AI 探索者 · 独立开发者 AI视频工作台:reelbenchai.com

有朋友问我 MiniMax H3 是如何固定角色的 可能大部分人只知道:H3 可以直接用参考图(Ref2VA)锁定人物。但这个方法会占用 ref 参考参数。最多是9个图像参数。 现在有个新方法 RefMod 可以提前生成角色数据,以获得不错的角色固定效果,还不占参考图参数。 github.com/Luisacaotica/Comf…
14
63
459
25,884
我胡汉三又回来了! 查了下网页版 Claude 能访问,Claude Code出现403 就是因为当前网络环境被cloudflare这个大善人给拦下了 具体是咋好的,我也说不清楚。不过有几点可以分享下: - 每调整一次网络参数,不要直接尝试去登录,可以先命令行用 curl -I api.anthropic.com 做测试,看到403就可以尝试登录。如果看到404就是不行,继续调整 - 上述命令跑出403之后,用 codex 清理 Claude Code 的登录日志 继续猥琐发育,看看还能苟多久😂
别吓我啊,重启了下电脑。然后让我重新登录! 结果我网页登录成功了,但Claude Code还是403!
4
3
1,202
别吓我啊,重启了下电脑。然后让我重新登录! 结果我网页登录成功了,但Claude Code还是403!
2
1
2,219
Codex 中新会话出现 Couldn’t load workspace settings 这种提示的,可以通过 codex cli 绕一下。 从cli进入项目目录,然后随便发个对话再回到app就可以用已经存在的会话继续了。
2
1,105
既然大家爱看,就再做个评测视频。 这次不分享提示词,分享全流程的创作思路。 剧本: - 在抖音搜00后,选自己觉得有趣的剧情情节,总结成一句话。可以复制链接到豆包,让豆包总结。 - 把上边总结的内容发给 Agent,让他放在古代场景进行二创并写出完整的剧本。检查核对剧本让Agent修改到自己满意 提示词: 基于 MiniMax H3 官方提示词 skill,二次封装个自己的提示词 skill,具体步骤: - 之前「x.com/eternityspring/status/…」这个推文有链接,复制给Agent - 让Agent参考 ComfyUI-H3-Continuum 的提示词要求重新封装一个新的skill 如果是做写实风格的视频,建议把评论区的画质提升提示词发给Agent,让他融入到新skill中。
用 MiniMax H3 + ComfyUI-H3-Continuum 做了个 110 秒的古装穿越短片《穿越娶亲,礼单炸了》 整条分三次生成,分别是 30 秒、40 秒、40 秒,最后拼起来。H3 单次最长 15 秒,每次里面靠 Continuum 把 10 秒一段接着上一段最后 22 帧往下续。 整片都是一次出,一刀没剪。只用Claude加了个字幕。
17
9
133
11,763
写实风 MiniMax H3 画质提示词: 正向 photorealistic cinematic shot, natural soft diffused lighting, realistic skin subsurface scattering, natural skin texture, no oily skin, subtle skin pores, natural eye catchlight, soft shadow transition, physically based rendering, PBR, real fabric texture, natural material reflection, fine film grain, shallow depth of field, smooth motion, stable facial features, consistent face across frames, no painted texture, no thick brush strokes, no oil painting effect, live action, real world scene, natural color grading, moderate contrast 反向 oil painting, thick paint, brush strokes, painted skin, plastic skin, shiny oily face, wax figure, mannequin, cartoon, anime, illustration, 3d render, CGI, harsh sharp shadow, over saturated color, flickering face, distorted facial features, blurry details, ugly, deformed, disfigured
1
5
1,083
看到别人的片子节奏炸裂、运镜丝滑,你想拆开看看到底怎么做到的——然后你拖着进度条分段,记景别,看运镜,一下午没了 借助我开源的 reelbench-skills 系列 skill 可以帮你省下很多时间。项目历时1个月已经1k star了 github.com/eternityspring/re… video-shots · 拉片 丢一条成片进去,自动拆出逐镜头分析表:时长、景别、类别、运镜、画面、节奏,全在一张表里。切点和时长由 ffmpeg 实测,不让模型瞎猜数字;模型只负责它真正擅长的语义判断。中间夹了 15 道质量门逐条对账。 x.com/eternityspring/status/… video-sync · 拉片二创 光有报告不够,还要能“带着分镜看片”。video-sync 把拉片数据直接贴回原视频——画面一边,分镜信息一边,镜头切了信息跟着切,镜头表自动滚动高亮当前这一镜。还加了八个维度的节奏分析:钩子 / 铺垫 / 递进 / 重音 / 转折 / 兑现 / 换气 / 收口,好片子为什么留人,一眼看穿。以前几小时的手动活,现在几分钟跑完。 x.com/eternityspring/status/… video-scrub · 去元数据 AI 生成的视频,元数据里可能带着 GPS、设备型号、账号 ID、软件标识、遥测数据这些你肉眼看不见的东西,文件一传播就跟着走。video-scrub 只保留画面和声音,把不该带走的元数据清理干净再输出——发布、分发、交付都更放心。处理完直接拿去用。 x.com/eternityspring/status/…
用 MiniMax H3 做视频发抖音的 一定要用我开源的reelbench-skills/video-scrub去掉视频元数据 github.com/eternityspring/re… 看我抖音号发了5个作品,前三个是没有去元数据,发布后一周内的播放都是没有破500的。 最近两个视频都用了skill去掉元数据,播放都破万了。 昨天x评测的视频同步在抖音已经5w+播放了 x.com/eternityspring/status/…
14
21
157
8,337
本来只是个评测视频,昨天心血来潮随手发了下别的平台,结果抖音跑了30w播放…… 说实话我一直挺懒的,偶尔同步内容到别的平台。要这么玩的话,得认真坐下别的平台了🤣
用 MiniMax H3 + ComfyUI-H3-Continuum 做了个 110 秒的古装穿越短片《穿越娶亲,礼单炸了》 整条分三次生成,分别是 30 秒、40 秒、40 秒,最后拼起来。H3 单次最长 15 秒,每次里面靠 Continuum 把 10 秒一段接着上一段最后 22 帧往下续。 整片都是一次出,一刀没剪。只用Claude加了个字幕。
37
9
106
21,763
用 MiniMax H3 做视频发抖音的 一定要用我开源的reelbench-skills/video-scrub去掉视频元数据 github.com/eternityspring/re… 看我抖音号发了5个作品,前三个是没有去元数据,发布后一周内的播放都是没有破500的。 最近两个视频都用了skill去掉元数据,播放都破万了。 昨天x评测的视频同步在抖音已经5w+播放了 x.com/eternityspring/status/…
AI 视频做完发布,播放个位数 可能不是视频质量差,是文件本身带着生成痕迹 平台一看就降权 我开源的 reelbench-skills 新增 video-scrub: 模拟录屏基于AI生成的视频重出一个全新的视频 人眼看着一模一样,机器读到的元数据变成 OBS(默认)录制 原片的标签扔掉,画面和声音原样保留 抖音实测,模拟转录后,推流好一点 有兴趣的可以拿去实测一下: github.com/eternityspring/re…
56
107
761
63,633
开整,准备做个读剧本的工具 借助开源的 tts 模型,用不同音色把剧本读一遍
剧本的质量从根本上决定了AI短剧的质量 而对于没有经验的人来说,很难直观的判定剧本质量 我想到一个好办法就是用tts模型把剧本读出来 要想把剧本读的有情感,就需要用tts模型认识的情感符号把一句话表现的更有节奏。我整理了一些常见的情感符号: 句末语调-收落、强调、疑问上扬: 。 ! ? 短中停顿-控制停顿长度和语义分组: , 、 ; : 延宕拖尾-犹豫、渐弱、留白、较长停顿: …… 转折中断-突然转折、打断、拖长: —— 组合语气-惊讶兼疑问: ?! !? 强度重复-加强语气,但模型表现不稳定: !! ?? 内容边界-通常产生明显停顿或重新起句: 换行 分段 弱提示-可能改变措辞节奏,但不是可靠的韵律控制: 引号 括号 书名号 等有空准备做个开源工具一键让 shuohao-skills 生成的剧本变成有声读物。
6
25
3,378
用 MiniMax H3 + ComfyUI-H3-Continuum 做了个 110 秒的古装穿越短片《穿越娶亲,礼单炸了》 整条分三次生成,分别是 30 秒、40 秒、40 秒,最后拼起来。H3 单次最长 15 秒,每次里面靠 Continuum 把 10 秒一段接着上一段最后 22 帧往下续。 整片都是一次出,一刀没剪。只用Claude加了个字幕。
上次分享Motion Context能用MiniMax H3生成多段衔接流畅的视频 现在有了更智能的方案 ComfyUI-H3-Continuum 可以一次性用 MiniMax H3 生成 30s 或者更长的视频 分块(Chunked)采样:把长视频拆成多个 5–8 秒的片段依次生成 原生音视频潜空间连续性(AV latent continuity):把上一个片段的视频和音频潜变量传给下一个片段,减少画面跳变和声音断裂 会话恢复 / 断点续传:中途中断可以接着跑,不用从头重来 支持参考图、首尾帧、参考音频等 H3-Continuum:像一个“长视频流水线”。你设好参数、提示词和时间线,它自动把视频分成多个 chunk 连续生成,并尽量保持画面和声音连贯。
70
87
996
178,209
《穿越娶亲,礼单炸了》第二段提示词: photorealistic cinematic shot, natural soft diffused lighting, realistic skin subsurface scattering, natural skin texture, no oily skin, subtle skin pores, natural eye catchlight, soft shadow transition, physically based rendering, PBR, real fabric texture, natural material reflection, fine film grain, shallow depth of field, smooth motion, stable facial features, consistent face across frames, no painted texture, no thick brush strokes, no oil painting effect, live action, real world scene, natural color grading, moderate contrast [Chunk 1] integrated_multimodal_description: [Shot 1] Live-action, cinematic Chinese costume comedy set in the Song dynasty, with natural soft daylight, natural color, fine film grain and a shallow depth of field. A medium-wide shot of the long gift table in the courtyard of a Song dynasty family mansion decorated for a wedding, with grey tiled roofs and upturned eaves, dark wooden pillars, red silk ribbons and round red paper lanterns along the veranda, a grey stone-paved courtyard and a long wooden gift table draped with red cloth holding a closed red ledger book, an inkstone and a brush, in soft overcast daylight. Behind the table sits the old gift clerk, a thin man about sixty-five with a long narrow face, a wispy grey goatee, sharp squinting eyes and a small black cloth cap, wearing a long dark-grey scholar's robe with a dark-blue collar, holding a brush. Beside the table stands the groom, a slim young man about twenty-four with a clean-shaven youthful face and straight eyebrows, black hair in a neat topknot under a black Song dynasty futou hat with two short flat wings, wearing a red round-collared wedding robe with wide sleeves and a large red silk flower ball on a red sash across his chest. A cheerful peddler about thirty with a round face and a grey cloth headscarf, wearing a short dark-blue cotton jacket, steps up and hands a small plain orange plastic gift card to the clerk with both hands. The camera holds a static shot. [Shot 2] At 00:03.000, the shot cuts to a medium shot of the old clerk with the grey goatee holding up the orange card. In a high, thin, sing-song old male voice (S2), he loudly announces: <d>[Chinese] 李货郎,随礼——腾讯视频,会员十年!</d> He closes his lips. The camera holds a static shot. [Shot 3] At 00:07.000, the shot cuts to a medium close-up of the peddler in the grey headscarf. He grins and, in a bright, cheeky male voice (S9), says: <d>[Chinese] 跳过广告,直奔洞房!</d> He waggles his eyebrows. The camera holds a static shot. overall_soundscape: Festive chatter of the wedding guests, the clerk's sing-song announcement, a few guests snickering. non_diegetic_music: A lively Chinese suona wedding tune with gongs and small cymbals plays at a brisk tempo, kept low under the voices. [Chunk 2] integrated_multimodal_description: Continuation of Chunk 1. [Shot 1] Live-action, cinematic Chinese costume comedy set in the Song dynasty, with natural soft daylight, natural color, fine film grain and a shallow depth of field, the medium close-up of a cheerful peddler about thirty with a round face and a grey cloth headscarf, wearing a short dark-blue cotton jacket continues for a moment as he grins and waggles his eyebrows, his lips closed. [Shot 2] At 00:01.500, the shot cuts to a medium close-up of the groom, a slim young man about twenty-four with a clean-shaven youthful face and straight eyebrows, black hair in a neat topknot under a black Song dynasty futou hat with two short flat wings, wearing a red round-collared wedding robe with wide sleeves and a large red silk flower ball on a red sash across his chest, standing beside the red gift table and clutching his head with both hands, his futou hat knocked slightly crooked. In an exasperated young male voice (S1), he says: <d>[Chinese] 大哥,这儿九百年后才通网!</d> He lets his hands drop. The camera holds a static shot. [Shot 3] At 00:05.500, the shot cuts to a medium shot of the gift table, where the old gift clerk, a thin man about sixty-five with a long narrow face, a wispy grey goatee, sharp squinting eyes and a small black cloth cap, wearing a long dark-grey scholar's robe with a dark-blue collar, lifts the lid of a small flat light-blue cardboard box. He peers inside, his squinting eyes going wide, and in a slow, puzzled, high, thin old voice (S2) says: <d>[Chinese] 赵家小娘子,随礼——此物,老夫不识。</d> He holds the open box up in one hand. The camera holds a static shot. overall_soundscape: Guests laughing, the soft scrape of a cardboard lid, then a sudden hush in the courtyard. non_diegetic_music: The suona wedding tune plays on, then stops abruptly as the clerk opens the box. [Chunk 3] integrated_multimodal_description: Continuation of Chunk 2. [Shot 1] Live-action, cinematic Chinese costume comedy set in the Song dynasty, with natural soft daylight, natural color, fine film grain and a shallow depth of field, the medium shot of the old gift clerk, a thin man about sixty-five with a long narrow face, a wispy grey goatee, sharp squinting eyes and a small black cloth cap, wearing a long dark-grey scholar's robe with a dark-blue collar, holding up a small flat light-blue cardboard box continues; he blinks at it, his lips closed. The courtyard is silent. [Shot 2] At 00:01.500, the shot cuts to a medium shot by the gift table. The grandmother, a small, sprightly woman about eighty with a kind wrinkled face and silver-white hair in a neat low bun with a green jade hairpin, wearing a dark-brown brocade jacket and holding a carved dark wooden cane with a dragon-head handle, hurries in, snatches the blue box out of the clerk's hand and hugs it to her chest. In a brisk, crackly old female voice (S4), she declares: <d>[Chinese] 杜蕾斯!没收!老身要抱重孙!</d> She taps her cane on the stones. The camera holds a static shot. [Shot 3] At 00:06.000, the shot cuts to a medium two-shot of the groom, a slim young man about twenty-four with a clean-shaven youthful face and straight eyebrows, black hair in a neat topknot under a black Song dynasty futou hat with two short flat wings, wearing a red round-collared wedding robe with wide sleeves and a large red silk flower ball on a red sash across his chest, and the silver-haired grandmother. The groom (S1) stares at her and stammers: <d>[Chinese] 奶奶……您也是?</d> The grandmother (S4) gives him a calm little smile and says: <d>[Chinese] 来了五十年喽。</d> She pats his arm. The camera holds a static shot. overall_soundscape: A beat of silence, quick footsteps and the tap of a wooden cane on stone, then the guests burst out laughing. non_diegetic_music: The suona wedding tune comes back in, bright and brisk, after the grandmother's line. [Chunk 4] integrated_multimodal_description: Continuation of Chunk 3. [Shot 1] Live-action, cinematic Chinese costume comedy set in the Song dynasty, with natural soft daylight, natural color, fine film grain and a shallow depth of field, the medium two-shot of the groom, a slim young man about twenty-four with a clean-shaven youthful face and straight eyebrows, black hair in a neat topknot under a black Song dynasty futou hat with two short flat wings, wearing a red round-collared wedding robe with wide sleeves and a large red silk flower ball on a red sash across his chest, and the grandmother, a small, sprightly woman about eighty with a kind wrinkled face and silver-white hair in a neat low bun with a green jade hairpin, wearing a dark-brown brocade jacket and holding a carved dark wooden cane with a dragon-head handle, continues for a moment as she pats his arm and he gapes at her. [Shot 2] At 00:01.500, the shot cuts to a medium shot of the gift table, where the old gift clerk, a thin man about sixty-five with a long narrow face, a wispy grey goatee, sharp squinting eyes and a small black cloth cap, wearing a long dark-grey scholar's robe with a dark-blue collar, holds up a colorful paper scratch-off lottery ticket with a silver scratch-off strip. In his high, thin, sing-song old voice (S2), he announces: <d>[Chinese] 表少爷,随礼——彩票一张!两块到一千万!</d> He closes his lips. The camera holds a static shot. [Shot 3] At 00:05.500, the shot cuts to a medium shot near the table. The aunt, a plump woman about fifty with a round face, thin arched eyebrows and a haughty look, black hair piled in a high bun stuck full of gold hairpins, wearing a magenta brocade jacket over a long dark-purple skirt and a thick shiny gold bangle on her right wrist, stands with her back to the table and, without looking, sneers over her shoulder in a loud, sharp middle-aged female voice (S3): <d>[Chinese] 谁家孩子这么穷酸——</d> She turns around and sees the aunt's son, a slim teenage boy about seventeen with a boyish face, black hair in a small topknot tied with a light-blue ribbon, wearing a light-blue scholar's robe with a white collar, standing at the table. He gives a sheepish grin and says in a young male voice (S5): <d>[Chinese] 娘,是我。</d> The aunt freezes, her mouth hanging open. The camera holds a static shot. overall_soundscape: The guests murmur and snicker, the aunt's sharp sneer cuts through, then an awkward silence. non_diegetic_music: A lively Chinese suona wedding tune with gongs and small cymbals plays at a brisk tempo, kept low under the voices.
1
4
4,632
《穿越娶亲,礼单炸了》第三段提示词: photorealistic cinematic shot, natural soft diffused lighting, realistic skin subsurface scattering, natural skin texture, no oily skin, subtle skin pores, natural eye catchlight, soft shadow transition, physically based rendering, PBR, real fabric texture, natural material reflection, fine film grain, shallow depth of field, smooth motion, stable facial features, consistent face across frames, no painted texture, no thick brush strokes, no oil painting effect, live action, real world scene, natural color grading, moderate contrast [Chunk 1] integrated_multimodal_description: [Shot 1] Live-action, cinematic Chinese costume comedy set in the Song dynasty, with natural soft daylight, natural color, fine film grain and a shallow depth of field. A medium shot in the courtyard of a Song dynasty family mansion decorated for a wedding, with grey tiled roofs and upturned eaves, dark wooden pillars, red silk ribbons and round red paper lanterns along the veranda, a grey stone-paved courtyard and a long wooden gift table draped with red cloth holding a closed red ledger book, an inkstone and a brush, in soft overcast daylight, near the gift table. The aunt, a plump woman about fifty with a round face, thin arched eyebrows and a haughty look, black hair piled in a high bun stuck full of gold hairpins, wearing a magenta brocade jacket over a long dark-purple skirt and a thick shiny gold bangle on her right wrist, stands stiffly beside the aunt's son, a slim teenage boy about seventeen with a boyish face, black hair in a small topknot tied with a light-blue ribbon, wearing a light-blue scholar's robe with a white collar, her face tight with embarrassment. Guests in robes around them snicker behind their sleeves. The camera holds a static shot. [Shot 2] At 00:02.500, the shot cuts to a medium close-up of the plump aunt with the gold hairpins. She yanks the thick shiny gold bangle off her right wrist, holds it high and, in a loud, sharp middle-aged female voice (S3), declares: <d>[Chinese] 这破纸要能换出一文钱,这镯子当场给你!</d> She glares, the bangle raised. The camera holds a static shot. [Shot 3] At 00:07.000, the shot cuts to a medium shot of the groom, a slim young man about twenty-four with a clean-shaven youthful face and straight eyebrows, black hair in a neat topknot under a black Song dynasty futou hat with two short flat wings, wearing a red round-collared wedding robe with wide sleeves and a large red silk flower ball on a red sash across his chest, standing by the gift table holding a colorful paper scratch-off lottery ticket with a silver scratch-off strip. He grins and, in a bright young male voice (S1), says: <d>[Chinese] 姨母说的啊!</d> The guests behind him cheer and clap. The camera holds a static shot. overall_soundscape: Guests snickering, the gold bangle sliding off a wrist, then cheers and clapping. non_diegetic_music: A lively Chinese suona wedding tune with gongs and small cymbals plays at a brisk tempo, kept low under the voices. [Chunk 2] integrated_multimodal_description: Continuation of Chunk 1. [Shot 1] Live-action, cinematic Chinese costume comedy set in the Song dynasty, with natural soft daylight, natural color, fine film grain and a shallow depth of field, the medium shot of the groom, a slim young man about twenty-four with a clean-shaven youthful face and straight eyebrows, black hair in a neat topknot under a black Song dynasty futou hat with two short flat wings, wearing a red round-collared wedding robe with wide sleeves and a large red silk flower ball on a red sash across his chest, grinning by the gift table continues. Beside him, the bride, a slim young woman in a red wedding robe embroidered with gold phoenixes, her head and face covered by a square red silk veil with short gold tassels, reaches out from her wide sleeve and puts a round copper coin with a square hole in the middle into his palm. [Shot 2] At 00:03.000, the shot cuts to an extreme close-up of a colorful paper scratch-off lottery ticket with a silver scratch-off strip lying flat on the red tablecloth, soft in the shallow focus. His fingers scrape the copper coin across the silver strip, and the silver coating flakes away in tiny curls. The camera holds a static shot. [Shot 3] At 00:06.000, the shot cuts to a medium two-shot. The groom in the red robe (S1) lifts the ticket high and shouts in a bright young male voice: <d>[Chinese] 中了!两块!</d> In the foreground, the aunt, a plump woman about fifty with a round face, thin arched eyebrows and a haughty look, black hair piled in a high bun stuck full of gold hairpins, wearing a magenta brocade jacket over a long dark-purple skirt and a thick shiny gold bangle on her right wrist, spins toward him and, in a sharp middle-aged female voice gone shaky (S3), asks: <d>[Chinese] 两块……是几文?</d> The camera holds a static shot. overall_soundscape: A hush falls over the courtyard, the coin scratching the ticket, then a burst of excited gasps. non_diegetic_music: The suona tune drops out for the scratching, leaving a soft, tense drumroll. [Chunk 3] integrated_multimodal_description: Continuation of Chunk 2. [Shot 1] Live-action, cinematic Chinese costume comedy set in the Song dynasty, with natural soft daylight, natural color, fine film grain and a shallow depth of field, the medium two-shot of the groom, a slim young man about twenty-four with a clean-shaven youthful face and straight eyebrows, black hair in a neat topknot under a black Song dynasty futou hat with two short flat wings, wearing a red round-collared wedding robe with wide sleeves and a large red silk flower ball on a red sash across his chest, holding up the ticket and the aunt, a plump woman about fifty with a round face, thin arched eyebrows and a haughty look, black hair piled in a high bun stuck full of gold hairpins, wearing a magenta brocade jacket over a long dark-purple skirt and a thick shiny gold bangle on her right wrist, in the foreground continues. From all around the courtyard, the wedding guests (S10) shout together in one big chorus of mixed voices: <d>[Chinese] 比一文多!</d> The aunt flinches. The camera holds a static shot. [Shot 2] At 00:03.000, the shot cuts to a medium shot of the old gift clerk, a thin man about sixty-five with a long narrow face, a wispy grey goatee, sharp squinting eyes and a small black cloth cap, wearing a long dark-grey scholar's robe with a dark-blue collar, rising from his chair behind the gift table and raising his brush high. In his high, thin, sing-song old voice (S2), he announces: <d>[Chinese] 姨母,随礼——金镯子一只!</d> The guests cheer. The camera holds a static shot. [Shot 3] At 00:07.000, the shot cuts to a medium close-up of the plump aunt with the gold hairpins clutching the thick shiny gold bangle in one hand. She narrows her eyes, looks around the courtyard and, in a low, suspicious voice (S3), asks: <d>[Chinese] 你们……怎么都知道?</d> The camera holds a static shot. overall_soundscape: A loud chorus of guests shouting, cheers and laughter, then the noise drops to a curious murmur. non_diegetic_music: The suona wedding tune bursts back in, loud and triumphant, then dips low under the aunt's question. [Chunk 4] integrated_multimodal_description: Continuation of Chunk 3. [Shot 1] Live-action, cinematic Chinese costume comedy set in the Song dynasty, with natural soft daylight, natural color, fine film grain and a shallow depth of field, the medium close-up of the aunt, a plump woman about fifty with a round face, thin arched eyebrows and a haughty look, black hair piled in a high bun stuck full of gold hairpins, wearing a magenta brocade jacket over a long dark-purple skirt and a thick shiny gold bangle on her right wrist, looking around with narrowed eyes continues for a moment, her lips closed. [Shot 2] At 00:01.500, the shot cuts to a wide shot of the courtyard of a Song dynasty family mansion decorated for a wedding, with grey tiled roofs and upturned eaves, dark wooden pillars, red silk ribbons and round red paper lanterns along the veranda, a grey stone-paved courtyard and a long wooden gift table draped with red cloth holding a closed red ledger book, an inkstone and a brush, in soft overcast daylight. All at once, every guest in the courtyard pulls a modern black smartphone with a dark, switched-off screen out of a wide sleeve and holds it up; the grandmother, a small, sprightly woman about eighty with a kind wrinkled face and silver-white hair in a neat low bun with a green jade hairpin, wearing a dark-brown brocade jacket and holding a carved dark wooden cane with a dragon-head handle, and the old gift clerk, a thin man about sixty-five with a long narrow face, a wispy grey goatee, sharp squinting eyes and a small black cloth cap, wearing a long dark-grey scholar's robe with a dark-blue collar, hold up phones too. In the foreground, the plump aunt in the magenta jacket stares, her eyes roll back and she faints backward into the arms of the aunt's son, a slim teenage boy about seventeen with a boyish face, black hair in a small topknot tied with a light-blue ribbon, wearing a light-blue scholar's robe with a white collar. The camera holds a static shot. [Shot 3] At 00:06.000, the shot cuts to a close two-shot of the groom, a slim young man about twenty-four with a clean-shaven youthful face and straight eyebrows, black hair in a neat topknot under a black Song dynasty futou hat with two short flat wings, wearing a red round-collared wedding robe with wide sleeves and a large red silk flower ball on a red sash across his chest, and the bride, a slim young woman in a red wedding robe embroidered with gold phoenixes, her head and face covered by a square red silk veil with short gold tassels, beside him. With two fingers she lifts one corner of her red silk veil, showing only her smiling lips and chin, leans toward him and, in a soft, playful young female voice (S6), whispers: <d>[Chinese] 相公,充电宝带了吗?</d> The groom turns to her, speechless. The camera pushes in with small amplitude at slow speed. overall_soundscape: The rustle of many sleeves, a stunned silence, a soft thump as the aunt faints, then a quiet whisper. non_diegetic_music: The suona wedding tune returns for one last bright flourish and ends on a single gong hit.
4
4,016
上个月刚把知名的开源 TTS 引擎试了个遍 这个月 Fish Audio 就开源了新一代的 Drama 3 主打多角色配音和长文本 用来做个读剧本的工具应该很不错,主要他有这三点优势: 第一,一段文本里用 <|speaker:N|> 标签就能分出最多 16 个角色,广播剧、有声书一次就能生成。 第二,可以在台词里直接插 [sighs]、[laughs]、[whisper] 这类标签,叹气、笑声、耳语会落在你标的位置。 第三,它把编辑当成续写:一句话里只改一个词,比如“周三”改成“周四”,把之前的生成结果一起传回去,就能只改这个词,其余部分的声音和语气保持不变,不用整句重来。 另外它支持用几秒参考音频加文字稿克隆声音,号称支持 83 种语言,输出 44.1kHz。 模型是 35B 参数的 MoE,每个 token 只激活 3B,推理走 SGLang-Omni 或 vLLM-Omni 官方在 H200 上测的单路首包约 89ms,实时率 0.13,也就是生成 1 秒音频约 0.13 秒 完整版要 48GB 以上显存,另外有 int4 量化版,权重约 21GB huggingface.co/fishaudio/dra…
9
4
77
6,553
周末了,有点想造轮子用 solidjs 写个做视频的无限画布了...
1
1,038
逼着升级 30刀+的x订阅啊,这grok额度几下就聊没了
2
996
上次分享Motion Context能用MiniMax H3生成多段衔接流畅的视频 现在有了更智能的方案 ComfyUI-H3-Continuum 可以一次性用 MiniMax H3 生成 30s 或者更长的视频 分块(Chunked)采样:把长视频拆成多个 5–8 秒的片段依次生成 原生音视频潜空间连续性(AV latent continuity):把上一个片段的视频和音频潜变量传给下一个片段,减少画面跳变和声音断裂 会话恢复 / 断点续传:中途中断可以接着跑,不用从头重来 支持参考图、首尾帧、参考音频等 H3-Continuum:像一个“长视频流水线”。你设好参数、提示词和时间线,它自动把视频分成多个 chunk 连续生成,并尽量保持画面和声音连贯。
有朋友问:MiniMax H3 一次最多生成 15 秒,做短剧的话,段和段之间怎么接才不突兀? 我推荐用 ComfyUI-H3-Motion-Context 的方案,项目中有直接可用的工作流示例。 常见做法是拿上一段的最后一帧去做图生视频。问题是一张静态图里没有运动信息,模型只能猜:人刚才往哪走、多快,全靠猜,所以接缝处经常卡一下或者动作突变。 Motion Context 的做法是:把上一段最后 22 帧(0.92 秒)原样钉在下一段开头,每一步采样都不动它们。模型看到的是真实的运动,不用从一张图去猜。 画面和声音都是直接从上一段的 latent 里切出来的,不用先解码成视频再编码回去,所以不会每接一次就损失一点画质。 生成完,用 Trim 节点把开头这 22 帧剪掉,两段直接首尾拼接,不需要做转场。我逐帧看了接缝前后,角色的姿势、位置、表情都对得上。 最关键的是提示词:下一段开头要先描述上一段是怎么结束的。同一个角色、同样的景别、同样的动作,保持一两秒,再写新的剧情。 项目地址: github.com/NikoDemon80/Comfy… 我不说的话,谁能看出这30s的功夫熊猫是两段生成的?
17
32
314
123,427
做视频的都知道,多人场景下人物占位是个头疼的问题 之前 shuohao-skills 解决了小说拆角色、排大纲、出场景与道具设定、写剧本、切分镜等问题 这两天我又加了个 shuohao-skills/novel-staging 来解决占位问题 github.com/eternityspring/sh… 输入剧本,就可以出一个能调整占位的简易工作台报告页,可以在报告页中调整占位之后: 把占位参考图交给出图模型生定位图 也可以直接把占位参考图喂给视频模型做位置参考 我也把skill上传在 agent.digen.ai/zh/skills/aut…有兴趣的可以直接在线尝试
小说做短剧系列 shuohao-skills 已经3.8k star了 前两天把已做完的skill做了解耦,并改为了纯文本任务 这意味着你现在可以零基础、零成本来体验 shuohao-skills 中的任意一个skill。比如 shuohao-skills/novel-outline: - 打开skill.md 并复制其url github.com/eternityspring/sh… - 在豆包或者Deepseek中上传你的小说,输入以下提示词: 按照 github.com/eternityspring/sh… 这个标准,把我发的内容给我生成大纲 - 以截图豆包中实测为例,他会返带有简单格式的大纲文本。对于这种情况,可以有两种用法: 1.用提示词“给我大纲报告 json 和 html报告”让他返个json,和html。方便你查阅和在其他地方使用大纲数据。 2.直接在当前对话框继续丢 shuohao-skills 中的剧本skill。
5
13
110
7,023