GitHub

AI 探索笔记AI ExplorationsAI Explorations

一些 AI 模型的实测，和可以直接上手玩的 demo。每一个都是静态网页，点开就能玩。

Hands-on tests of AI models, with interactive demos you can play with right in the browser.

打开 demoOpen demo

语音合成Speech synthesisTTS 对比TTS comparison盲听Blind test交互Interactive

Drama 3 对 IndexTTS 2.5：开口快十倍，情绪更到位Drama 3 vs IndexTTS 2.5 — speaks 10× sooner, with better emotion

Fish Audio 新开源的 Drama 3 和 IndexTTS 2.5，用同样两个音色比了三个场景：像发微信语音那样聊天、演十种情绪、念一篇口播稿。Drama 3 用社区 int4 量化版跑在 40GB 的 A100 上，开口中位数 0.26 秒，IndexTTS 要 2.6 秒。两段语音电话能按真实延迟回放，所有 A/B 都能开盲听。

Fish Audio's newly open-sourced Drama 3 against IndexTTS 2.5, same two voices, three scenes: chatty WeChat-style voice messages, ten emotions, and a narration script. Drama 3 runs as the community int4 quant on a 40GB A100 and starts speaking in 0.26 s (median) versus 2.6 s for IndexTTS. Two voice calls replay with their real latency, and every A/B pair can be blind-tested. The page is in Chinese.

Drama 3 int4 (fishaudio/drama-3-int4)IndexTTS 2.5SGLang-Omni 0.1.8Whisper turboemotion2vec_plus_large

要点Notes
  • 打电话：IndexTTS 生成和播放差不多快（实时率 1.02），模拟通话里 20 句有 14 句播到一半要停下来等；Drama 3 实时率 0.26，一句没卡。On calls: IndexTTS generates about as fast as it plays (RTF 1.02), so 14 of 20 lines stalled mid-playback; Drama 3 runs at RTF 0.26 and never stalled.
  • 情绪：句首加 [sad]、[angry] 这类标签，emotion2vec 判对 11/18，不加 9/18，IndexTTS 的情绪向量 7/18；厌恶谁都没做出来。Emotion: with tags like [sad] or [angry] up front, emotion2vec agreed on 11/18, 9/18 without tags, and 7/18 for IndexTTS emotion vectors. Nobody managed disgust.
  • Drama 3 的语速跟着参考音频走：2.4 秒的快语速参考让它每秒说 6 个字，听着像在赶；IndexTTS 两个音色都稳在每秒 4 字左右。Drama 3 copies the pace of its reference clip: a fast 2.4-second reference made it speak about 6 characters a second, which sounds rushed; IndexTTS stays near 4 for both voices.
  • int4 版在 A100-40GB 上占约 28GB 显存；为了塞进去关掉了预填充 CUDA graph，4 路并发时每路开口涨到 2.6–2.7 秒。The int4 build takes about 28GB on an A100-40GB. Fitting it meant turning off prefill CUDA graphs, and with 4 concurrent requests each one took 2.6–2.7 s to start.
  • SGLang-Omni 0.1.8 还不支持多角色和续写式编辑，模型卡上这两项暂时用不了。SGLang-Omni 0.1.8 doesn't support multi-speaker or continuation editing yet, so those two model-card features aren't usable for now.

打开 demoOpen demo ↗

打开 demoOpen demo

视频生成Video generation环绕视频Orbit video抠像Matting交互Interactive

天台起舞：她一直在跳，拖着转整个天台Rooftop Dance — spin the scene while she keeps dancing

接着《朋友圈》第 4 格：苏曼在天台派对上一直在跳舞，拖动画面，整个天台绕着她转。背景和人分开生成——空天台用 MiniMax-H3 + ORB360 LoRA 绕一圈，跳舞用 H3 首尾帧出一段能无缝循环的 5 秒视频，再用 SAM 3.1 抠出来叠在中间，音乐也是 H3 一起生成的。

A follow-up to panel 4 of The Feed: she keeps dancing at the rooftop party while you drag to spin the whole rooftop around her. The background and the dancer are made separately — an empty-rooftop orbit from MiniMax-H3 + the ORB360 LoRA, and a seamlessly looping 5-second dance from H3 first/last-frame video, cut out with SAM 3.1 and layered in the middle. The music comes from H3 too.

Qwen Image 2.1MiniMax-H3 FL2V (8-step turbo)MiniMax-H3 Ref2VAORB360 LoRA (step 1500)SAM 3.1

要点Notes
  • 两层分开，所以她永远正面朝你：拖动转的是背景，镜头不会绕到她背后。Two separate layers, so she always faces you: dragging spins the background, and the camera never goes around behind her.
  • 让 ORB360 绕一块空地转：2 个种子里 1 个真的绕了一圈、首尾接得上；另一个把整个画面原地翻滚，城市都倒过来了。Asking ORB360 to orbit an empty spot: one of two seeds circled properly and looped cleanly; the other rolled the whole frame, skyline upside down.
  • 人糊是因为太小：1024×768 里她的脸只有约 76 像素，一跳就糊成蜡像；整张图提到 1536×1152 后，同样的显示尺寸下脸部清晰度约是原来的 3 倍。The blur came from size: at 1024×768 her face was about 76 px and smeared as soon as she moved; at 1536×1152 the face is about 3× sharper at the same display size.
  • 想裁成竖构图让她更大，结果 5 次全都推镜头（推近 1.09–1.6 倍），提示词写了不准推也没用；横构图 4 次都纹丝不动。Cropping to a tall portrait to make her bigger backfired: all 5 runs pushed in (1.09–1.6×), even with a no-zoom prompt; all 4 landscape runs held perfectly still.
  • SAM 3.1 按 woman 抠人不带她手里的酒杯，要再按 cocktail glass 抠一次补上。SAM 3.1’s “woman” mask leaves out the cocktail glass in her hand; a second “cocktail glass” pass fills it in.

打开 demoOpen demo ↗演示视频Video

打开 demoOpen demo

图像生成Image generation人物一致性Character consistency环绕视频Orbit video交互Interactive

朋友圈：一页可以转的漫画The Feed — a comic page you can spin

Qwen Image 2.1 画同一位名媛一周的朋友圈：六套夏日和度假造型，换了衣服，脸、卷发、泪痣、首饰都对得上；每格再用 MiniMax-H3 + ORB360 LoRA 生成 360° 环绕视频，拼成一页可以拖着转的漫画。周日那格衣架上的吊牌可以点。

A socialite’s week on her feed, drawn with Qwen Image 2.1: six summer and resort outfits, and her face, hair, beauty mark and jewelry hold steady through every change. Each panel becomes a 360° orbit with MiniMax-H3 + the ORB360 LoRA — drag to spin, and click the clothes rack on Sunday.

Qwen Image 2.1MiniMax-H3 Ref2VAORB360 LoRA (step 1500)three.js

要点Notes
  • 换装不换人：六套衣服、白天黑夜、浓妆素颜，脸、栗色卷发、右眼下的泪痣和金耳环都对得上。Outfits change, she doesn’t: six looks, day and night, full makeup and none — same face, chestnut waves, beauty mark and gold hoops.
  • 环绕时皮肤、薄纱、亮片都稳，背面补得合理：游艇补出船舱，悬崖泳池补出别墅，天台补出酒吧。Skin, sheer fabric and sequins hold up through the orbit, and the back views make sense: a yacht cabin, cliffside villas, a rooftop bar.
  • 大片平地是弱点：网球场 4 次里 3 次镜头升到头顶、画面翻倒，加“镜头保持水平”的提示词也救不回来，只能换种子；坐在地毯上那格加了这句就好了。Big flat ground is the weak spot: on the clay court the camera rose overhead and flipped in 3 of 4 runs, and a “keep the camera level” prompt didn’t save it — only a new seed did. The same line did fix the panel where she sits on a rug.
  • 身边的道具会被“粘”在人身上：漂浮早餐篮在一个种子里转到侧面就跑到她脚下，像站在篮子里；换个种子，篮子才老老实实留在甲板上。Props near the subject can get glued to her: in one seed the floating breakfast tray slid under her feet as the camera came around, as if she were standing in it; another seed kept it on the deck.

打开 demoOpen demo ↗演示视频Video

打开 demoOpen demo

图像生成Image generation环绕视频Orbit video交互Interactive

寻猫：一页可以转的漫画Lost Cat — a comic page you can spin

Qwen Image 2.1 画六格剧情漫画，每格用 MiniMax-H3 + ORB360 LoRA 生成 360° 环绕视频，再用 three.js 拼成一页：拖动任意一格就能绕着角色转，点一格推近看，第二格的寻猫启事还藏了彩蛋。

Six comic panels drawn with Qwen Image 2.1, each turned into a 360° orbit video with MiniMax-H3 + the ORB360 LoRA, then assembled into one three.js page: drag any panel to circle the characters, click to zoom in — and try the Lost Cat poster in panel 2.

Qwen Image 2.1MiniMax-H3 Ref2VAORB360 LoRA (step 1500)three.js

要点Notes
  • 每一格的环绕都成立：转一圈回到原图，背面补得合理，角色前后一致，双人格也能一起转。Every orbit closes the loop: plausible back views, consistent characters, even two people turning together.
  • 官方提示词会把室内背景补成户外（草地、雪）；给每格写清 360° 环境后，卧室和菜市场都修好了。The stock prompt turns indoor scenes into outdoor ones (grass, snow); describing each panel’s surroundings fixed the bedroom and the market.
  • 提示词改不了图里已有的遮挡：主角背后贴着墙或广告牌，转到背面就看不到人，出图时要给背后留空。Prompts can’t remove occluders already in the image — leave space behind the subject when drawing the panel.

打开 demoOpen demo ↗演示视频Video