Gordon Chen

Ph.D. @ MMLab, NTU

[Gordon Chen @ChenGordon19]

@Rancho99437786 @realNingYu @debfx @liuziwei7 @qhnmoon If the query tile is the same size as the frame, the method reduces to standard self-attention. Smaller tiles reduce peak attention memory from O(N^2) to O(BN) where B is the tile size, at the cost of additional tiling overhead and repeated global K/V access. Compute is the same.

[Gordon Chen @ChenGordon19]

For the first time, open-source video diffusion models such as Wan can generate high-fidelity 4K videos without any additional fine-tuning, entirely at inference time!

🚀 Excited to share our recent paper, CineScale! 🎉

🌐 Project Page: https://t.co/3SdAqCv2Vn

[Gordon Chen @ChenGordon19]

@MKay88905412917 @realNingYu @debfx @liuziwei7 @qhnmoon We’re working on adding ComfyUI support for CineScale. Stay tuned!

[Gordon Chen @ChenGordon19]

@Otnashii @toyxyz3 @realNingYu @debfx @liuziwei7 @qhnmoon Yes. Although CineScale was not developed specifically for MiniMax H3, it is theoretically applicable to all RoPE-based video models. Since both H3 and Wan2.2 use 3D RoPE over temporal and spatial dimensions, adapting CineScale to H3 should be straightforward!

[ハカセ アイ(Ai-Hakase)🐾生成AI・ComfyUI @ai_hakase_]

🎬 AI動画の「運任せ」を卒業！キャラ崩れを防ぐ神技術「Prompt Relay」

「顔が変わる⋯」「背景がワープする⋯」そんな悩みはもう過去のものです！
最新のAI動画技術で、一貫性を完全に支配できるようになります✨

👇️ このスレッドで解説していきます！

#PromptRelay #LTXVideo #動画生成AI

[Gordon Chen @ChenGordon19]

@hckinz @sunbaolong_2001 Check out our Project Page: https://t.co/bacWSZuJbJ

[有趣的80后程序员 @sunbaolong_2001]

Wan 2.2 is changing the game for AI Video action control! 🎬🚀
I tested LTX 2.3 vs Wan 2.2 using the "Prompt Relay Timeline" technique.
Result? Wan 2.2 performed a 5-action sequence with incredible precision.
Character consistency stays 💯 while movements shift frame by frame.
Full comparison & Workflow: 👇
🔗 https://t.co/R35mOFsaxV
#Wan22 #LTX23 #ComfyUI #AI视频 #PromptRelay

[Stable Diffusion Tutorials @SD_Tutorial]

LTX 2.3 + prompt relay 🙃🧐 (setup+workflow)

>For more Controlled event driven video generation
>with Kijai's custom node...

detailed post👇👇
https://t.co/diYhHihJ9G

[R@aiaicreate @aiaicreate]

LTX2.3を用い、Prompt relayとKeyframesを組み合わせたComfyUIワークフロー。「2027年のChatGPT自己認識」をテーマとした生成例。両者の併用は設定がシビアな反面、成功時の高い効果を指摘。詳細なワークフロー公開。

#ComfyUI #LTX23

URLはリプ���️

[Ziqi Huang @ziqi_huang_]

🎬 VChain is accepted to #ACL2026 Findings! 🚀

#VChain is an inference-time framework that injects reasoning signals (*Chain-of-Visual-Thought*) from MLLMs to guide video generation for better consistency.

🔗 Website: https://t.co/qbm4usAAVb
📄 Paper: https://t.co/vHphF0rP75

[R@aiaicreate @aiaicreate]

ComfyUIでLTX 2.3モデルとPrompt Relayノード活用。低VRAM（6GB）環境向け動画生成ワークフロー。タイムライン別プロンプト制御による詳細���動画調整と画像生成含むフルパイプライン構成。RTX 3060、16GB RAMでの30〜40分生成事例も。

#ComfyUI
#AI動画

URLはリプ⬇️

[Esha Sharma @aistudynowcom]

I just ran a new zombie chase test using our LTX 2.3 @LTXStudio Prompt Relay @ComfyUI workflow. I purposely made this scene extremely difficult. We added full-body running, multiple zombies, store shelves, scattered objects, and fast camera movements.
Standard video models usually fail right here. They break the environment completely.
But look at our result. Prompt Relay kept the exact action and the entire digital scene perfectly consistent.

[T8star @T8star_Aix]

Ai动画230-开源向闭源进军！从时间戳到物理运动双重控制视频生成！KJ复现LTX2.3版Prompt Relay，VBVR更新240K版，工作... https://t.co/yXQUXPcgPS 来自 @YouTube

#ai #人工智能 #ComfyUI #T8star #promptrelay

[Esha Sharma @aistudynowcom]

I just built a new ComfyUI workflow combining LTX 2.3 and Prompt Relay. It gives us absolute control over our video edits.
You don't have to guess when an action occurs anymore. We can generate a 20-second video, and you decide the exact second every single event happens.
Youtube Video is coming. I'll post the video for you soon.

[Ziwei Liu @liuziwei7]

🤩Excited to see that our *Prompt-Relay* enables fine-grained temporal control in video generation for multi-event storytelling @ChenGordon19

🧑‍💻*Prompt-Relay* has already been incorporated into Wan and LTX.

- Project: https://t.co/l4C3NdEMRQ
- Code: https://t.co/bAPuzAw0Qt

[Gordon Chen @ChenGordon19]

@mikeromana1 @SlipperyGem Yes our code builds prompt relay around wan2.2. We may add comfyUI support at a later time :) https://t.co/am2MzrN8HU

[Gordon Chen @ChenGordon19]

@cfryant @toyxyz3 Prompt Relay allows you to control exactly what happens at each second of the video :). I’ve linked my Project Page here: https://t.co/bacWSZuJbJ

[Gordon Chen @ChenGordon19]

@toyxyz3 I recommend using additional prompts to form transitions between events. As two prompts that are totally unrelated may cause cuts in between segments. Having these additional prompts will make the video more seamless :)

[AI Search @aisearchio]

What a crazy week in AI! 🚀

Opus 4.7
Qwen 3.6
Ternary Bonsai
GPT Rosalind
AniGen
Happy Oyster
Lyra 2
HY World 2.0
Prompt Relay
MotifVideo
OmniShow
Gemini 3.1 TTS
& more!

Watch the full recap:
https://t.co/a1CYNnvDaG

[AIQUEST @AiquestAcademy]

Text-to-video models suck at storytelling. Prompt a sequence of events and the concepts just bleed together into a messy hallucination.

Prompt Relay fixes this. It modifies cross-attention at inference time so each video segment only sees its assigned prompt.

The result is perfect timeline control:
• Zero retraining
• Zero extra compute
• Plug-and-play

Project demo page is available.
#AI #VideoGeneration