Fashion Workflow Revisited: We compare the latest AI video generation models in Sept 2025 as part of our 2D > 3D> AI Fashion workflow review
- Sep 8, 2025
- 3 min read
The Fashion Guild Podcast
Introduction
From 2D Pattern to AI Video: What’s Changed in a Year?
We’ve just dropped a new podcast episode putting today’s video-generation models through their paces on a real fashion workflow, and the results might surprise you. If you care about taking a garment from 2D pattern to 3D model, into photo-real stills and finally into moving image, this is the one to watch. Grab a brew and hit play below.
Last year's workflow
Over the past year we’ve been refining our workflow to support businesses with their workflow to incorporate digital, 3D and AI.
Almost a year ago to the day, we showed how we could take a complex jacket-and-dress pattern in Style3D, build a virtual model of it in 3D, upscale to a clean, photo realistic image, and use the latest models at the time to turn that into believable video. Last year, the blockers were clearly with the AI video stage of this model: ghosting, “morphing” anatomy, and styling drift that turned modern editorial into 1980s catalogue. Watch the original video here.
What changed this year?
This year we tightened the pipeline. We kept Style3D as a 2D>3D software of choice, but we took the draft render from this, and appled it to a classic upscaling (within Freepik), plus this year gave it a second pass with Topaz Gigapixel (cloud). This allowed us to feed the video models a sharper, less noisy reference image.
We then fed the upscaled image to the latest in video generation models. Some were version upgrades from those available last year, others were brand new models that weren't even launched this time last year.
We tested 6 different models this year, which were
Wan 2.2
Runway4
Minimax 0.2
Seedance 1.0 Pro
Kling 2.1
Google Veo 2 and Veo 3
So what improved?
A fair bit - just not uniformly. Several contenders cleaned up the worst artefacts: better fabric behaviour, far less ghosting, and more natural micro-movement (hands, hair, hem). In side-by-side runs, different models excelled or failed in different areas - but all showed some improvement. For example, Runway 4 showed nicer fabric rendering than our 2024 tests but still suffered from stiff, “robotic” hands and tonal drift (even hair colour). Mininax impressed on knit and fur texture but wrestled the odd camera movement when all we wanted was a clean studio take. Cling 2.1 delivered one of the most consistent faces/eyes across a full 10-second clip—yet hands again felt plasticky in places. Google Veo3, despite cost and hype, introduced pose and anatomy errors we couldn’t overlook.
The business reality?
On the bright side, AI Images are now reliably photoreal - and great for pose/scenery variations. These are precision images, built off actual sewing patterns and 3D models - making them perfect for use by fashion professionals who don't have the time or budget for full multi-day photoshoots.
But fashion-grade video remains the bottleneck for fidelity, control, and economics. Our verdict is similar to last year’s: 3D design is production-ready, image generation is campaign-ready with the right guardrails, but video still needs another turn of the crank - unless you have the time and buget to keep running multiple attempts to various generation models until you hit lucky with a few seconds of usable footage.
Watch the full episode of our latest podcast to see the full results of our tests, settings, and commentary—and tell us which models you want us to test next in the youtube comments.
👋 About The Fashion Guild
The Fashion Guild is a global consultancy and innovation hub committed to advancing creativity, efficiency, and sustainability in the fashion industry. Co-founded by industry veterans Peter Gallagher-Witham RCA MDes and Jon Smith, we help brands remain competitive by integrating the best of AI and 3D technology.
