An experiment with a workflow that hands an AI a single storyboard sheet and tells it to "make this a film." It shows how to save credits and time with one prompt instead of generating scene by scene.
00:00What you'll learn
- Give the AI one storyboard and say "make a film" → a result from a single prompt, with no scene-by-scene generation. What this video covers: how to make the storyboard, how to put yourself in as the main character, whether to use Nano Banana 2 or GPT Image 2, and the strengths and weaknesses of each.
01:00Making the storyboard (with an LLM)
- He generates the prompt with Claude (or ChatGPT, any LLM). Example: "I need a prompt for GPT Image 2. Make a 5×5 panel storyboard (3×3, 5×3, whatever you like). An astronaut wakes up alone in a cryopod on a space station and tries to send a signal to Earth. Turn it into a short story."
- Claude invents the story to fit the reference (if you want a specific story, just write it yourself).
[01:00~] Nano Banana 2 vs GPT Image 2
- He runs the same prompt through both models (16:9, 4K). Nano Banana 2 = 10 panels, somewhat comic-book styled. GPT Image 2 = more photorealistic → he prefers GPT Image 2 (better results).
01:39Putting yourself in as the main character
- The method: take selfies from multiple angles (front, profile, various expressions), feed them to Claude and ask for a "character sheet prompt for an astronaut in a spacesuit" (with or without a helmet) → run that prompt through Nano Banana 2 / GPT Image 2 to produce the character sheet.
- Then feed those photos plus the storyboard prompt to Claude and ask it to "make the same storyboard with the attached character" → you get a nearly identical prompt to run through both models. Here too he prefers GPT Image 2 (personal taste).
02:56Other storyboard examples (easier without inserting yourself)
- Nike ad: the protagonist unboxes shoes → laces them up → the room suddenly becomes a stadium → running → winning → a back shot with Nike "Just Do It." GPT Image 2 is more photorealistic.
- Parrot (Pixar style): flying over New York and heading home. GPT Image 2 has stronger detail, contrast and saturation.
- Sherlock Holmes (manga style): the detective examines a crime scene → his thought process → he spots a figure behind a curtain.
- Parkour kid: leaping between buildings and running down the street, then gets distracted by a beautiful woman and slams into a wall.
- He compares the Nano Banana 2 and GPT Image 2 versions of each (pick to taste) and asks viewers which they prefer in the comments.
04:39Storyboard → video (Seedance 2.0)
- Converting with Seedance 2.0 (he uses Hit Film; other places work too). Settings: 14 seconds, 16:9, 720p. Two input images (the storyboard image + his own astronaut character sheet).
- The prompt is very simple: "Generate a photorealistic cinematic video following the attached storyboard panel by panel, maintain character consistency using the attached character sheet, no music, sound effects only."
- The result: it follows the frames well except for the last panel (looking out the window) — he's satisfied.
[05:53~] Generating the rest of the storyboards
- Nike: using the GPT Image 2 version, with an even simpler prompt ("cinematic video for Nike, panel by panel," 12 seconds). The first pass transitioned from living room to stadium too fast → regenerating fixed the transition but glitched the finish line → another regeneration gave him his favorite (glitches are easy to fix in the edit).
- Parrot: too cute and pretty to regenerate — he'll fix it in post instead.
- Parkour: regenerated several times. In the final version the figure doubles on impact and there are two women in red dresses → so he mixes part of a previous generation with part of this one in the edit.
- Sherlock: "You've been standing there since before I arrived" — not bad as a teaser for a longer format.
09:35Improving it in the edit
- Parrot: the nonsensical part where it flies in a window, is already inside, and exits through another window gets cleaned up with cuts.
- Parkour: he mixes the opening of one generation with a few shots from another into a new video.
10:35Pros and cons of this method (conclusion)
- The good: it saves time and credits (far fewer credits than frame-by-frame generation) and it's fast because it's a single generation. It's excellent for checking and proofing a concept and for showing a client what a future ad or short-form piece could be.
- The limits: if you want more control (a specific transition between frames, say), frame-by-frame generation — or building one segment at a time from two or three images — is still better. It costs more time and credits, but the control and the output are better.
- He closes by asking for thoughts in the comments and recommending his video on making cinematic AI ads.
📌 Bottom line
- Instead of generating scene by scene, this workflow hands an entire storyboard sheet to an AI (Seedance 2.0) and has it follow the panels to produce video — saving substantial credits and time.
- The pipeline: generate the storyboard prompt with Claude → image it with Nano Banana 2 / GPT Image 2 (he prefers the more photorealistic GPT Image 2) → build a character sheet from selfies to insert yourself as the lead → feed the storyboard and character sheet as two images to Seedance with a prompt saying "panel by panel, maintain character consistency, sound effects only."
- The output follows the frames reasonably well, but errors like transition pacing, glitches and duplicated figures mean you regenerate several times or mix multiple generations in the edit.
- The method is ideal for concept proofing and client pitches, but when precise control (a particular transition) matters, frame-by-frame generation remains the better choice despite costing more credits and time.