只是等我们听见。</d> [Shot 2] At 00:03.400, the camera cuts to a macro close-up of the tea surface. Concentric ripples become a precise star map, tiny points of light connect across the cup, and the receiver needle snaps toward a distant signal. The astronomer leans into the eyepiece, her face illuminated by moving constellations, and replies: <d>[Chinese] 信号锁定了。</d> [Shot 3] At 00:06.600, the shot cuts outside to a sweeping arc around the teahouse as the vast moon slowly opens into layered mechanical petals, revealing a luminous blue core. The two women step onto the balcony without changing clothes or positions relative to each other; the tea master raises her cup while the astronomer sends one narrow beam from the receiver into the core. The camera pulls out with large amplitude at slow speed until the teahouse becomes a warm point against the unfolding lunar structure.
overall_soundscape: Porcelain touches wood with delicate clicks above the soft hiss of boiling water and the papery flutter of the doors. A low receiver hum gains harmonics as the signal locks, followed by distant metal plates shifting across the moon and a soft pressure wave moving through airborne dust.
non_diegetic_music: Slow guzheng harmonics over sustained glassy synthesizer tones at 68 BPM. A restrained taiko pulse enters when the lunar petals open, then resolves into one long bowed-cymbal shimmer.
下载 prompt
SAMPLE 02 / 07 · SEED 1101

Dachshund at the Skatepark腊肠犬滑板场

02-dachshund-skatepark
0.00 / 15.04 s
OriginalReference
17.6 MB · 361 framesMP4 ↓
PSNR ∞ dB · SSIM 1.0000
DiT MXFP8A8W8
16.5 MB · 361 framesMP4 ↓
PSNR 14.46 dB · SSIM 0.5748
Sage AttentionQK8 / PV8
17.4 MB · 361 framesMP4 ↓
PSNR 16.28 dB · SSIM 0.6080
VAE MXFP8A8W8
17.6 MB · 361 framesMP4 ↓
PSNR 38.12 dB · SSIM 0.9567
VAE NVFP4A4W4
17.8 MB · 361 framesMP4 ↓
PSNR 35.87 dB · SSIM 0.9447
DiT MXFP8 · O BF16O BF16
17.2 MB · 361 framesMP4 ↓
PSNR 15.42 dB · SSIM 0.5904
DiT MXFP8 · QKVO BF16FFN MXFP8
17.4 MB · 361 framesMP4 ↓
PSNR 15.44 dB · SSIM 0.5890
MXFP8 · 首步原版1 Original + 3 MXFP8
17.7 MB · 361 framesMP4 ↓
PSNR 18.83 dB · SSIM 0.6687
Sage · 首步原版1 Original + 3 Sage
17.7 MB · 361 framesMP4 ↓
PSNR 22.65 dB · SSIM 0.7823
完整 Prompt / Full generation prompt
integrated_multimodal_description: [Shot 1] Photorealistic unedited video, a medium tracking shot follows the action with natural, imperfect jitter under flat, overcast daylight. Finn, a mid-30s East Asian man with short black hair falling loosely to the sides and a lean athletic build, wears a black crew-neck t-shirt with the white wordmark "el.cine" printed across the chest, light beige chino trousers, and black shoes. He balances sideways in a skater's stance directly on the back of Frank, a small chocolate-brown dachshund with a long low body, stubby legs, and floppy ears. They balance at the top edge of a smooth, empty concrete skatepark bowl, bordered by a metal rail, with green hills and a tree line visible under a grey cloudy sky. A sparse crowd of casually dressed spectators sits along the far rim. Finn's black shoes press flat into Frank's fur, with absolutely nothing between them, as Finn makes small, constant ankle corrections to keep his balance while the dog's tail wags. They tip forward and drop down the steep curved wall of the bowl, plunging downwards as Frank's stubby legs blur into motion, carrying Finn effortlessly in a long, fast arc across the concrete floor. Finn crouches low, his knees pumping and arms extended wide with a dead serious expression, while the spectators turn their heads in unison to follow the ride, some raising phones to film. Racing up the opposite curved wall, man and dog fly off the top edge into the open air. Mid-flight, Finn reaches down and grabs Frank's long body with one hand; the dachshund hangs suspended for a beat, all four paws paddling in the air with his tongue out, while the crowd visibly erupts in applause. On the descent, Finn places the dog perfectly back under his feet, his shoes landing square on the furry back, and they slide down the transition. Leveling out on the flat bottom, they glide to a complete stop. Finn hops off, stepping one foot down from the dog's back directly onto the bare concrete, and raises both hands toward the cheering crowd. Frank immediately scampers in a joyful circle around Finn's legs, his tail held high, leaving nothing but empty concrete where they rode.
overall_soundscape: A continuous rush of wind and the pronounced, rapid clatter of dog claws scrabbling against smooth concrete dominate the foreground, accelerating into a loud, frantic clicking as the pair gains speed down the bowl. A distinct physical thud marks their landing back on the concrete surface, followed immediately by the clearly heard, rhythmic panting of the dachshund. Throughout the sequence, the loud, clear whoops, cheering, and enthusiastic clapping of the surrounding crowd rise and fall in the background, entirely devoid of any rolling wheel or bearing hum sounds.
non_diegetic_music: N/A
下载 prompt
SAMPLE 03 / 07 · SEED 1101

Rain in the Attic阁楼雨景

03-attic-rain
0.00 / 15.04 s
OriginalReference
13.2 MB · 361 framesMP4 ↓
PSNR ∞ dB · SSIM 1.0000
DiT MXFP8A8W8
13.0 MB · 361 framesMP4 ↓
PSNR 17.23 dB · SSIM 0.5935
Sage AttentionQK8 / PV8
13.0 MB · 361 framesMP4 ↓
PSNR 19.63 dB · SSIM 0.6590
VAE MXFP8A8W8
13.3 MB · 361 framesMP4 ↓
PSNR 41.07 dB · SSIM 0.9706
VAE NVFP4A4W4
13.5 MB · 361 framesMP4 ↓
PSNR 38.65 dB · SSIM 0.9625
DiT MXFP8 · O BF16O BF16
13.4 MB · 361 framesMP4 ↓
PSNR 18.12 dB · SSIM 0.6052
DiT MXFP8 · QKVO BF16FFN MXFP8
13.1 MB · 361 framesMP4 ↓
PSNR 19.68 dB · SSIM 0.6531
MXFP8 · 首步原版1 Original + 3 MXFP8
13.4 MB · 361 framesMP4 ↓
PSNR 21.33 dB · SSIM 0.7044
Sage · 首步原版1 Original + 3 Sage
13.0 MB · 361 framesMP4 ↓
PSNR 24.21 dB · SSIM 0.7926
完整 Prompt / Full generation prompt
integrated_multimodal_description: [Shot 1] 3D-animated, indie art-house animation aesthetic with semi-realistic painterly rendering, a static medium shot. A young woman in her early 20s (S1) features slender proportions, soft skin with heavy subsurface scattering, freckles across her nose and cheeks, and a natural blush. Her thick, wavy auburn hair is rendered in painterly clumps with a warm rim light. She wears a soft mustard-yellow knit cardigan over a cream slip dress, with bare feet, and holds a chipped ceramic mug in her right hand. She sits curled in the window seat on the left of a warm, cluttered attic bedroom. The room has worn wooden floors, stacks of books, a rumpled linen bed in the midground, dried flowers in jars, and unlit string lights on the wall. Outside the closed window, a hazy grey-gold sky looms over terracotta rooftops. Hazy diffused lighting creates a heavy soft bloom, with warm lamplight illuminating the right side of her face and shallow depth of field blurring the background. She reads a book on her lap, her shoulders relaxed, then pauses and glances up toward the glass as fat raindrops streak the pane. [Shot 2] At 00:04.000, the camera cuts to a close-up and pushes in. A slow smile spreads across the woman's (S1) freckled face. She lowers the ceramic mug out of frame, then uses both hands to push the window fully open. Warm damp air rushes in, lifting loose strands of her hair. In the background, pages of an open book flutter wildly in the breeze. [Shot 3] At 00:07.000, the camera cuts to a wide static shot. The woman (S1) leans right out into the rain through the open window frame on the left, her eyes closed, face tilted up, and arms spread on the wooden sill. The heavy rain instantly soaks her hair and cardigan, darkening the fabric, while her shoulders drop in a deeply relaxed posture. [Shot 4] At 00:09.000, the camera transitions to a close-up and pushes in slowly. Rendered in slow motion, heavy raindrops burst against her upturned face and shoulders. Each droplet catches the hazy grey-gold light, with mist and raindrops suspended in the light beams. Her hair clings to her wet cheeks, and mist drifts off the rooftops behind her. Her chest rises visibly as she takes a single deep breath. [Shot 5] At 00:12.000, the camera cuts back to a medium static shot inside the bedroom. Back in real time, the woman (S1) laughs warmly, shaking her head to send a spray of water droplets flying from her hair. She pulls her soaked body back inside, grinning broadly, and flops backward onto the rumpled bed in the center-right of the frame with her arms spread wide. The camera holds on her happy, dripping face as she sinks into the mattress.
overall_soundscape: Quiet ambient room tone and a distant, low thunder rumble. The distinct tap and patter of the first raindrops hitting the glass pane are clearly heard, transitioning into a loud splash and wind hiss as the window slides open. The sharp, rapid flutter of paper pages flapping is prominent in the foreground. The audio then drops into a hushed, muffled hiss of rainfall, punctuated by a deep, clearly audible inhale. This abruptly bursts back into the loud, steady drumming of heavy rain on terracotta tiles, a bright, cheerful chuckle from the woman, and a pronounced fabric thump as her body hits the mattress.
non_diegetic_music: Solo acoustic guitar, slow tempo, picking a delicate melody with a gentle atmospheric cello pad underneath, no percussion, remaining steady throughout.
下载 prompt
SAMPLE 04 / 07 · SEED 1101

Doorways Between Worlds穿门切换世界

04-doorway-worlds
0.00 / 15.04 s
OriginalReference
12.7 MB · 361 framesMP4 ↓
PSNR ∞ dB · SSIM 1.0000
DiT MXFP8A8W8
13.2 MB · 361 framesMP4 ↓
PSNR 16.24 dB · SSIM 0.5628
Sage AttentionQK8 / PV8
12.9 MB · 361 framesMP4 ↓
PSNR 18.22 dB · SSIM 0.6326
VAE MXFP8A8W8
12.7 MB · 361 framesMP4 ↓
PSNR 40.45 dB · SSIM 0.9671
VAE NVFP4A4W4
13.2 MB · 361 framesMP4 ↓
PSNR 38.54 dB · SSIM 0.9603
DiT MXFP8 · O BF16O BF16
13.4 MB · 361 framesMP4 ↓
PSNR 15.91 dB · SSIM 0.5575
DiT MXFP8 · QKVO BF16FFN MXFP8
13.0 MB · 361 framesMP4 ↓
PSNR 17.10 dB · SSIM 0.6140
MXFP8 · 首步原版1 Original + 3 MXFP8
13.0 MB · 361 framesMP4 ↓
PSNR 20.28 dB · SSIM 0.7021
Sage · 首步原版1 Original + 3 Sage
12.6 MB · 361 framesMP4 ↓
PSNR 23.15 dB · SSIM 0.8031
完整 Prompt / Full generation prompt
integrated_multimodal_description: [Shot 1] Cinematic, wide tracking shot positioned at waist height, maintaining a constant perpendicular side-profile view that captures the full body of the subject. A 20-year-old Neapolitan woman (S1) walks steadily from left to right across a sunlit Naples apartment, her shoulders shifting naturally with each stride and her gaze fixed calmly forward. She has a dark brown chin-length bob parted in the middle with the ends flipped slightly out, blue eyes, a straight nose with a slightly rounded tip, a clear jawline, natural rosy lips, small facial moles, slight redness around her nostrils, and realistic skin texture. She wears a black sleeveless turtleneck, loose ivory linen trousers, small gold hoop earrings, and black ankle boots. The room around her features warm afternoon sunlight, wooden shutters, colorful ceramic tiles on the floor, and delicate lace curtains. She reaches a standard wooden door on the right wall, grasps its metal handle, and pushes it open. She steps over the threshold, her arms swinging lightly at her sides. The camera trucks right, gliding seamlessly through the wooden doorframe directly behind her. Without any break in the camera's motion, the environment beyond the doorframe changes entirely, and she immediately emerges from an identical wooden door on the left side of a new room—an underwater glass palace. The wooden door slowly swings shut behind her. Outside the massive glass walls, large schools of fish and glowing jellyfish drift through deep blue water, illuminated by shimmering shafts of light. She does not pause, her expression remaining perfectly neutral as she maintains her exact pace and posture across the glass-floored room. Reaching the far right wall, she grasps the handle of another identical wooden door, pushes it open, and steps through. The camera smoothly trucks through this second doorframe. As the camera crosses the threshold, the setting transitions flawlessly into a futuristic neon metropolis. She emerges from the identical door on the left side of this new room, framed against large windows streaked with rain, where shifting neon lights and flying cars illuminate the background. She continues her steady walk to the right as the continuous take holds on her profile.
overall_soundscape: Distinct rhythmic footsteps on ceramic tiles, accompanied by the loud clack of a metal door handle and a pronounced wooden creak as the first door opens. A soft, audible whoosh marks the transition through the doorway, followed by a loud click as the door shuts behind her. The acoustic space changes to feature a deep, dominant underwater hum and crisp, clear footsteps echoing on a hard floor. Another distinct click and loud creak sound as she opens the second door, leading to a smooth whooshing transition. The clip concludes with the audible patter of rain against glass and the low, sweeping hum of passing flying cars, layered over her steady footsteps.
non_diegetic_music: Subtle, continuous synthesizer drone, slow tempo, ethereal and atmospheric, providing a steady surreal undercurrent to the spatial transitions without dominating the physical mix.
下载 prompt
SAMPLE 05 / 07 · SEED 1101

Ronin in the Void虚空女浪人

05-ronin-void
0.00 / 15.04 s
OriginalReference
11.5 MB · 361 framesMP4 ↓
PSNR ∞ dB · SSIM 1.0000
DiT MXFP8A8W8
11.7 MB · 361 framesMP4 ↓
PSNR 18.03 dB · SSIM 0.7147
Sage AttentionQK8 / PV8
11.7 MB · 361 framesMP4 ↓
PSNR 21.52 dB · SSIM 0.7868
VAE MXFP8A8W8
11.6 MB · 361 framesMP4 ↓
PSNR 40.84 dB · SSIM 0.9649
VAE NVFP4A4W4
11.7 MB · 361 framesMP4 ↓
PSNR 38.91 dB · SSIM 0.9582
DiT MXFP8 · O BF16O BF16
11.6 MB · 361 framesMP4 ↓
PSNR 18.20 dB · SSIM 0.7221
DiT MXFP8 · QKVO BF16FFN MXFP8
11.3 MB · 361 framesMP4 ↓
PSNR 19.74 dB · SSIM 0.7522
MXFP8 · 首步原版1 Original + 3 MXFP8
11.7 MB · 361 framesMP4 ↓
PSNR 21.27 dB · SSIM 0.7816
Sage · 首步原版1 Original + 3 Sage
11.6 MB · 361 framesMP4 ↓
PSNR 25.06 dB · SSIM 0.8553
完整 Prompt / Full generation prompt
integrated_multimodal_description: [Shot 1] 2D-animated, a wide shot transitioning to a medium shot from a low camera angle as the camera performs an arc shot at slow speed. An infinite black void is filled with thick, rolling grey smoke, drifting black ink haze, and glowing red ember ash mixed with floating crimson thread particles. A female ronin assassin (S1) stands motionless in the center. She has pale skin, long flowing black hair, and wears layered black silk robes with oversized sleeves, tightly cinched with vibrant red waist cords. A sleek black katana rests in a black scabbard at her hip. Cold white volumetric lighting cuts through the dark void, heavily accented by the deep red ambient glow of the floating embers. She remains perfectly still in a restrained stance, her right hand resting lightly on the katana's hilt, as the smoke rolls smoothly across the unseen floor beneath her feet. [Shot 2] At 00:03.500, the camera cuts to a close-up, pushing in at fast speed toward the assassin (S1). Her eyes narrow sharply. In a sudden blur of motion, she executes a rapid quickdraw slash, pulling the sleek black katana from its scabbard. A glowing crimson slash trail erupts from the metallic blade, slicing the rolling smoke in half. She drops into a low stance and dashes forward aggressively. Her layered black silk robes and oversized sleeves snap backward, demonstrating heavy, delayed fabric physics, while her long black hair whips fiercely behind her head. [Shot 3] At 00:06.200, the shot transitions to a medium wide shot, panning right at fast speed to follow the action. The assassin (S1) leaps into the air, performing rapid airborne rotational strikes. The sleek black blade reflects the cold white light, leaving overlapping ink-like energy arcs and brilliant crimson thread-shaped distortions in its wake. The glowing sword impacts the surrounding black ink haze, scattering bright red embers violently outward in a spray. Her body twists powerfully in mid-air, the red waist cords and black robes trailing dramatically with the cinematic motion. [Shot 4] At 00:09.800, the camera cuts to a wide shot, zooming in at fast speed toward the assassin (S1). She lands on the unseen floor and unleashes a massive, devastating crimson-infused finishing slash straight ahead. A giant, jagged crimson energy arc erupts from the blade, visually fracturing the infinite black void into shattered geometric shards of dark matter and blinding white light. The rolling smoke, flying debris, scattered ember particles, and shattered crimson thread energy suddenly freeze in extreme cinematic slow motion. The assassin (S1) is captured in a striking dark silhouette against the glowing red and white fracture lines, holding her extended katana follow-through pose motionless as the glowing threads remain perfectly suspended in the air around her.
overall_soundscape: A low, rumbling wind and a soft hiss of swirling smoke mix with the subtle crackle of burning embers. This ambient tension is shattered by a loud, metallic scrape as a blade is quickly drawn, followed immediately by an aggressive, tearing whoosh and the rapid, heavy thud of quick footsteps. Multiple sharp, high-pitched swishes and crisp, clearly heard fabric snaps dominate the foreground as the spinning blade slices the air. A deafening, explosive boom and a deep, resonant bass crash mark the massive final impact, abruptly transitioning into a sustained, high-frequency ringing that stretches out over the frozen slow-motion sequence.
non_diegetic_music: A fast-paced score featuring heavy, driving taiko drums and a rapid, solo shamisen melody, steadily increasing in rhythmic density and dynamic volume, before abruptly cutting off into a sudden, sustained low-frequency synthetic sub-bass tone with no additional instrumentation.
下载 prompt
SAMPLE 06 / 07 · SEED 1101

Anime Maid Livestream动漫女仆直播

06-anime-maid-livestream
0.00 / 15.04 s
OriginalReference
20.8 MB · 361 framesMP4 ↓
PSNR ∞ dB · SSIM 1.0000
DiT MXFP8A8W8
21.5 MB · 361 framesMP4 ↓
PSNR 15.94 dB · SSIM 0.4284
Sage AttentionQK8 / PV8
20.9 MB · 361 framesMP4 ↓
PSNR 16.56 dB · SSIM 0.4656
VAE MXFP8A8W8
20.8 MB · 361 framesMP4 ↓
PSNR 37.60 dB · SSIM 0.9610
VAE NVFP4A4W4
21.3 MB · 361 framesMP4 ↓
PSNR 35.06 dB · SSIM 0.9506
DiT MXFP8 · O BF16O BF16
21.1 MB · 361 framesMP4 ↓
PSNR 15.73 dB · SSIM 0.4298
DiT MXFP8 · QKVO BF16FFN MXFP8
20.5 MB · 361 framesMP4 ↓
PSNR 15.82 dB · SSIM 0.4306
MXFP8 · 首步原版1 Original + 3 MXFP8
20.6 MB · 361 framesMP4 ↓
PSNR 17.82 dB · SSIM 0.5239
Sage · 首步原版1 Original + 3 Sage
20.8 MB · 361 framesMP4 ↓
PSNR 20.31 dB · SSIM 0.6460
完整 Prompt / Full generation prompt
integrated_multimodal_description: [Shot 1] Mixed-media style blending a photorealistic setting with a 2D-animated character, the camera maintains a POV medium close-up shot that shakes slightly, incorporating subtle autofocus breathing, tiny exposure adjustments, realistic high ISO grain, and slight JPEG compression artifacts. An English livestream UI overlays the view, displaying a red "LIVE" badge, a viewer count, floating pink hearts, a follow button, a comment input bar, and rapidly scrolling text reading "omg she’s actually anime", "best maid cafe ever", "bro this feels unreal", "the ketchup heart", "she looks so sleepy lol", "I’d go every day", and "why does this look fake and real at the same time". In the immediate foreground, an off-screen customer's hand grips a silver spoon over a worn wooden table edge beside a condensation-covered water glass and a partially visible black bag. The background reveals a cramped, realistically unkempt Akihabara maid café inside an old multi-tenant building, lit by fluorescent ceiling lights that cause occasional highlight clipping. Laminated menu POPs, idol posters, a drink bar machine, a cluttered register counter tangled with cables, and stacked plastic trays fill the dimly lit space. An on-screen young female 2D-animated maid (S1) steps into the midground, featuring delicate facial proportions, soft sleepy pink eyes, and long blonde hair tied with pink ribbons. She wears a pastel pink frilly maid dress with a white lace headpiece and a matching apron. She leans forward, her dress rustling gently, and places a plate of yellow omurice onto the table. Retrieving a red plastic bottle from her pocket, she slowly squeezes it to draw a ketchup heart onto the egg. She quietly confirms the order with a soft nod, pausing to look directly at the smartphone camera with a heavy-lidded, sleepy expression. Raising her hands chest-high into a half-hearted heart gesture, the maid (S1) quietly murmurs, <d>[Japanese] moe moe kyun</d>. She then lazily drops her hands back to her sides, entering a tiny, awkward pause. She blinks slowly, taking small, idle breaths as her blonde hair, pastel apron, and pink ribbons sway subtly, standing perfectly still like an exhausted worker surviving a long shift.
overall_soundscape: A steady, low ambient hum of a drink refrigerator and fluorescent lights establishes the cramped cafe space, punctuated by muffled, distant chatter. In the foreground, a distinct ceramic clatter rings out as the omurice plate makes contact with the wooden table. This is followed by a clear, squelching wet sputter from the plastic ketchup bottle being squeezed over the food. Throughout the clip, the pronounced rustling of stiff frilly fabric and crisp apron ties can be clearly heard as the maid shifts her weight and moves her arms, ending with her quiet, close-proximity voice and the subtle, rhythmic intake of her idle breathing.
non_diegetic_music: N/A
下载 prompt
SAMPLE 07 / 07 · SEED 1101

Ninja Obstacle Course忍者障碍赛

07-ninja-obstacle-course
0.00 / 15.04 s
OriginalReference
19.2 MB · 361 framesMP4 ↓
PSNR ∞ dB · SSIM 1.0000
DiT MXFP8A8W8
19.2 MB · 361 framesMP4 ↓
PSNR 15.91 dB · SSIM 0.4985
Sage AttentionQK8 / PV8
19.4 MB · 361 framesMP4 ↓
PSNR 17.35 dB · SSIM 0.5454
VAE MXFP8A8W8
19.2 MB · 361 framesMP4 ↓
PSNR 37.67 dB · SSIM 0.9540
VAE NVFP4A4W4
19.2 MB · 361 framesMP4 ↓
PSNR 35.20 dB · SSIM 0.9373
DiT MXFP8 · O BF16O BF16
19.5 MB · 361 framesMP4 ↓
PSNR 16.08 dB · SSIM 0.5041
DiT MXFP8 · QKVO BF16FFN MXFP8
19.6 MB · 361 framesMP4 ↓
PSNR 16.52 dB · SSIM 0.5129
MXFP8 · 首步原版1 Original + 3 MXFP8
19.6 MB · 361 framesMP4 ↓
PSNR 18.25 dB · SSIM 0.5772
Sage · 首步原版1 Original + 3 Sage
19.1 MB · 361 framesMP4 ↓
PSNR 20.34 dB · SSIM 0.6631
完整 Prompt / Full generation prompt
integrated_multimodal_description: [Shot 1] Cinematic cel-shaded 3D anime, a medium tracking shot moves horizontally right alongside a young female ninja warrior athlete featuring white skin, sharp elegant features, fierce intense eyes, and long dark hair tied in a tight, high ponytail. She has a lean, toned athletic build and wears a black fitted crop top, red fitted athletic shorts, black shoes with red accents, black wristbands with a red stripe, and gold hoop earrings. She sprints and leaps at full speed across five small, angled, floating red platforms that bounce slightly under her weight. Bright midday natural lighting with heavy, hard-edge shadows illuminates the outdoor competition course against a clear blue sky. A massive crowd of spectators lines both sides of the course, their colorful clothes blurring in the vibrant background as they cheer. [Shot 2] At 00:03.200, the camera cuts to a low-angle tracking shot that follows the athlete from a side profile as she launches herself toward a set of suspended red flying bars. Her body fully extends in mid-air, her dark ponytail whipping behind her. She grabs the first bar, her toned arm and core muscles flexing with heavy cel-shaded definition, and powerfully swings herself forward to catch the next progressively spaced bar. The cheering crowd raises their hands in excitement in the deep background. [Shot 3] At 00:06.500, the shot transitions to a wide tracking shot positioned slightly above the course. The athlete drops from the final flying bar, landing perfectly onto a series of hexagonal floating steps hovering just inches above a shimmering blue pool of water. Each step wobbles slightly upon impact as she navigates the varying heights and angles with precise, rapid footwork. The bright midday sun reflects off the water's surface, while the flanking crowd leans over the barriers, their animated faces tense with anticipation. [Shot 4] At 00:10.000, the camera cuts to a low-angle tracking shot from behind as the athlete sprints at full speed toward the final obstacle: a massive, curved blue warped wall. She hits the steep incline without breaking stride, her legs pumping furiously as she scales the vertical curve. As her upward momentum slows near the top edge, she extends her arms upwards, her fingertips grazing the rim. [Shot 5] At 00:12.500, the camera cuts to a medium close-up, and the lens zooms out into a wide low-angle hero shot. Her black-and-red-accented hands grip the top edge of the warped wall, and she pulls her body up over the ledge with a fierce exertion. She immediately reaches out and violently slams her palm down on a large red buzzer mounted on a sleek pedestal. A massive explosion of thick, swirling red smoke bursts outward from the buzzer's base. The camera holds as she stands tall at the summit, her arms raised in victory with a fierce, confident smile on her face. Her ponytail settles against her back, framed by the billowing red smoke and the crisp blue sky, while the massive crowd below erupts in absolute frenzy.
overall_soundscape: A loud, continuous roar of a cheering crowd dominates the ambient background, swelling and falling with the physical action. Rapid, heavy footsteps thud against bouncy platforms, transitioning into sharp, metallic clanks and a distinct whoosh of air as the athlete swings between the flying bars. The dull, hollow clatter of shoes landing on wobbling plastic steps mixes with the subtle slosh and splash of water underneath. A quick sequence of rapid, gripping footfalls slaps against the curved wall surface, leading to a hard physical grunt and the loud scrape of hands gripping the top edge. A sharp, heavy plastic smack rings out as the buzzer is hit, instantly followed by a deep boom and a loud hissing burst of compressed gas from the smoke explosion, while the crowd noise aggressively spikes into a deafening, frenzied roar.
non_diegetic_music: High-energy, fast-paced electronic dance music with a driving, heavy bass drum beat and aggressive synthesizer stabs, remaining loud and relentlessly rhythmic throughout the entire sequence to underscore the intense athletic momentum.
下载 prompt

04. Validation & Reproducibility

VALIDATION STATUS: PASS
70 / 70 视频完整解码PASS
560 份 rank / request 调用审计PASS
十组共同配置一致PASS
上游源码与计算运行环境一致PASS
两组 VAE 的输入 latent 与原版一致EXACT MATCH
两组 VAE 解码音频与原版一致EXACT MATCH

上游 FastVideo commit:3196835913ce9e97b23888b32643a1207cc5ee70
模型:FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
Checkpoint revision:b65818d41939b5085451074fe8ca8b799f8d4921

上述核验确认执行与产物一致性;kernel qualification 是数值与接口检查,视频质量需另行评估。所有 MP4 保留原始字节,SHA-256 可在核验记录中查阅。公开配置中的机器绝对路径已省略。

完整核验记录   /   执行配置与结果   /   Kernel qualification   /   方法说明   /   Prompt token counts

双视频对照

0.00 / 15.04 s

用时间轴定位同一时刻;可使用播放器全屏检查细节。按 Esc 关闭。