The Less You Write Prompts, the Higher the Success Rate (Nano Banana Pro Best Practices) - Part 5: How to Create AI Short Dramas - Blog 2026/02/03
Continuing from last time.
This is the 5th installment on how to create AI short dramas.
This is the 5th installment on how to create AI short dramas, and this time it is Episode 7.
It was released yesterday.
First, please take a look at the latest short video.
AI Drama Episode 7 "Victory":
Escape Skull Island! Hikari's Natural Unrivaled Power
Runtime: 58 seconds
In this day and age, any video professional could easily produce this kind of video content (AI drama).
Due to the technological evolution of image and video generation, the barrier to entry has lowered so much that there is no longer any inherent advantage to using AI. It is terrifying.
However, for independent creators like us, being able to quickly bring a story we've created to life is like a dream.
Since one person can handle every task in video content production, I believe anyone who has their own book (script) should definitely give it a try.
Scene where a giant robot appears
I reused a reference from a battle scene I had already generated.
Prompt:
Please expand the top, left, and right sides of the screen. It is fine if the two people become smaller.
Quality & Style Modifiers: Hyper-realistic, 8k resolution, UHD, RAW photo, shot on 70mm, detailed skin texture, visible pores, subsurface scattering, cinematic lighting, volumetric lighting, depth of field, sharp focus, bokeh background, masterpiece, photorealistic.
Remix with the giant robot reference.
Prompt:
An orange, super-giant robot has descended from the sky. At its feet are two people fighting.
Quality & Style Modifiers: Hyper-realistic, 8k resolution, UHD, RAW photo, shot on 70mm, detailed skin texture, visible pores, subsurface scattering, cinematic lighting, volumetric lighting, depth of field, sharp focus, bokeh background, masterpiece, photorealistic.
Change the character's pose.
Prompt:
Please pose the two people at the robot's feet as crouching with their backs turned.
What you should be careful about is,
A giant orange robot descended from the sky. At its feet, the two people who were fighting are surprised by the massive vibration and are looking up at the robot.
If you give instructions for the robot's movement and the characters' poses at the same time like this, the generation accuracy will drop, leading to repeated trial and error.
Reducing the number of instructions (reducing the prompt) and repeating local edits will result in a higher success rate and, consequently, faster completion of the work.
A giant orange robot descended from the sky. At its feet are two people fighting.
Video generation uses Kling AI.
Prompt for Kling:
A colossal orange robot descended from the sky and landed forcefully on the ground. It crouched low, then sprang upright. Violent tremors and massive clouds of dust. Countless tiny fragments of shattered houses scattered everywhere. Massive amounts of sand and dust. Massive amounts of smoke. Flames in the background roared violently.
Japanese:
A colossal orange robot descended from the sky and landed forcefully on the ground. The robot crouched down and immediately stood up. Violent tremors and massive clouds of dust. Countless tiny fragments of shattered houses scattered everywhere. Massive amounts of sand and dust. Massive amounts of smoke. Flames in the background are roaring violently.
Please check the movement in the short video above.
The following parts are the gravity effects. Phrases like "countless tiny fragments scattered everywhere" and "massive amounts of sand and dust" are the key to AI VFX.
Violent tremors and massive clouds of dust. Countless tiny fragments of shattered houses scattered everywhere. Massive amounts of sand and dust. Massive amounts of smoke. Flames in the background roared violently.
Write as few prompts as possible
If you want to create AI dramas quickly and without spending money, write as few prompts as possible (reference + local editing is best).
Breaking down scenes and repeating local edits to gradually refine and complete a single image is the fastest and most cost-effective workflow.
The fewer instructions you put in your prompt, the higher your success rate will be.
Attempting to generate everything in one go using long prompts takes time for trial and error and consumes a large amount of credits, which costs money.
As the number of instructions in a prompt increases to generate every element of a scene at once, the generation accuracy decreases. In other words, you end up generating repeatedly to get the expected result, consuming a lot of time and credits.
This is a generation method you want to avoid in AI dramas.
Generating with Nano Banana Pro 4K
While "2K" is fine for standard use, "4K" is essential for AI drama production. From a cost-efficiency perspective, the optimal approach is to use "2K for iterations and 4K for the final version."
Since Nano Banana Pro in Dzine AI, which I use for simultaneous lip-syncing of multiple people, allows for 4K selection, I will give it a try.
Note: Nano Banana Pro in Adobe Firefly and Photoshop is 1K / 2K.
In image generation with a high density of information like the one below, image quality will be sacrificed, so you should generate in 4K.
Naturally, 4K consumes twice as many credits as 2K.
Note: In the official Gemini API price list, 1K and 2K are treated the same, and only 4K is priced differently.
Nano Banana Pro 2K: 2752 x 1536 px
Nano Banana Pro 4K: 5504 x 3072 px
Last week's report meetingAs explained in the report, there is a significant difference between 2K and 4K depending on how you write your prompts.
Although the internal implementation (whether it is native 4K generation or generation + super-resolution) is not officially specified, the decisive difference is that details not depicted in 2K are newly added.
In 4K mode, the inference capability is fully utilized for writing in details, so images generated in 4K from the start have an overwhelmingly higher amount of information than those scaled up from 2K using AI upscaling later.
An example where 4K is dramatically effective:
Infographics and diagrams
In 2K, small Japanese text within complex charts or flowcharts will become blurred and unreadable, but in 4K, there is a higher probability that it will be generated with sharp resolution.
Creating assets intended for cropping
Artwork for stock photo sales
Crowds and detailed backgrounds
When the screen is packed with elements, such as a "panoramic view of a cyberpunk city" or a "group photo at a party," 2K will simplify the details, whereas 4K has a higher probability of executing a detailed depiction.
If you are using Nano Banana Pro as a synthesis asset for Photoshop, you will need to generate it in "4K" on another platform before bringing it into Photoshop, as the Photoshop version of Nano Banana Pro does not allow you to select 4K.
For character bust shots or close-ups of faces, there may not be much difference, and the effect will vary depending on the content of the image and how the prompt is written, so some trial and error is necessary.
Since the descriptive capability of Nano Banana Pro 4K changes depending on how the prompt is written, it is better to set aside some time for verification at the beginning.
The following is a 4K generation, but since Creative Upscale is not required, it saves more credits than upscaling from 2K using Topaz or similar tools.
[Call for Participants]
Would you like to produce an "AI drama"?
This is being distributed as the February 2026 version of the "Would you like to produce an 'AI drama'?" event held last August. We will use Zoom, following the same style as our monthly report meetings.
I will hold it sometime early this month, and I will announce the date and time on note as soon as they are decided.Just like the report meetings, I would like to accept participation via Peatix.
Date and Time: February 6th (Fri) from 8:00 PM to 9:00 PM
Participation application page (Peatix) is currently being prepared
Generative AI used: Nano Banana Pro (image generation), ElevenLabs (voice generation), Suno AI (BGM generation), Kling AI (video generation), Domo AI (lip-sync)
App used: Adobe Firefly Boards (storyboarding)
As with last time, this is not a "course" to learn something.
It is a fun activity where we try to create a simple AI drama together, so it is not content that explains things carefully for beginners.
*I intend this to be an inventory of the AI drama production knowledge accumulated in my own head.
I will leave the live stream archive up, but the provision of AI drama data is for participants only, same as last year.
This is the archive from last August.
Looking at it now, it feels nostalgic. I am surprised that only half a year has passed.
Actually, "Hikari no Tennen Muso" was a sample (assignment) for this live streaming series. I never expected it to survive as updated content...
A place to experiment with Nano Banana Pro techniques
"Hikari no Tennen Muso" serves as a testing ground for the latest AI models and Nano Banana Pro techniques. Since it focuses on short videos, there are no detailed technical explanations, but you can check it out as "sample footage" of the latest techniques.
Create "one minute" "every day" by "one person" "like drawing a 4-panel manga"
In the case of AI dramas, it becomes "one minute" in "three days"A place to actively test cutting-edge AI technology
This initiative will run until March 31, 2026 (a decision on continuation will be made in late March)
Related Magazine:
Last updated: February 3, 2026 (Tue) / Published: February 3, 2026 (Tue)