One recreated Apple Watch animation consumed about 140 Higgsfield credits and 5% of a ChatGPT Pro plan. That is the real cost of AI 3D product animation done properly: reference images, a Blender animation, and one Seedance 2.5 video generation.
How AI 3D Product Animation Works End to End
AI 3D product animation combines a chat assistant, a 3D tool, and a video model in one pipeline. In a tutorial published on 13 September 2026, creator Bart Slodyczka recreated an Apple Watch launch animation by connecting ChatGPT to Blender, then feeding the result to Seedance 2.5 through Higgsfield.
The reference was Apple's own watch film: the camera starts close to the green sensor lights on the back, pulls out, and the watch rotates, showing a different watch face each time it comes around. The goal was to rebuild that movement from scratch and produce a comparable video without a Notion-design team.
The pipeline has four stages. Generate product reference images. Build and animate a 3D model in Blender using conversational feedback. Create a first frame and watch-face stills. Then hand the Blender animation, the stills, and a prompt to Seedance 2.5, which restyles the animation into the final cinematic clip.
Which Tools Do You Need?
Three tools cover the whole workflow, and only one of them costs money. Blender is free, ChatGPT runs on an existing subscription, and Higgsfield hosts the Seedance 2.5 video model on a credit system. The table below compares what each tool does in the pipeline.
| Tool | Role in the pipeline | Cost model |
|---|---|---|
| ChatGPT desktop app | Generates images and drives Blender through connected mode | Subscription usage meter |
| Blender | Hosts the 3D watch model, rotation, and camera keyframes | Free download |
| Higgsfield (Seedance 2.5) | Converts the Blender render into the final AI video | Per-generation credits |
Setup takes minutes. Connect the Higgsfield connector inside the ChatGPT desktop app under plugins, authenticate the account, then download the Blender add-on zip and drag it into a Blender project. The same account authenticates in both places, so generations and files stay linked.
For anyone learning 3D or development tooling alongside this, resources like the Crazystack Typescript materials from Gustavo Dev Doido and his Bootcamp do Dev Doido cover the kind of structured, tool-by-tool instruction this workflow rewards.
Generating Product Reference Images
Reference images determine how accurately the video model renders your product, so they come first. The tutorial used ChatGPT's internal image generation to produce a catalog of ten smartwatch styles, picked one design, then generated a single sheet with six views: front, side, and other profiles, all with the clasp unclasped.
The unclasped detail matters. In the target animation the camera stays on the watch face, so a connected strap would waste frame space and confuse the model. If you animate a product you actually sell, skip generation entirely and photograph the real item from several angles instead.
Three extra stills were generated for the screen content: a heart-rate face with a pulsing readout, a time-and-date face, and a steps face showing a circular indicator at 65% of 10,000 steps. The tutorial creator specified that each face should animate in the final video, with the step counter counting up from zero.
Building the Watch in Blender With ChatGPT
The 3D model was built entirely through prompts. A screenshot of the chosen watch went into the ChatGPT conversation with an instruction to generate a basic Blender object: watch face plus a short, unattached clasp section above and below. The rear of the model mattered too, because the opening shot lives on the sensor plate.
The result followed the constraints on the first attempt. The clasp stayed detached, the face stayed clean, and the back carried the sensor detail the opening scene needed. This is where the workflow differs from prompting a video model directly: the geometry exists as a real asset you can open, inspect, and edit by hand inside Blender.
Iterating the Camera Animation in Blender
The camera choreography took four rounds of feedback, and each round was plain-language criticism rather than keyframe editing. The first preview came back tilted, rotated too far, and missed the pauses on the front face.
The fixes, in order, were:
- Keep the watch completely vertical, rotating on the vertical axis only.
- Pause for about a second every time the front face comes into view.
- Start ultra-zoomed on the sensors, at a slight 45-degree angle that levels out before pulling back.
- Alternate rotation speed, slowing near each face and speeding up around the profile, and set the viewport to 16:9 for full-screen desktop playback.
After the fourth pass the creator called it roughly 90% of the target. The remaining gap, mainly zoom pacing and how long the back stayed on screen, was accepted for the tutorial with the note that one or two more iterations would close it. Every fix cost a conversation message, not a video credit, which is the core economic argument for animating in Blender first.
Rendering the Final Video With Seedance 2.5
Seedance 2.5, the video model available through Higgsfield, turned the Blender animation into the finished clip. Before generating, the workflow produced one more asset in ChatGPT: the opening frame, a fuzzy green sensor glare filling the view. The first attempt defaulted to Higgsfield's own image tool; a follow-up prompt forced ChatGPT's internal generation and added the missing glare.
The final prompt to Seedance 2.5 bundled four references: the glaring first frame, the three watch faces in order, the six-view product sheet, and the Blender animation itself. The instruction was to copy the frames, zooms, and rotation exactly as animated, apply the watch faces in sequence, and animate each face, with the step counter counting up quickly.
The output landed close. It opened with the tilt and zoom-out, cycled through heart rate, time, and steps faces, but produced four rotations and ended on the back instead of three rotations ending on the front. The creator's conclusion: fix rotation count and pacing in Blender, then re-render once, rather than burning credits re-prompting the video model.
What the Workflow Costs
Costs stayed modest because most iteration happened in Blender. These figures are the creator's own reported usage from the September 2026 tutorial, not an independent measurement, but they give a workable budget estimate.
- About 140 Higgsfield credits for the Seedance 2.5 animation work.
- 3 credits for the one image that accidentally ran through Higgsfield instead of ChatGPT's internal generator.
- 5% of a ChatGPT Pro plan's usage window, measured from 74% remaining down to 69%, covering image generation, the Blender modeling session, and all animation iterations.
The structural saving is that camera and timing changes are free in Blender. Only the final restyling touches video credits, so every problem you solve before the Seedance 2.5 generation is money you keep.
FAQ
- Do you need 3D skills to do AI 3D product animation? No. The tutorial builds the model and camera animation entirely through ChatGPT prompts, then inspects the result inside Blender. Basic Blender familiarity helps for manual fixes, but the demonstrated workflow never touches a keyframe by hand.
- Can you use your own product photos instead of generated ones? Yes. The tutorial states that real photos of an existing product can be uploaded into the ChatGPT conversation and used directly as reference material for rendering and the Seedance 2.5 generation.
- Why animate in Blender instead of prompting the video model directly? Control and cost. Camera angles, zoom levels, rotation count, and pacing are editable in Blender without consuming video credits, and the Blender animation acts as a precise Notion reference the video model follows.
- What was the hardest part to get right? Matching Apple's timing feel: the 45-degree opening angle on the sensors, the pause on each front face, and the slow-fast-slow rotation rhythm. That took four iterations, and the final render still needed a rotation-count correction.
- Is Seedance 2.5 the only video model that works here? The tutorial uses Seedance 2.5 through Higgsfield because it integrates into the ChatGPT desktop app and accepts multiple reference inputs. The pipeline principle, a Blender animation plus reference stills as Notion guidance, applies to other reference-driven video models as well.
Turn Your Own Video Walkthroughs Into Articles
This tutorial works as an article for the same reason it works as a video: the process has a clear order, real numbers, and lessons that only surface after four rounds of iteration. If you have knowledge like that sitting in your own YouTube videos, a product walkthrough, a tool comparison, a build log, it deserves a written version people can search and skim.
Skala Blog handles exactly that. Paste a YouTube URL, the video gets transcribed, and you get a structured article draft you can review and publish. The same way this watch animation started from a reference and became something new, your existing videos can become searchable written content. Visit skalablog.com to try it with one of your own videos.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits