shenzhen
Joined February 2007
- Tweets29.6K
- Following211
- Followers19.7K
- Likes11.6K
Datou retweeted
Super hyped to finally ship this! You can now play with interactive 3D prototypes of AI hardware directly in your browser.
If you want to learn about the hottest AI devices on the market, looking at flat 2D images isn't enough—you need to see the actual build.
It covers almost every trending AI device worth learning about, complete with 3D models, GitHub repos, and official sites.
Connecting Shenzhen to the world.
Enjoy🏊
As an AI hardware founder in Shenzhen, I sift through tons of smart devices daily. Most of them are pure noise.
So I spent the last 2 months putting together almost every single AI hardware project on the market that is actually worth seeing.
Click into any product to get its official website + GitHub repo.
Enjoy! 👇
Datou retweeted
kale-img-1 fine-tuning results
1. @ostrisai / 8.10 / moss
2. @AmandineFlachs / 7.61 / moss
3. @suhaaspk / 7.20 / moss
4. @Assu_Bhadhu / 6.90 / moss
5. @max_pe2002 / 5.35 / hdo1
6. @hasnain__dev / 3.51 / sdxl
(same prompt, seed, resolution)
ALT A large comparison grid from the kale-img-1 fine-tuning challenge. Each of its 20 rows shows one subject: a reference picture from our dataset (soft 3D children's illustrations on white), then six entries' versions, ranked by score: ostris 8.1, amandine flachs 7.6, su 7.2, assu_bhadhu 6.9 (all MOSS-based), max 5.3 (HiDream-O1), hasnain 3.5 (SDXL). Subjects: sea otter, toucan, kid planting a seedling, girl skipping rope, two kids building a sandcastle, alarm clock, xylophone, steam train, pagoda, treehouse, pancakes, cactus, fire truck, cello, sea turtle, monarch butterfly, windmill, kid astronaut, teapot, kid playing king. Every picture has a score badge and a pass or fail mark. Ostris's column has a gold border and passes every row. Amandine and su each drew a four-roofed pagoda, assu_bhadhu put writing on the train, max has blotchy backgrounds, hasnain often gets compositions wrong. Same prompts, seeds and resolution for all, scored blind by three judge passes.
paying $5k to whoever fine-tunes kale-img-1, an img model for kids
dm me
i'll send u the dataset
u pick a base model we can license commercially
u comment with ur result here
ur result gets judged on:
- recreating 5 imgs from the dataset almost 1:1
- making 5 new imgs in our style
ALT Promotional poster for Kale inviting people to help create “KALE-IMG-1,” an image model for kids. The design uses a bright yellow grid background with playful 3D-style icons in white rounded cards, including a castle, panda, helicopter, cupcake, child playing soccer, globe, parrot, child playing violin, dolphin, double-decker bus, tortoise, and tropical island. A large brown central panel reads “help us create KALE-IMG-1, an img model FOR KIDS,” with a white label below stating “best submission = $5k.” The Kale logo appears in the lower-right corner. The overall style is colorful, friendly, and child-focused, combining educational and imaginative visual themes.
Datou retweeted
5 MiniMax H3 speedups on one RTX 5090
Stock 20-step: 129.0 s
larryvrh Turbo: 61.4 s (2.1x)
LightX2V Turbo: 55.8 s (2.3x)
FastH3 V2 +VSA: 44.7 s (2.9x)
DMAD 4-step: 31.7 s (4.1x)
Same diner clip, 5 s with sound, 1344x768. Settings and quality in the reply.
Datou retweeted
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
Datou retweeted
I’ve developed TARA — Temporally Aligned Reference Attention, an experimental method for MiniMax H3.
Research, implementation and video examples:
github.com/alisson-anjos/ai-…
Native video references can capture an action but drift out of sync with the source. TARA aligns the reference and output timelines inside attention, helping preserve performance and timing while using native reference conditioning, without Add Guide.
I’ve also developed optional precompiled kernels for Comfy Kitchen and sol-attn, with TARA’s temporal attention bias fused directly into the attention computation. This avoids the slower per-frame compatibility path. BFSNodes automatically downloads the matching binary on supported systems, compatibility and performance across GPUs are still being validated.
Early tests look promising. Interestingly, my guide-latent-trained LoRA currently seems to work better with TARA than the unfinished TARA-trained checkpoint. This isn’t a controlled comparison; I’m training a fresh LoRA from scratch to investigate further.
There’s a TARA node in BFSNodes for native ComfyUI workflows. If you’d like to try it, feedback, settings and short comparisons are very welcome, especially from different GPUs!
Datou retweeted
【MiniMax H3のキャラクター表現を改善するオリジナルLoRA学習とパイプライン検証】
動画生成モデル「MiniMax H3」でオリジナルキャラクターの品質を向上させるため、LoRA学習と各種パイプラインを比較検証した知見が共有されています！
RefMod（latent references）単体では、映像全体が過度に磨きすぎた不自然な印象になりがちです。そこで、静止画8枚、音声、会話クリップ14本を組み合わせてキャラクターLoRAを構築することで、t2v本来の自然な質感を維持したままキャラクターの一貫性を担保できるのが大きなポイントですね。
学習環境にはModalのH200 GPUを使用し、約41分で完了しています！推論環境はRTX 3090で、LoRAの適用強度は0.8がスイートスポットとのことです。1.0にするとファンデーションを厚塗りしたような見た目になってしまいます。
リファレンス画像がライティングを強く引き継ぐ点や、トレーニングプレビューを過信してはいけないといった実践的な教訓も参考になります！
#MiniMaxH3 #StableDiffusion
Datou retweeted
MiniMax H3 のキャラ入れ替え LoRA
(Character Swap v1.1) を試しました
結果は動画を見ての通り、
元の動画の動きを忠実にコピーしつつ、
キャラクタの一貫性も維持できてます
以前に LoRA なし V2V で何とかしようとして
苦労したのが嘘のよう
ワークフローなどはリプ欄に
羽根付きキャラは、途中で一瞬羽根が消えたりしてますが、元動画との造形（体形）が違うので、これは仕方ない気はします
むしろ、造形が違うキャラでも入れ替えが出来ているのは、凄いんじゃないでしょうか
Astra を使っても MiniMax H3 の V2V が
失敗する原因をいろいろ探ってましたが、
どうやら
「元動画の動きが速すぎる」
と失敗することが多いようです
それで、元画像の速度を 0.8 倍とかにして
V2V (ref2va) で生成し、その後に 1.25 倍すると
結構うまくいくことが分かりました
これを Astra 君に伝えて、
「そうやって生成して！」とお願いして
できたのが下の動画です
比較で Wan3.0/SeeDance 2.5 の生成動画も
つけました
この動画、通常の速度で MiniMax H3 で V2V
しても全然うまくいきませんでした
0.8 倍にしたら（この動画は）いけましたが、
動画によってはもっと遅くする必要が
あるかもしれません
xitter.cf/IsekaiBardGuild/status…
Datou retweeted
🔥炸了！国产AI 3D模型Hi3D V3.0正式发布，细节反超GPT-6 Astra！V1.0同步开源！
同一张人物图对比：GPT-6 Astra只能搭出基本形体，Hi3D Agent+V3.0在轮廓、体积和局部细节上明显更贴近原图，直接拉开差距。
核心亮点：
🔹业界首个可商用2048³超高精度国产自研模型，纹理最高8K
🔹V1.0正式开源，基于Sparc3D，训练数据扩至约200万，切片成功率90%，已进生产流
🔹Agent闭环：意图理解→生成→持续修改→工程交付，支持MCP，可直接进Blender/游戏引擎
开源地址：hi3d.ai/hi3d-1-0-open-source
GitHub：github.com/hi3d-official/Hi3…
官网：hi3d.ai/
#科技 #国产AI #Hi3D #3D生成 #Agent #开源 #GPT6Astra
Datou retweeted
Today we are releasing JST-2, the world's first full-stack physics-native world foundation model.
Many world models learn in pixel/video space. We went back to first principles and built the entire model stack in 3D.
Generate a 3D scene in seconds and expand it in real time. Create characters with motion built in, all through one general architecture.
Datou retweeted
字节的回应：公司总算力中，只有约10%用于视频生成模型的训练和推理。
中美国情差异很大，以Token消耗为例 ，中文在线披露的一组数据：
国内所有Token消耗份额中，AI短剧与视频生成高居首位，占比达55%；电商和营销又占了24%；在美国列Token消耗第一的软件开发只有15% 。
2026年前8个月，中国上线了大约43万部微短剧，是去年全年的13倍，其中超过90%都使用了AI技术。全国微短剧用户已超过8亿，2025年市场规模突破1000亿元。
大西洋对岸美国的数据是这样的：美国58%的Token消耗用于编程开发和企业自动化。
一边在写代码、搞企业自动化，一边在拍短剧、做广告、卖东西。
有人称：AI时代，中国人再一次把火药做了烟花。
但我不这么认为，核心原因是：美国的企业端软件生态决定了做coding能赚很多钱，所以OpenAI放着那么多C端用户不培养也要去和Anthropic争coding市场；但在中国自始至终都没有形成很好的企业软件市场，在B端单纯靠coding市场很难赚钱，中国单纯靠去搞coding赚钱根本不现实。在C端娱乐市场，尤其是在AI生成视频这块，本身创意视频传播力就强，再加上用户广泛，所以比较繁荣，竞争也激烈。
Datou retweeted
感谢分享网址，于是我也用这个网站测了一下 DeepSeekR1，对，你没看错，就是 2025 年 1 月推出的 DeepSeekR1 模型。根据分析我们可以看到 DeepSeekR1 的指纹与 Claude Sonnet 5.5 高度重合，达到了 99.4%的相似度。那么是不是就说明 DeepSeekR1 蒸馏了 Claude Sonnet 5.5 呢？🤣🤣
This quoted post is unavailable.
Datou retweeted
呃😅袁腾飞这种想法不是古典派润人思想吗？已经有无数人实践过了，结果就是你以为你养了个美国人，其实你养了个住在你家里的美国陌生人。
以前很多家长砸锅卖铁移民，或者拼命把孩子生在美国，图的是什么？图的是一口纯正的加州腔，图的是孩子不再背负第一代移民的口音包袱，觉得只要“语言零障碍”，孩子就能丝滑融入美国主流社会。
这逻辑乍看没毛病。但这些人忘了一个最致命的代价：当孩子真正“融入”了美国主流，他第一个融入不进去的，就是你那个中国式的家。
家庭是靠什么维系的？不是靠你说的英语有多标准，而是靠情感表达方式、价值观和代际伦理。在中国家庭里，父母的爱是“唠叨”、是“干涉”、是“我为你牺牲了这么多”；但一个土生土长的美国孩子（ABC/ABD），接收到的底层逻辑是“个人主义、边界感、独立与自由”。
然后你用中国式的自我牺牲去爱他，他用美式的心理防御机制来反抗你。你们之间也许没有语言障碍（有些情况有），但有文化鸿沟。他觉得你越界、操控、有毒；你觉得他冷漠、自私、没良心。
能在美国主流社会混得开的华人孩子，往往都有一个共同点：他们极度认同美国社会的游戏规则。这套规则的核心之一，就是成年即脱离原生家庭的羁绊。18岁搬出去，自己背学贷，逢年过节发个短信，这在美国父母眼里叫“独立”，在你眼里可能这就叫“白养了”。
而且吧，你以为没口音就能当白人？太天真了。你在美国主流社会是个“永远的外国人”，你的孩子哪怕英语是你的母语，在白人眼里依然是个“Asian”。而因为你不教中文、不教中国文化，他甚至融不进华人圈子。孩子成了一个精神上的孤儿，一个父母眼里的“异类”。这种身份认同的撕裂，才是很多ABC抑郁、迷茫的根源。
你想要他懂美国，又想让他懂你；
你想让他独立自由，又想让他承欢膝下；
你想让他逃离华人阶层的底层内卷，又想让他保持中国传统的孝道温存。
天下哪有这么两头占便宜的事？一切皆有代价。
Datou retweeted
Sharing our new work: UniForge3D generates detailed 3D meshes from sparse-view inputs with high fidelity! The code has been released and welcome to try it!
Paper: arxiv.org/abs/2610.15357
Project Page: xishuxishu.github.io/uniforg…
Code: github.com/xibushu/UniForge3…
Datou retweeted
Looking for something to play on Switch 2? According to Metacritic, The Drifter is in the top 10 highest rated for the platform EVAAA! :O
Datou retweeted
The Character Swap LoRA is out! Feel free to try it and share what you make.
It’s far from perfect, it’s just a LoRA, but it can work surprisingly well in some cases, like the examples below. I hope to work on an improved version soon together with @akatz_ai.
LoRA:
huggingface.co/Alissonerdx/M…
ComfyUI workflow: huggingface.co/Alissonerdx/M…
And there's an AI-generated video (video, voice, but it's my real voice, just in English hahaha) here hahaha to explain the Workflow planner node and some other things about LoRa: youtube.com/watch?v=D037QMeQ…