PixVerse 总结与 OpenAI「Prompt to Production」直播的五个要点
PixVerse 总结与 OpenAI Prompt to Production 直播的五个要点
PixVerse 与 OpenAI 于 10 月 28、29 日合办两场 Prompt to Production 直播,演示如何用 ChatGPT、Codex 和 PixVerse 插件从创意简报走到成片。
PixVerse· @PixVerse · X·2026-11-20 22:19· 1 小时前AI 评分55
Five Takeaways from Prompt to Production with OpenAI
AI 导读
PixVerse 回顾 10 月 28、29 日与 OpenAI 合办的 Prompt to Production 直播。
正文 · 原文

On 28 and 29 October, PixVerse and OpenAI held two live Prompt to Production sessions, one for APAC and one for North America, attended by hundreds of people from around the world. Paul Hsu, who leads marketing for PixVerse in the US, hosted both. Tyler Ryu from OpenAI’s applied AI engineering team joined him each time, alongside creators from the PixVerse Creator Partner Programme: Aldo and Hidayat from Indonesia and Serge Donaldson from the United States. All speakers demonstrated how they moved from a creative brief to a finished video using ChatGPT, Codex, and the PixVerse plugin.
The sessions build on product integrations PixVerse has already built with OpenAI. The PixVerse plugin, natively available in ChatGPT and Codex, gives users an end-to-end workflow for video, image, and audio, and it recently added support for Canvas, PixVerse’s AI video project workbench. The PixVerse CLI turns Codex into an AI content studio, giving users access to video, image, and audio generation in one place. Both are available today, and the ongoing Prompt to Production video series on the PixVerse YouTube channel carries the work beyond the webinars.
How to set up the PixVerse plugin in ChatGPT and Codex

Tyler's walkthrough covered the setup in five steps.
- Install the app on your computer. Use the Codex app or ChatGPT’s agentic mode, in its local version, so it can reach files on your machine. To check whether it is running on your computer or in the cloud, look at the top of the chat: a cloud icon means you are in the cloud version, which cannot access files on your computer.
- Search “PixVerse” in Plugins. Two listings appeared during the sessions. The older one was labeled legacy and due to be retired, so choose the new one.
- Sign in to PixVerse. Installing prompts you to log in, which gives the agent a direct, authenticated connection to your PixVerse account.
- Call it by name. Type @PixVerse and describe what you want in plain language. Tyler’s demo asked for a running delivery robot shaped like a box with legs, and the finished video was ready in about two minutes.
- Let the bundled skills do the prompting. The plugin includes skills written by PixVerse’s developers that encode how to prompt PixVerse’s models. Tyler’s example: a bare request to flip a pancake can produce a pancake that simply turns into a cooked one, while a prompt written with the skill describes the spatula motion step by step and the flip works.
Plugin listings and menus change quickly, so what you see in the app may differ slightly from what the speakers showed on 28 and 29 October.
Takeaway 1: Use PixVerse video as an animation reference in Codex
Tyler builds games, and AI-animated characters tend to move oddly. His fix was to ask Codex for reference images of the character, feed them into PixVerse to generate a walking video, then have Codex trace that motion onto the game character.
From there he refined the result in plain language, for example pointing out that the knee stayed bent in the front view. By his estimate, the whole sequence took about 20 minutes, compared with two to three hours of hand animation or a motion-capture studio that costs significantly more.
He also showed a second use where the video becomes a reference that other tools can build on. A PixVerse video of a delivery robot became a 3D model rebuilt in Blender, then a small playable demo game where the robot animates as you click around. His takeaway was that AI gets you most of the way there, and a video model can supply the last touch of motion that is hard to get otherwise.
Watch the full tutorial on YouTube: Sharing by OpenAI: Creating a game with the PixVerse plugin in Codex.
Takeaway 2: Keep references simple for consistent characters in ChatGPT and Codex
Consistency came up repeatedly in the Q&A of both sessions, and the speakers gave some tips.
Hidayat's rule is that references do not need heavy detail. Three views of a character (front, back and side) plus one face close-up are enough, and the same approach works for objects, vehicles, buildings and scenery. He then tells Codex to check the local reference folder before it generates anything.

Serge added several refinements. Keep characters out of location references, because the model will otherwise try to put them in the frame. If a wide reference includes a small face, cover it with an opaque circle so the model uses your close-up instead. Give wardrobe and props their own reference sheets: a character who should be holding a knife can end up holding a phone if the prop is left to chance. For the strongest continuity between clips, extend from the previous video, which costs more credits. To add on, Tyler suggested keeping all references in one project folder, so Codex remembers the layout of a scene.
Takeaway 3: Storyboard first in ChatGPT, then spend PixVerse credits late

Aldo’s example was a short, funny ad about a commuter wearing noise-canceling earbuds who never notices a T-Rex behind her. He built reference images first, then a storyboard, and only then generated video. By his figures, a storyboard costs about 30 credits (USD 0.20) and a 30-second video about 1,000 (about USD 6.60). When the office setting felt wrong, he moved the scene to a train station at the storyboard stage for roughly 30 credits. When the ending disappointed him, he regenerated only the last 10 seconds for about 350 credits (less than USD 2.50), roughly a third of the cost of redoing the whole 30-second video.
He also found that over-prompting made his character stiff, and cutting the prompt down fixed it. His other habit is to test at 480p and upscale later. These figures come from his own project, so your costs will vary.

Takeaway 4: Block the shot in 3D with ChatGPT before you generate video
Serge works on animated series set in real and imagined places, and his problem was that video models would not respect the space. His solution is to give ChatGPT reference images of a location, such as a pier or a bay with a tilted lighthouse, and ask it to build a simple three.js model. He moves the camera around, frames the shot he wants, and saves that frame. Then he pairs the frame with his character references in PixVerse Canvas and asks PixVerse to apply the look to that composition.

The 3D model does not need to be detailed, because all it has to settle is framing and composition. Tyler pointed out that this kind of previsualization used to belong to big studios working in game engines, and it now fits a home budget. Serge described the result as feeling more like directing than prompting.

Takeaway 5: Write skills once, and keep a ChatGPT or Codex project per job
Aldo runs his work like a studio. A workspace holds his skills, his knowledge files and his projects, and each project has its own folder and its own chat so every decision stays in the conversation. A skill is a playbook for one job, such as prompting, story, thumbnails or captions. It is written once and reused, and whenever something breaks, he writes the fix back into the skill so it improves with each project. His thumbnail skill already knows his fonts, colors and layout, so he never explains them again.

Serge works the same way in Codex. He makes a project for each body of work and keeps related chats inside it for better context. He also tells the agent what a file is and lets it name the file and choose the folder.
Interesting questions from the floor
Attendees asked a mix of practical questions, and one came up in both sessions: what is the difference between ChatGPT and Codex? Codex started as a developer tool, and ChatGPT’s agentic mode runs on the same engine without the technical interface. Both run the PixVerse plugin.
Several of the PixVerse-related questions came down to how the two products share an account. Because the plugin connects to your PixVerse account, generation limits follow your PixVerse plan whether you generate on PixVerse directly or through ChatGPT, and anything you generate through ChatGPT also appears in PixVerse. Videos you made earlier on the PixVerse website, however, do not appear in ChatGPT.
Canvas, PixVerse’s node-based project workspace, drew questions too, and Paul showed how a reference image in one node feeds the next video and how separate clips can be stitched into a short film.

Just start
Aldo's advice for beginners was to start, expect a rough first attempt, and keep iterating.
Serge asked whether creativity lives in the mouse clicks or in the idea that made you learn the tools in the first place, and for him it is the idea. Agents, he said, let him get more of it out of his head.
Both creators started with an idea, and the tools helped them get it onto the screen. The combination of ChatGPT or Codex with PixVerse makes the process quicker to iterate on, which is how most of the work shown in these sessions got finished.
Watch the Prompt to Production with OpenAI series on YouTube
The sessions have been distilled into short tutorial videos on the PixVerse YouTube channel, as part of the Prompt to Production with OpenAI playlist.
来源:PixVerse · x.com
来源:RadarAI · aihot.news