公开快照:保留原页面结论与全文prompt;只播放已授权模型原片。GT参考、首图和原始附件保留本地。

返回全部方法 30秒港口 · SANA / Solar / HY
返回全部实验

同一路线,三种原生接口

2026-10-14 · 独立三方法 campaign · SANA / Solar 动态视频,HY 静态回访

本轮完成:5条视频已验收,完整场景均未通过

完整结论与原因边界(附件保留本地) · 已测配置选择与复现证据(附件保留本地)

四条30秒动态原片(两方法各R0/R1)均已完成真实生成、远近端哈希、全帧解码、实际播放和独立视觉对照。唯一对照变量是追加公开相机时段,未来GT画面未进入模型。

方法本轮选择实际任务结论
SANA-WMR0/R1打平,保留R0未完成封闭候船室进出与回访;停止prompt支线。
SolarWM Stage2R1,局部改善恢复部分室外与入厅形态,但早入厅、晚出厅、原区身份未解决;完整路线失败。
HY-World2保留静态诊断配置,视觉失败中段黑空洞/模糊碎片,有限结构回见;完整静态路线失败,搬运N/A。

动态搬运多数离屏,结果记UNKNOWN;没有把相似箱子当作正确身份或交付证据。下面按模型查看原片与对应输入。

R0历史观察(R1见下方独立结论):像素结论:SANA主要维持开放棚廊;Solar约2秒已入室内且片尾仍室内。媒体有效不等于动作正确。独立逐阶段视觉报告(附件保留本地) · 浏览器播放回执(附件保留本地)

共同输入与完整动作 prompt

固定首图仅近似配准;不能声称照片中的每个箱、人、车与Three.js参考完全对应。入厅约13.7秒、出厅16.9秒、回到码头约25秒。部分搬运离屏,不把不可见动作自动判失败。

展开实际全文 prompt
A continuous photorealistic first-person view of this working harbor. The camera is an observer's eyes, separate from the orange-clad worker; do not add a visible camera person. At the opening, observe the wooden crate and cargo cart. The orange-clad worker's task is to load the wooden crate onto the cargo cart, push the loaded cart toward the berthed vessel, and unload the crate at the receiving area. The camera turns smoothly and walks at normal speed along the safe corridor into the waiting room, briefly looks at the floor tiles, blue seats, service counter and departure board, then turns and walks back out through the same doorway to revisit the harbor. Workers and cargo handling continue naturally off screen while the camera is inside; the crate, worker and cart need not appear in indoor views. On returning, look toward the cargo area and the berthed vessel. Keep a single uninterrupted shot, realistic materials and stable surroundings, with no teleportation, wall crossing or sudden cuts.
固定写实首图

参考/输入图像附件保留本地,公开版不包含GT。

SANA世界时间=(输出帧号+1)/16;Solar=输出帧号/16。逐帧对比需按世界时间对齐,不能直接同帧号。

验收合同(附件保留本地) · 输入独立审查(附件保留本地)

R1:只补公开时间段,其他条件固定

两方法原生CPU/token门已通过。Solar与SANA R1均已真实生成、回传并全帧解码;两方法已实际播放并独立对照完成。原动作prompt保持不变,仅在结尾补充:

Camera timing: outdoors 0-13.7s; waiting room 13.7-16.9s, turn back at 15.5s; outdoors again 16.9-30s, revisit the cargo area at 25s.

不修改首图、相机/K、seed、权重、采样或真实历史。只检验时序约束是否改善,不承诺恢复隐藏室内几何。没有收益即停止此支线。

SANA:R0 与时间提示 R1

R0 · 原动作提示

R0原始输入、prompt、配置与日志(附件保留本地) · R0哈希与解码(附件保留本地)

R1 · 仅追加时间段

R1:1280×704,480帧 / 16FPS / 30秒;生成与解码通过,实际播放结束。独立审阅:与R0在任务上打平,仍未进入封闭候船室或正确回访。保留R0原基线,停止prompt支线;搬运UNKNOWN。

SANA R1全文输入、prompt、配置、日志(附件保留本地) · SANA R1哈希与解码(附件保留本地)

Solar:R0 与时间提示 R1

R0 · 原动作提示

R0原始输入、prompt、配置与日志(附件保留本地) · R0哈希与解码(附件保留本地)

R1 · 仅追加时间段

R1:864×480,480帧 / 16FPS / 30秒;生成exit0,全帧解码0;独立审阅:R1有局部收益,6–10秒和末段恢复室外港口;但2秒早入厅未改,出厅晚约5–7秒,25秒原装卸区身份未确认。选R1为Solar本轮较好的已测配置,完整路线仍失败;搬运UNKNOWN。停止继续调prompt。

R1全文输入、prompt、配置、日志(附件保留本地) · R1本地哈希与解码(附件保留本地)
SANA 同刻对照图

参考/输入图像附件保留本地,公开版不包含GT。

参考/输入图像附件保留本地,公开版不包含GT。

SANA两版实际播放回执(附件保留本地) · R1独立视觉报告(附件保留本地) · Solar两版实际播放回执(附件保留本地)

Solar 同刻对照图

参考/输入图像附件保留本地,公开版不包含GT。

参考/输入图像附件保留本地,公开版不包含GT。

假设与停止条件(附件保留本地) · 冻结批量配置(附件保留本地)

本次实际运行的输入与参数

项目SANA-WM R0SolarWM Stage2 R0
输入图与相机写实首图;481个公开c2w+裁切后K;输出丢首帧物理RGB仅首帧;480个公开c2w;原生固定K,未消费自定义K
输出1280×704 / 480帧 / 16FPS864×480 / 480帧 / 16FPS;内部489帧尾裁9帧
seed与精度stage1=3407,refiner=42;BF16,attention FP32noise=42,selection=42;BF16
采样stage1 4步,CFG=1;refiner 3步;flow shift=8Stage2 NFE4;flow shift=5;本分支没有negative/unconditional CFG
真实生成历史stage1 KV/GDN状态;refiner滑窗KV11与初图sink持续6个chunk KV缓存,每chunk 3个latent;已生成clean chunk提交缓存
GPU与耗时单A80080GB;外层1190秒,native e2e485秒 / stream428秒;报告峰值49GB单A80080GB;外层405秒,提交全程409秒;采样净时未单独记录
tokensGemma输入408/506,未截断;native编码后BOS+后299位置;refiner202/1024UMT5 228/512,未截断;无system/chi前缀

这些是实际运行配置,无跨模型公平速度/质量排名。源码、权重revision及每项证据SHA见完整运行事实清单(附件保留本地)。

SANA 实际stage1前缀(官方chi模板)
Given a user prompt, generate an "Enhanced prompt" that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:
- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.
- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.
Here are examples of how to transform or refine prompts:
- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.
- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.
Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:
User Prompt: 

此前缀与公开prompt一起编码;不是另一次模型生成的改写。refiner直接编码公开prompt。SANA负prompt为空。

Solar 配置中的负prompt与实际消费区别
色调艳丽,过曝,静态,细节模糊不清,字幕,风格,作品,画作,画面,静止,整体发灰,最差质量,低质量,JPEG压缩残留,丑陋的,残缺的,形态畸形的肢体,静止不动的画面,杂乱的背景

以上文字存在于train配置,但本次Stage2推理未编码它。train real_guidance_scale=3也不是本次推理CFG,不把配置存在当实际消费。

批量运行与恢复已实测

R1两任务已由同一冻结批量入口真实运行完成。随后以相同配置、相同队列恢复,原进程身份及产物哈希保持一致;新增子进程为0,没有重复GPU推理。

实际恢复跳过回执(附件保留本地) · R1源码、权重、硬件、输入与耗时事实(附件保留本地)

SANA R1外层1008秒,native e2e441秒 / stream375秒;Solar外层400秒,采样净时未单独记录。不把外层等待时间当GPU忙时,也不据此估算费用。

HY-World2:仅静态空间回访

原生世界构建已实际完成(2941秒,含2000步高斯场景训练)。公开路线静态视频已真实渲染并回传(单卡126秒、480帧全帧解码通过),独立视觉验收失败;不评价搬箱,30秒播放长度不代表动态30秒预测。

输入为同一固定写实首图与下列静态全文prompt;不接收未来GT画面、深度或隐藏候船室布局。WorldNav自动探索只用来构建世界,另用公开相机路径渲染回访。

HY 实际静态全文prompt
Eye-level industrial cargo yard in daylight, matching the provided image. A worker in orange coveralls and a yellow hard hat is next to a metal flatbed cart and a large wooden crate on a pallet. Preserve the crate/cart placement and visible canopy posts, corrugated green warehouse walls, wet paved ground, cargo cranes and container stacks. Extend the surrounding static space with consistent surfaces and lighting, including connected pedestrian access toward a passenger waiting room outside the initial view. The observer is separate from the worker. Keep the initial view and visible layout stable; this is one static 3D environment, not animated cargo handling.

实际渲染锚点为本次生成场景的 panorama_L00_A0000,照片首视角配准未建立。路线只保留相对形状与转向,最大平移半径归一到0.25×native scene_scale;不能声称米制同世界路线复现。首个输入合同的配准设想不作为实际已完成证据。

HY源码、11组件权重来源、环境、参数与耗时(附件保留本地) · 固定HY输入包(附件保留本地) · 修正后的静态渲染配置(r02)(附件保留本地)

HY独立视觉报告(附件保留本地) · 实际播放回执(附件保留本地)

HY Pano实际正向包装/system与默认负向全文

positive

<|im_start|>system
Describe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>
<|im_start|>user
Picture 1: <|vision_start|><|image_pad|><|vision_end|>Create a **ERP** panoramic expansion of the provided image. Preserve the original style, lighting, and fine details seamlessly throughout the extended areas, extend according to: Eye-level industrial cargo yard in daylight, matching the provided image. A worker in orange coveralls and a yellow hard hat is next to a metal flatbed cart and a large wooden crate on a pallet. Preserve the crate/cart placement and visible canopy posts, corrugated green warehouse walls, wet paved ground, cargo cranes and container stacks. Extend the surrounding static space with consistent surfaces and lighting, including connected pedestrian access toward a passenger waiting room outside the initial view. The observer is separate from the worker. Keep the initial view and visible layout stable; this is one static 3D environment, not animated cargo handling. 8k UHD, masterpiece, razor-sharp details.<|im_end|>
<|im_start|>assistant

negative

<|im_start|>system
Describe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>
<|im_start|>user
Picture 1: <|vision_start|><|image_pad|><|vision_end|>低分辨率,低画质,模糊。杂乱的背景,结构扭曲,模糊纹理,物体融合。构图混乱。过度光滑,画面具有AI感。人脸畸形。巨大物体,巨大建筑,近景特写,近景压迫,比例失调。车,车辆。画面上方的树叶。<|im_end|>
<|im_start|>assistant
WorldNav原生模板与34份实际探索描述

根据本次生成图重新caption;不是未来GT,WorldStereo不直接消费原用户全文。负向车辆词与cart目标的冲突保留披露,不据此归因全部缺失。

native_templates/src/navi_utils.py(附件保留本地)

native_templates/src/vlm_utils.py(附件保留本地)

native_templates/traj_generate.py(附件保留本地)

worldgen/scene/render_results/reconstruct_0/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_0/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_4/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_4/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_6/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_6/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_7/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_7/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_9/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/reconstruct_9/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_cart_3/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_cart_3/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_green_container_0/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_green_container_0/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_green_container_1/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_green_container_1/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_green_container_2/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_green_container_2/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_pillar_4/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/target_pillar_4/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/view0/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/view0/traj2/traj_caption.json(附件保留本地)

worldgen/scene/render_results/view1/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/view1/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/view1/traj2/traj_caption.json(附件保留本地)

worldgen/scene/render_results/view2/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/view2/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/view2/traj2/traj_caption.json(附件保留本地)

worldgen/scene/render_results/wonder_0/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/wonder_0/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/wonder_1/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/wonder_1/traj1/traj_caption.json(附件保留本地)

worldgen/scene/render_results/wonder_6/traj0/traj_caption.json(附件保留本地)

worldgen/scene/render_results/wonder_6/traj1/traj_caption.json(附件保留本地)

832×480 · 480帧 / 16FPS / 30秒;真实渲染、媒体与实际播放通过。独立视觉FAIL:中段大面积黑空洞、墙面破碎模糊;25秒恢复部分码头结构,但完整路线不可用。保留失败并停止调尺度/重训。

343项原始证据、实际探索prompt、输入与几何(附件保留本地) · 本地哈希与全帧解码(附件保留本地) · 实际恢复跳过回执(附件保留本地)

评分参考 · 未提供给模型

参考/输入图像附件保留本地,公开版不包含GT。

事实与恢复

当前状态(附件保留本地) · 恢复入口(附件保留本地)

旧八方法和旧部署结果保留在总览各自区块,不与本轮混算。