A 15 s vertical ad for a fictional hoodie: the chain we used and where it fails
Transparency: this ad is a portfolio piece of our own, made to test a video chain. Everything is AI-generated. The model, the brand (grawwww) and the hoodie are fictional: it is not a product for sale and not a client's ad. The commercial part, at the end, says what we do and do not offer.
Music: MiniMax-Music3
The problem
A vertical product ad asks for two hard things at once: the garment must move with believable volume and folds, and the embroidered logo must be legible. We tried the tools separately and each one failed at something.
What we did
Four shots and an end card, 15 s in all. The chain:
- Our own starting images, already perfect. The exact hoodie on a hanger, our model wearing that same hoodie in the street and on a rooftop, and a close-up of the crisp embroidery. What the starting image gets wrong, the video does not fix.
- Blender previs for the first three shots. A clay hoodie turning about 70 degrees; in the street, the camera pulls back and drops to a low angle; on the rooftop, a 35-degree orbit. On CPU, a few minutes per shot.
- A motion pass with MiniMax H3 (Ref2VA). It reinterprets the previs: clay becomes soft fabric and the mannequin becomes the model. It is an intermediate step: the final pixels are painted by Wan.
- Final render with Wan 2.2 Fun Control. It uses the depth of that pass (computed with Depth Anything V2 Small) as guidance. On the hanger it repaints everything; in the street, on the rooftop and on the logo it repaints only 25 %, to keep cars, buildings and embroidery.
- Edit and upscale. Generated at 720x1280, hard cuts, no speed changes, an end card in real type, and an upscale to 1080x1920 with NVIDIA RTX Video Super Resolution.
- Music generated with MiniMax-Music3, under a licence that allows commercial use.
Results
- 15 s vertical at 1080x1920, 4 shots and an end card.
- The hoodie comes out soft, with natural folds side-on, and the last shot ends on a readable wordmark.
- Time: about 5 to 8 minutes of GPU per shot, and about 25 to 30 minutes for the whole ad.
What still fails
- With people, the motion pass follows the idea of the previs (approach, low angle, orbit) but not the exact camera path.
- The embroidered logo is still a bit soft at this size, and small in the street shot.
- Wan alone does not hold small text in motion: the embroidery drifted from black to brown or the letters garbled ("grawrcww").
- A full repaint of a shot with a real environment makes the environment worse: it invented worse cars and turned a scooter into a pedestrian. Hence the 25 %.
- A pixelated cloud is left in the rooftop sky.
- We tried speeding up the people shots 1.2x because the models walk a little slow; it looked too fast and we removed it.
- We do not paste the logo in post. We tried it and it does not look integrated.
What a client gets
This chain (Blender previs, a motion pass with MiniMax H3 and a final pass with Wan 2.2 Fun Control) is now available for brand work and is delivered as a finished video, made from images of your product. This ad is a piece of our own: the hoodie is fictional and not for sale. What you can see working is on the 5 s clip page and the Brands page.
Write to us with your case: hello@grawwww.xyz. We answer in under 48 hours.
All the video and images in this article are AI-generated. Fictional model, brand and hoodie. Music: MiniMax-Music3.
Credits and licences
- Motion pass: Powered by MiniMax H3 (MiniMax). MiniMax H3 is licensed under the MiniMax H3 Community License Agreement.
- Final render: Wan 2.2 Fun Control (Apache-2.0), with depth from Depth Anything V2 Small (Apache-2.0).
- Previs: Blender. Upscale: NVIDIA RTX Video Super Resolution. Music: MiniMax-Music3.
- AI-generated: all the video and images. Fictional model, brand and hoodie.