跪拜 Guibai
← Back to the summary

Turning a Child's Superhero Drawings into a 1-Minute AI Short with Seedance 2.5 Flash

Foreword

At nine in the evening, my son was drawing at his desk: on the left was the Monkey King, on the right a mecha warrior. He looked up and asked me, "Dad, who do you think is stronger?"

I couldn't answer right away. When I was a kid watching cartoons, I also used to wonder: if superheroes from different worlds fought, who would win? But back then there were no such technologies; I could only hold my toys and voice the characters myself, imagining a big battle.

Looking at my child, I seemed to see my younger self. At that moment, I realized: I should use the tech in my hands to do something for him. I wanted to use Seedance 2.5 Flash to turn my child's imagination into a 1-minute AI short film, making the Monkey King and the mecha warrior really battle before his eyes.

A child's aesthetic plus a programmer's obsession: Let's just do it!

1. Seedance 2.5 Flash: Double generation speed, price on par with the previous generation

For this 1-minute AI short film, I chose Seedance 2.5 Flash. It is a video generation model released by ByteDance's Volcano Engine on November 12, 2026. Its official positioning is direct: the closest effect to the flagship Seedance 2.5, with double the generation speed and a lower price. In the Seedance 2.5 family, the flagship handles 1080P and 4K high-quality production, while Flash targets high-frequency creation and batch content production, an option for "quickly getting ideas out with less trial-and-error cost."

In terms of parameters, it supports two resolutions, 480P and 720P, with a maximum of 30 seconds per video segment, and can take up to 10 images, 5 videos, and 5 audio clips in a single request. For the most troublesome part of short films, voiceover and sound effects, it provides a ready-made solution: native audio generation — when the picture comes out, the sound comes out too; lip sync, ambient sound, and sound effects arrive together, no need for a separate post-production dubbing alignment process. It also has video editing capability, allowing the same footage to be quickly adjusted into different versions, which is practical for content that requires repeated tweaks of pacing and shots.

Price is the key to making this approach work. Seedance 2.5 Flash's pricing is basically on par with the previous generation Seedance 2.0: about 0.44 yuan/second for 480P, about 0.99 yuan/second for 720P, and the 720P price is about 34% lower than that of the flagship Seedance 2.5. In other words, at the price of the previous generation's flagship, you get close to the current flagship's quality and double the generation speed.

0.png

2. Complete video generation workflow

2.1 Task thinking: telling a story in 1 minute

This time I didn't just want to generate "two heroes fighting" but a complete, warm story: a child draws, says he wants to see the Monkey King and mecha warrior battle; the father working nearby uses a large model to turn the battle into an animation; the son is moved after watching it — in his heart, the father who can make his fantasies come true is the real superhero.

With the story ready, duration was the first practical issue. Thirty seconds obviously couldn't hold the full line of "drawing — talking — battle — embrace," so I set the final video at 1 minute and split it into two segments: the first covers the son drawing and the father-son conversation, the second covers the superhero battle and the final embrace.

The real difficulty of multi-segment generation is not the effect of a single segment, but the seams: how the characters, scenes, and lighting transition naturally between the two segments, making it look like a single piece. Around this problem, I executed the following three steps.

2.2 Task steps

2.2.1 Define characters first: lock character consistency with three-view images

To ensure the characters and superheroes remain consistent across both segments, the first step is not to write video prompts, but to generate three-view reference images of the character designs first. With unified reference images, the two later segments have a common anchor.

I first chatted with DeepSeek, simply describing the character settings to it: the boy with a round face, short black hair, blue hoodie; the father in a gray sweater and thin-frame glasses; the Monkey King with golden fur, tiger-skin skirt, and golden staff; the mecha warrior in dark red bronze gold with a hexagonal chest plate. DeepSeek helped me expand these rough settings into specific visual detail prompts.

1.png

Then I sent these prompts to Doubao and used Doubao's image generation function to produce four sets of three-view character images.

2.png

2.2.2 Then solve the seam: use "light" to stitch the two segments together

Once the designs were set, I wrote the prompts for the two segments. To ensure natural transition, I had multiple rounds of interaction with DeepSeek, and it gave me an eye-opening solution: generate the two segments independently, but use light as the seam.

The specific approach: change the ending of the first segment to: "Warm golden light particles suddenly burst out, the glow fills the whole frame"; change the opening of the second segment to: "Warm golden glow spreads from the center of the frame, revealing the boy's face lit by light (reference @image5), he blinks, and the camera pulls back..."; then add 3–5 frames of white/golden flash between the two segments.

This way the seam is completely covered by the light, and the audience perceives a "transition" rather than a "frame jump." The design was striking, so I adopted it directly.

3.png

2.2.3 Finally, run the generation: DeepSeek Harness + Seedance 2.5 Flash

Only now did I actually call Seedance 2.5 Flash to generate video. If you want to try the effect first, you can directly use conversational generation at the Volcano Engine Seedance 2.5 Flash Experience Center.

4.png

But videos generated in the Experience Center have watermarks, so for the final output I switched to calling the API via Python. In the intelligent era, such API-calling code doesn't need to be written by hand — I passed the Volcano Engine API documentation for Seedance 2.5 Flash model calls directly to the DeepSeek Harness agent (the agent I currently use most frequently; see my collection DeepSeek Harness from Beginner to Advanced).

I created a new workspace, put the three-view reference images generated by Doubao and the two prompts into it, and explicitly asked DeepSeek Harness to first learn the API call method for Seedance 2.5 Flash from the official documentation, then generate the two video segments according to the reference images and prompts, and finally stitch them together.

5.png

6.png

DeepSeek Harness, after reading the official documentation, wrote the Python wrapper code: create a task, poll for status, download the result — all the way through.

3. Effect of the generated video

The final video looks like this:

The result is stunning, already reaching the production level of AI short films. Seedance 2.5 Flash can be foreseen to become one of the main tools for AI short films in the future; this performance is truly impressive.

4. Flaws and shortcomings

If I score the result on a 10-point scale, I give it 8 points. The deductions are mainly in two areas:

First, cross-shot consistency. Recurring props occasionally "multiply" — for example, the Monkey King's golden staff starts as one, but inexplicably becomes two after flying into the sky, which looks a bit jarring.

Second, native audio is a bit chaotic. The audio generation saves the post-dubbing workflow, but the stability and clarity of the voice still have room for improvement.

Still, flaws don't hide the merits. Seedance 2.5 Flash remains excellent: 30-second direct output, native audio, reference-image character locking, fast generation, and low price. With a limited budget and the need for batch output, this cost-performance ratio is very solid.

5. Summary

The final video is exactly 1 minute long. The last shot freezes on the father-son embrace, with the two heroes standing side by side in the background. After my son fell asleep at night, I watched the film again in the living room, still moved.

The next morning, my son was still relishing yesterday's video. He asked me, "Dad, can you generate a flying cat for me?" I smiled and said, "Yes."

The most charming thing about AI is probably that it doesn't dream for us, but turns "what if" into "real." You can also hand that shot you've never been able to make to Seedance 2.5 Flash. The boundary of imagination is always tested by doing.