Co-founder & CTO @pika_labs | ex @StanfordAILab @Stanford | Co-inventor of DDIM and SDEdit
Joined March 2020
- Tweets453
- Following1.8K
- Followers9.3K
- Likes3.1K
Pinned Tweet
So proud of our team 🩵 Pika H3 Flex debuts as the #1 open weight model on DesignArena’s t2v board—and generates 15 seconds of video + audio in as little as 7 seconds.
Can’t wait to see what you create. Give it a try and feel free to DM me your thoughts! 😁
Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.
It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.
So, yeah. It’s kind of a flex 💅
Congrats @DrJimFan !
We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing Flora, a suite of open models to parse the physical world and turn raw footage into rich training signals.
Flora already powers our in-house data engine, labeling 200K hours of videos a day on 500 GPUs. We created it to scale the frontier for robot learning. Now you own the weights too.
Flora gives you:
- High-quality, timestamped captions to curate and filter human egocentric and robot data;
- Dense progress rewards for quality control and RL post-training;
- Rubrics scoring for real-world benchmarking
Try it today, links in thread:
Chenlin Meng retweeted
We launched Pika H3 Flex a few days ago, and creators have been pushing it to incredible places. A few of our favorite creations so far:
1. Every cut in this montage came from one 15-second generation. Pika H3 Flex handles multi-shot like a pro here.
nitter.gq/GlitterPixely/status/2…
Chenlin Meng retweeted
Amazing to see all the momentum around Pika H3 Flex already!
Everything we put into this model was designed to give creators the flexibility and speed to explore with precision.
In that spirit, we’re making Pika H3 Flex open source from day 1. You can find the code on GitHub, the weights on @huggingface, and start playing with it in @ComfyUI and diffusers right now.
Code: github.com/Pika-Labs/Pika-H3…
Weights: huggingface.co/Pika-Labs/Pik…
Diffusers:
huggingface.co/docs/diffuser…
ComfyUI: docs.comfy.org/tutorials/vid…
Chenlin Meng retweeted
Pika 把原本很慢的 MiniMax-H3，通过蒸馏 + 后训练，变成了只跑 6 步就能出比较高质量、运动自然，并且自带同步立体声的视频模型。
但万万没想到 huggingface 社区的AK @_akhaliq 大佬把进一步把它和更激进的 2/4 步蒸馏融合的基础——官方已经压到 6 步了，社区再继续压到 4 步甚至 2 步。进一步提升效率：
-4 步版本：约 1.5–1.66× 更快，且作者测试中画质不输甚至更好（尤其叠在 Singularity 上）。
-2 步草稿：约 2.4× 更快（适合快速试构图/运动），但偶尔会失败，音频也较差。
-8 步 HD：稍慢一点，但细节更高。
使用建议（ComfyUI）：
-下载对应文件到 ComfyUI/models/loras。
-在常规 H3 工作流中加载 LoraLoaderModelOnly，强度设为 1.0。
-采样步数匹配文件（2/4/8），CFG=1.0，推荐 euler / lcm / res_multistep + simple scheduler。
-Flow shift 按文档设置（视频 FL2VA 约 6.0，Ref2VA 约 12.0，音频约 3.0）。
huggingface.co/akhaliq/Pika-…
解释：
-视频扩散模型（比如 MiniMax-H3）生成视频时，需要一步步“去噪”，每一步都要让模型算一次。
-原始 MiniMax-H3 大概需要 40～50 步 才能生成比较好的视频。
-步数多 = 生成慢（一个 5 秒视频可能要几分钟甚至更久）。
-蒸馏就是一种训练技巧：用一个已经训练好的“老师模型”（50 步版本）去教一个“学生模型”，让学生模型学会用更少的步数达到接近老师的效果。
-Pika 把 H3 蒸馏成了 6 步 版本（Pika H3 Flex），所以生成速度快很多，同时尽量保住画质，支持同步立体声音频。
Chenlin Meng retweeted
解説動画（YouTube）
youtu.be/7cog89wVP_A
Reddit投稿
reddit.com/r/comfyui/comment…
Chenlin Meng retweeted
Using #H3 Pika Flex 6 step ref lora from @pika_labs + my anime upscale pipeline, uses ref images + voice clips (cleaned up with MelBand) Claude added the subs + stitched the 3 scenes together. I just finished Little Witch Academia. I already miss this world and the characters.
Chenlin Meng retweeted
Tried it myself, and it's a really impressive model. Congrats @chenlin_meng and the Pika team!
Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.
It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.
So, yeah. It’s kind of a flex 💅
Chenlin Meng retweeted
Pika H3 Flexでさらに検証しましたが、結論、これはゲームチェンジャーかもしれません。この速度とコストでこの品質を出せるなら、私のフローを劇的に改良してくれそうな感覚です。
主に、Seedance 2.5でお馴染みのR2V（キャラシートや音声を参照で入れて生成）に複数人のキャラや音声を同時に入れて、色々制御できるか試しました。まだキャラが2人一緒に喋ったり、違うキャラが喋ったり、同じキャラが2人出てきたり、制御しきれていない箇所も多々ありますが、それでも全体を通して、かなりのクオリティで出してくれました。しかも前回問題だった日本語のイントネーションも、入力した音声をそのまま使ってリップシンクやダンスをさせることができました。
これでElevenLabsなどのちゃんとした日本語が使えそうです。しかも複数人のキャラと音声を自在に選んで制御できるので、できることの幅が大幅に広がりました。
動画生成は全てClaude Code上でOpus 5.5がColabのCLIを使って自動で行っています。私がしているのは基本こんな動画にしてという指示や生成結果にイチャモンをつけて修正してもらうことくらいです。
コストも動画の生成だけなら1秒あたり約0.5〜1円（15秒の動画1本で約17円）と、かなり安いです。
ツール
検証した動画AI
・Pika H3 Flex（MiniMax H3 ＋ Pika LoRA、ComfyUI 公式ワークフロー）
・音声をそのまま使う：MiniMaxH3AddGuide ノード（ComfyUI）
・実行：Google Colab G4（RTX PRO 6000）with Colab CLI on Claude Code
素材
・セリフ：ElevenLabs v4
・歌声の差し替え：ElevenLabs ボイスチェンジャー
・笑い声：ElevenLabs SFX
・曲：Lyria
・ボーカル分離：Demucs
動画
・検証／集計／構成／コード：Claude Code（Opus 5.5）
・映像：HyperFrames
・字幕タイミング／音声チェック：Soniox STT v5
Try out Pika H3 Flex on @sunra_ai !
Pika H3 Flex is now on Sunra.ai
The entire hotel staff versus one pika. The concierge, the bellboys, the chef and the maid chase it through the lobby, the lift, the kitchen and the pool. So fast you can't catch up.
→ 5 to 15 second clips with sound generated in sync
→ Start from a prompt, or set the first and last frame
→ Upscale to 1080p, 2K or 4K
Chenlin Meng retweeted
Pika H3 Flex is now on Sunra.ai
The entire hotel staff versus one pika. The concierge, the bellboys, the chef and the maid chase it through the lobby, the lift, the kitchen and the pool. So fast you can't catch up.
→ 5 to 15 second clips with sound generated in sync
→ Start from a prompt, or set the first and last frame
→ Upscale to 1080p, 2K or 4K
Chenlin Meng retweeted
What a fun video 🐰
Pika H3 Flex is now on Sunra.ai
The entire hotel staff versus one pika. The concierge, the bellboys, the chef and the maid chase it through the lobby, the lift, the kitchen and the pool. So fast you can't catch up.
→ 5 to 15 second clips with sound generated in sync
→ Start from a prompt, or set the first and last frame
→ Upscale to 1080p, 2K or 4K
Chenlin Meng retweeted
testing @pika_labs #pika H3 Flex with speed dial, 0.7 speed + my upscale 0.5m to 2mp pipeline (derope fast, animesr, face detailer, and anime cadance) + my edge runners style lora. All 1 click t2v wf. Had to create a low vram node to get it to not OOM.
github.com/matlowai/ComfyUI-…
Chenlin Meng retweeted
Creative control + faster generation = a serious upgrade for AI video. 🔥
Pika H3 Flex can generate 15-second clips in as little as 7 seconds, supports up to 12 creative references, separate audio stems, precision editing, and 4K finishing.
Even better? Open weights, 1.8x faster than H3 Max, and up to 80% cheaper.
Built for creators who want to experiment, refine, and bring their vision to life without slowing down. 🚀
@pika_labs
Check - pika.art/pika-h3-flex?utm_so…
🤝 Paid partnership
Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.
It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.
So, yeah. It’s kind of a flex 💅
Chenlin Meng retweeted
Seedance may be the best model out there, but the way everyone is pushing H3 is amazing!
The latest iteration comes from Pika and it’s a huge step forward:
Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.
It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.
So, yeah. It’s kind of a flex 💅
Chenlin Meng retweeted
🚀 AI video creation just got a major upgrade!
Meet Pika H3 Flex — designed to make video generation faster, editing more precise, and creative workflows smoother.
🎬 Generate videos in seconds
🎧 Split audio into 5 separate stems
🖼️ Use up to 12 references
⚡ Iterate faster while spending less
More control, fewer limitations, and endless creative possibilities. 🔥
🤝 Paid partnership
Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.
It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.
So, yeah. It’s kind of a flex 💅
Chenlin Meng retweeted
AI video creation just got a serious upgrade.
@pika_labs is taking AI filmmaking to the next level with Pika H3 Flex.
Built for creators who want more than a good first generation, it offers up to 5 separate audio stems, support for 12 reference assets, precision editing, and 4K finishing.
15-second videos in as little as 7 seconds? That's a major step toward faster AI filmmaking.
Excited to explore what's possible with Pika H3 Flex.
pika.art/pika-h3-flex.
Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.
It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.
So, yeah. It’s kind of a flex 💅
Chenlin Meng retweeted
Pika finetune of Minimax H3. Pretty good quality from my initial tests.
huggingface.co/Pika-Labs/Pik…
Chenlin Meng retweeted
Created my 15-second fantasy short film, “The Last Dragon Egg,” using Pika H3 Flex.
I used the model to bring the dragon hatching scene, frozen kingdom, and epic fantasy twist to life.
🤝 Paid partnership · 🤖 Made with AI
Meet Pika H3 Flex, debuting today as the #1 open weight model on DesignArena’s t2v board.
It’s built for professional iteration, with exportable audio stems, resolutions up to 4k, and faster-than-realtime generation. It can produce a 15-second clip with soundtrack in as little as 7 seconds.
So, yeah. It’s kind of a flex 💅