The argument in one line.
You can generate AI videos of any length by splitting a GPT Image 2 storyboard into 15-second Seedance 2.0 clips, using character reference sheets for consistency, and chaining clips together with last-frame extraction to create seamless transitions.
Read if. Skip if.
READ IF YOU ARE…
- A content creator with basic video editing experience who wants to produce 30+ second AI films without learning complex VFX software.
- Someone experimenting with AI video tools who has character reference images ready and wants a systematic pipeline to chain multiple Seedance clips together.
- A filmmaker or animator exploring GPT Image 2 storyboarding as a rapid pre-production step before animating with Seedance 2.0.
- You're comfortable with prompt engineering and have access to Higgsfield AI, GPT Image 2, and Seedance 2.0 tools.
SKIP IF…
- You're working with live-action footage or real actors — this workflow is AI-native and won't integrate existing video assets.
- You need frame-perfect consistency or photorealistic output; the pipeline accepts some stylistic variation between clips as a tradeoff for speed.
The full version, fast.
Long AI films become tractable when you stop trying to generate them end-to-end and instead chain 15-second clips that already know what comes next. The pipeline runs in three stages: generate a 12-panel storyboard with GPT Image 2 from a one-sentence prompt, build character reference sheets for every recurring figure, then animate the storyboard four panels at a time in Seedance 2.0 on Higgsfield, tagging the reference sheets inside the prompt so faces, bodies, and props hold across cuts. For action-heavy sequences where cuts jar, extract the last frame of each clip and feed it as the opening frame of the next. To go past 45 seconds, ask GPT Image 2 to generate the next 12 panels using the prior storyboard as reference, then repeat. Add no music, no subtitles to every prompt so editing stays clean.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:19
01 · Cold open
Plays both finished AI films (44s fight scene, 71s short) before naming any tools. Proof-first hook.
00:19 – 02:55
02 · Step 1: AI storyboard with GPT Image 2
Inside Higgsfield, uploads two reference photos (scientist in hazmat suit + robot companion) and prompts for a 12-panel storyboard at 16:9. Discusses fixing repetition by re-prompting individual panels.
02:55 – 08:48
03 · Step 2: Animate rows with Seedance 2.0
Seedance maxes at 15s, so the storyboard is cropped into rows of 4 shots layered on a 16:9 canvas. Each prompt assigns a per-shot time range using the storyboard's own text descriptions. Adds 'no music, no subtitles' tail. Introduces the 4-column character reference sheet for cross-clip consistency.
08:48 – 09:56
04 · First 45 seconds assembled
Three 15-second clips animated and combined into a continuous 45-second sequence covering the full 12-panel storyboard.
09:56 – 10:44
05 · Step 3: Extend the storyboard
Upload original storyboard + character reference sheets back into GPT Image 2; prompt for the next 12 panels with a continuation hint. Animate the new page the same way for a 90s total (trimmed to 71s for repetition).
10:44 – 15:00
06 · Step 4: Seamless transitions for action scenes
For motion that crosses clip boundaries (a chokehold mid-fight), use the Video Frame Extractor tool to save the last frame of clip N and upload it as the first-frame seed for clip N+1. Demonstrated on the fight-scene example.
15:00 – 15:47
07 · Higgsfield eligibility note + outro CTA
Side note about Higgsfield rejecting some uploaded reference images for copyright reasons (retry tends to work). Plays the full fight-scene example. Ends pointing to his 10-practical-tips video.
Lines worth screenshotting.
- Splitting a 12-panel GPT Image 2 storyboard into rows of four and animating each row as a separate 15-second Seedance clip is the simplest method for generating long-form AI video of any length.
- A character reference sheet generated from GPT Image 2 is the solution to the consistency problem that makes multi-clip AI films look disjointed.
- Last-frame seeding — using the final frame of one clip as the starting reference for the next — is what creates the seamless visual continuity that distinguishes a film from a clip collection.
- GPT Image 2's strength is reasoning about narrative structure, which is why a one-sentence prompt produces a 12-panel storyboard with text descriptions that can be fed directly into the video generator.
- Editing a storyboard in-place by tagging the reference image and describing the change — 'adjust shot 11 so it is not a repeat of shot three' — is faster than regenerating the entire storyboard.
- Adding 'no music and no subtitles' to the video prompt is a post-production decision made at generation time — skipping it creates downstream editing problems when stitching clips together.
- A 44-second scientist-and-robot sequence and a 71-second fight scene are both produced through the same three-step pipeline: storyboard, animate rows, chain via last-frame seeds.
Steal the 'turn the limit into the structure' move.
How to teach a constrained AI tool
The 15-second cap was Seedance's biggest problem. Tao turns it into the unit of the entire workflow — and the whole tutorial has a shape because of it.
- When demoing an AI tool with a hard ceiling (15s clips, 4 image refs, 4k token context), build the lesson around how that ceiling becomes the structural unit. 'Four shots per 15-second clip' is the entire video's spine.
- Open with the receipts. Play the finished 44s + 71s clips before naming a single tool. Demo-first hook earns the 15 minutes that follow.
- Name your workflows. 'Character reference sheet,' 'last-frame seed,' 'four-shots-per-clip' are all coinable phrases viewers can re-use. This is how a tutorial becomes a meme that other creators cite.
- Bury the most valuable thing 70% in. The transition-fix at 10:44 is the part of this video most likely to go viral as a short — Tao left it where the algorithm has to reward the watch-time first. Mod-Boss / JoeFlow tutorials should do the same.
- Tail-append your prompt boilerplate. 'No music, no subtitles' is a copy-paste rule that gets used 100% of the time. JoeFlow vocab + Mod-Boss session templates already do this; lean harder.
- Use ratio-of-effort to amplify the pitch: 'one sentence prompt → 12-panel storyboard' is the math that hooks the audience. Always state the leverage explicitly.
Terms worth knowing.
- Storyboard
- A grid of sequential image panels, each with a brief description, that maps out the visual shots of a film or video before production begins.
- GPT Image 2
- OpenAI's image generation model with strong reasoning capabilities, used here to produce multi-panel storyboards with text descriptions for each scene.
- Seedance 2.0 (Cdance)
- A video generation AI model available on Higgsfield AI that animates a reference image or storyboard panel into up to 15 seconds of video.
- Higgsfield AI
- An AI media platform that hosts multiple image and video generation models — including GPT Image 2 and Seedance — in a single interface.
- Character reference sheet
- A set of images showing a character from multiple angles or in multiple poses, used to help an AI video model maintain visual consistency across different generated clips.
- Last-frame chaining
- A technique for creating seamless AI video sequences by extracting the final frame of one clip and using it as the starting frame of the next generation, ensuring visual continuity between shots.
- Video frame extractor
- A tool that pulls individual still images from a video file at specified timestamps, commonly used to isolate the first or last frame of a clip for use as a reference image.
- Character consistency
- The challenge of keeping an AI-generated character looking the same — same proportions, colors, and features — across multiple separate video or image generations.
- 16:9 aspect ratio
- The standard widescreen format for video (width is 16 units for every 9 units of height), used by most modern TVs, monitors, and online video platforms.
Things they pointed at.
00:45toolGPT Image 2
02:58toolSeedance 2.0 (ByteDance)
12:30toolVideo Frame Extractor
15:21channel10 Practical Tips for Realistic AI Videos (his prior video)
Lines you could clip.
03:46
“There just isn't enough time inside those fifteen seconds to animate this entire storyboard. So I'm gonna split the storyboard up.”
05:32
“We need an additional character reference sheet so that when we generate the long AI video sequence of them, they actually stay consistent throughout the entire scene.”
04:42
“Just a one sentence description is gonna be enough to create a full storyboard.”
12:42
“Tell the AI to use the screenshot I just saved as the first frame and generate those next four shots starting from that initial screenshot.”
13:33
“Using this method, you can generate endless continuous shots for your AI films.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphorstory
00:00Here's the easy way to create AI videos. In this tutorial, I'm showing you exactly how I generated this realistic forty four second long fight scene and how I made the seventy one second long short film, and I'll show you how to extend them even more to be as long as you want. This is by far the simplest and easiest method to generate long AI videos.
00:19The first step is to use AI to generate a storyboard just like this one. It's gonna divide our long AI videos into these separate scenes. The tool we're gonna use for this is the newest GPT two image model.
00:32What this tool is amazing at is reasoning and generating text, and that's gonna come in really handy for us when we're trying to generate these lawn storyboards. To use GPT image two, I'm inside Higgs Field AI, and you can find it on the home page, or we can go to the list of image generators and find the new GPT image two model.
00:52For the first example I'm creating, I'm gonna generate an lawn AI video based around these two photos of the scientist in the hazmat suit and his robot companion who are exploring this toxic forest. So we have the GPT image two model chosen, and what I'm gonna do is upload those two reference images of my characters.
01:13Let's put those inside here. Next, I'm gonna ask the g p t image two model to create a full storyboard for us. The prompt can be super, super simple.
01:21Just a one sentence description is gonna be enough to create a full storyboard. Now you have some other options, uh, like choosing the quality, the resolution of the storyboard, and also the aspect ratio.
01:35I'd recommend setting this to 16 to nine. And then we just have to generate the AI storyboard. This is the full built out story.
01:43It's got 12 complete shots in it that tell a full story, and underneath each individual shot, there's a small text description. Those little text descriptions are gonna be super useful later on because we're gonna use those as part of the prompts for our AI video generator.
01:59I did notice sometimes that when you're generating a bunch of separate panels like this, there is some repetition. So if you look at this image in panel three, which is the robot gesturing upwards, it ends up being the same image as shown inside panel 11. So we need to make a minor edit to the storyboard so we don't have this repetition.
02:18So inside Hicksfield, what I'm gonna do is hit reference, which is gonna let me edit this storyboard. And so there's that storyboard added on into our image references.
02:29Let's delete all the other ones, um, and also change the prompt. What we wanna do is change panel 11 so that it's not a repetition of panel three, and that's basically exactly what I'm gonna tell the AI to do.
02:44Just adjust shot 11 so it's not a repeat of shot number three. And looking at the result, it's swapped shot 11 to one of the scientists, uh, inspecting some kind of a toolkit. So now that we have the storyboard sequence, how do we leverage this and turn it into a lawn AI video?
03:02We're gonna use a AI video model called c dance two point o. So if we have a storyboard, and what cdance can do for us is to animate all of these in a single video generation at the same time.
03:15To use the cdance video model, I'm gonna look through the video models on Higgs field, and the top one is cdance two point o. So for this AI video model, the maximum duration of each video generation happens to be fifteen seconds.
03:30So then how do we create a super long AI video sequence of our entire storyboard? Well, if we try to fit this entire storyboard, all 12 shots into a single fifteen second video, there just isn't enough time inside those fifteen seconds to animate this entire storyboard.
03:49So I'm gonna split the storyboard up and crop out each individual role.
03:55So this is the first role I've cropped out where I'm just taking the first four shots of that storyboard. And what we're gonna do is animate these four shots inside the fifteen second video sequence, and it should be able to produce a pretty good result for us. A quick technical point, you will need to layer the cropped roll on top of a 16 by nine image.
04:17This is just to make sure that Higgs Field can actually use this image as a reference. For the prompt that I'm gonna use, what I'm gonna rely on is the pregenerated text descriptions that's already inside the storyboard.
04:31So what we're gonna do is tell it to generate a scene using the shots in the uploaded film storyboard. Then for each of the shots inside this fifteen second video generation, I'm gonna give it a time frame of when I want the shot to happen.
04:46For example, the first four seconds. And then for the description of what happens in the shot, I'm basically just gonna copy the text description that's already described in the storyboard, and I'm gonna fill out the rest of the prompt using the same method.
05:02You'll also wanna add in this line at the bottom that says no music and no subtitles. I found that that just makes it much easier later on when we're putting everything together. Before we go and use this prompt though, there's an really important addition that we need to add into this to preserve character consistency.
05:21So if we just used the prompt exactly written like this, we'll get results like this, which does basically animate the scene as expected except the robot doesn't look quite the same. For starters in this sequence, the robot's legs look super long for some reason. He should be much shorter than that, and there's a lot of different variation in the way that the character looks between different video generations.
05:46We need an additional so that when we generate the long AI video sequence of them, they actually stay consistent throughout the entire scene. To create a character reference sheet, I'm still gonna use the GPT image two model, and I'm gonna upload the original shot of my robot in the forest.
06:04And then in the prompt, I'm gonna tell it to create a character reference sheet. I'm gonna put the prompt in the description so you can go and copy it, and the prompt is gonna give us this really useful character reference sheet to help us maintain consistency inside the videos. Now that we have all these different assets, we can start putting together our long AI video sequence.
06:24So first off, I'm gonna upload into Higgs Field the first four shots of our storyboard along with a character reference sheet of our robot. Let's drag that inside there.
06:34And inside the prompt, we're gonna write it exactly like I previously described. So first telling it to generate a scene using the shots inside the uploaded film storyboard. And here, I'm actually gonna make a reference to the film storyboard, uh, uploaded image.
06:52So I'll type at, which lets me tag different references, and then I'm gonna write out each of the individual scenes. Just basically copy in the text descriptions already in the storyboard.
07:05And then inside this prompt, I'm also gonna need to reference our character sheet for the robot as well so that the AI video knows exactly what the robot should look like.
07:16So here, I'll add an additional tag next to the word robot for our character sheet.
07:23I'm actually also gonna do this in a few other places as well. And then in the settings, I'm just gonna make sure that I'm using the full fifteen second video duration.
07:35Now let's generate the video and see what it looks like.
07:53It actually looks like it animated five shots here instead of the four shots that are prompted for, but that's not a bad problem to have. It still followed all the shots inside the storyboard accurately and the robot looks great.
08:09And then using the same exact method, I went ahead and animated the rest of the storyboard. So here's the animation sequence for roll number two of the storyboard, which is shots five through eight. And here are the last four shots of the storyboard.
08:30And now we have three separate fifteen second video clips that are in sequence to each other, which means that we can combine them together into a full forty five second AI video sequence, but it doesn't stop here. We can actually extend our storyboard as many times as we want to. So looking at the sequence of shots inside the storyboard, what if we wanted to extend this so that they start exploring the forest even more?
08:56What we can do is actually use GPT image two to generate the next 12 panels of the storyboard as well. So here's what we're gonna do. Inside the g p t two image generator, we're gonna upload the original storyboard reference.
09:11So this is the first 12 shots that we just animated. And then I'm also gonna upload a character reference sheet for my scientist and also the character reference sheet that I generated for the robot. Then inside the prompt, I'm gonna ask it to simply generate the next 12 panels of the storyboard, and here's the prompt I'm gonna use.
09:30Generate the next page of the storyboard from image one, which continues the story with 12 panels using the uploaded images. The robot reveals hidden knowledge of the forest, guiding the scientists to a deeper, more dangerous core where the source of the toxic outbreak lies.
09:47And this is the extended storyboard that GPT image two has created for us. And then I can use the same exact technique of animating four separate shots at the same time to turn this storyboard into three fifteen second video generations as well.
10:08And if you do the math, using the two storyboard pages, we'll end up with six fifteen second video clips, which is a ninety second video sequence. I found that some of the sequences that I generated were a little bit repetitive, so I ended up trimming it down to a seventy one second video instead.
10:28But you can use this method and extend your videos for as long as you want. I'm gonna put a link in the description for Higgs Field AI if you wanna go and generate your own long video sequences using GPT image two and c dense two point o.
10:44Now, one of the challenges we're gonna run into this especially when animating a more dynamic action packed scenes, for example, this fight scene that I created, is that because of how much action there is, I found that if we animate each of the roles inside the storyboard separately by themselves, it's much harder to combine them together seamlessly.
11:06So let's see how we can fix this problem. First off, for my fight scene, these are the characters and the environment that I want them to be in, and this is the storyboard and the prompt that I used inside GPT image two to create this.
11:21Now when I go and animate each row separately of the storyboard, each of the individual animations look really, really good. So this is the first role animated, and keep an eye on what happens in the last scene.
11:36The bounty hunter suddenly ambushes the female character and has her in this chokehold. Now when I go and animate the second row of the storyboard, it starts off with the first frame of them already engaged inside a fight scene.
11:51If we look at this individually, it's a really amazing looking AI video. However, if we try to combine those two separate fifteen second clips together, the transition is gonna look a little weird.
12:03So starting with the first video sequence and suddenly jumping into the second video sequence. That transition right there, it doesn't really make sense.
12:12She's in a choke hold and suddenly is free and engaged in a fight scene. So then how do we get the transitions to be seamless inside our lawn AI video? What we're gonna need to do is give the AI some extra information when generating each individual video clip.
12:30So first, I'm gonna use this tool called video frame extractor, and I'm gonna save that last image frame of our video sequence. Now when we go and generate the next four shots of our storyboard sequence, actually tell the AI to use the screenshot I just saved as the first frame and generate those next four shots starting from that initial screenshot.
12:54This is what it's gonna look like inside Higgs field first with all of our image references uploaded. And then inside the prompt, I'm gonna tell it to generate a scene using the uploaded film storyboard fight sequence starting with this image frame of the female character getting attacked.
13:13And then the rest of the prompt is written just like how it was done before, basically describing what happens inside each storyboard sequence.
13:22Now when we go and generate this scene, it should generate the next four shots starting with the initial image frames. And using this method, you can generate for your AI films. Here's a quick note about using Higgs field.
13:37When you upload image references, what Higgs field is gonna do is check the eligibility of each image.
13:46This is to avoid any copyright issues. Now if you're using, like, a celebrity or a scene from a movie or something, it's gonna get denied.
13:57But sometimes when you upload images of your own characters as well, it can also get denied. So the first time that I tried to upload my reference sheet for my bounty hunter character, it actually got determined as not eligible.
14:13But then when I tried again, the next time it was determined as eligible to be used. So if you try to do this and you upload an image of yourself or your characters and it gets denied, just try uploading that image reference a few different times and eventually it might work.
14:30Let's take a look at what the full animated fight scene looks like so you can get an idea of what this method is capable of.
15:21If you also want a complete breakdown of 10 practical tips to generate the most realistic possible AI videos, go watch his guide right here.
The Hook
The bait, then the rug-pull.
Tao Prompts opens by playing the finished pieces — a 44-second fight scene and a 71-second short film — before he names a single tool. The pitch is the proof, not the promise. Then the two title cards drop: GPT Image 2 for the storyboard, Seedance 2.0 for the motion.
Named ideas worth stealing.
02:55concept
Storyboard → Rows of 4 → Chain by Last Frame
The whole pipeline. Generate a 12-panel storyboard, split into rows of 4 shots, animate each row inside one 15-second Seedance clip with per-shot time ranges, chain clips together using last-frame seeds.
Steal forany 'use a constrained AI tool to make long-form output' tutorial — same shape works for AI music, AI animation, AI voice
03:50concept
Four Shots Per 15-Second Clip
Turn Seedance's 15-second cap into a structural feature by cropping the storyboard into 4-shot strips and prompting the model with explicit time ranges ([00:00-00:04], [00:04-00:07], ...) plus each panel's text description verbatim.
Steal forany constrained-window video tool — Pika, Runway, Kling
05:45concept
4-Column Character Reference Sheet
Use GPT Image 2 with a prompt that specifies four vertical columns showing front view, left profile, right profile, back view of the character. Plain barren background. Reusable as a character bible across image and video generations.
Steal forany AI workflow needing consistent characters across multiple generations
06:45concept
Tagged References in Prompts
Higgsfield lets you @-tag uploaded images inside the prompt (@image_1, @image_2). Lets you say 'the robot @char_ref_sheet walks along @storyboard_1' so the model knows which reference grounds which entity.
Steal forgeneral prompt-engineering pattern: named slots beat positional references
12:30concept
Last-Frame Seed for Seamless Transitions
When motion crosses a clip boundary (chokehold, fall, swing), extract the literal last frame of clip N with the Video Frame Extractor tool and upload it as the first-frame anchor for clip N+1. Eliminates the jump-cut tell.
Steal forany 'stitch short AI clips into a longer sequence' workflow — the #1 fix that separates 'AI demo reel' from 'short film'
04:57concept
Tail-Append 'No Music, No Subtitles'
Standard tail string added to every Seedance prompt so the model doesn't bake music or subtitle artifacts into the clip that you'd have to remove in post.
Steal forany video-gen prompt — clean output is easier to edit
How they asked for the click.
VERBAL ASK
10:35link
“I'm gonna put a link in the description for Higgsfield AI if you wanna go and generate your own long video sequences using GPT image two and Seedance two point o.”
Soft affiliate drop right after the 71-second payoff lands — value-delivered first, ask second. End screen at 15:21 cross-promotes his '10 practical tips' video.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
