Designing a Luxury Handbag in NVIDIA Omniverse with Claude

Hanen Bdioui

4 November 2026 15 mins to read

I built a fully configurable luxury handbag, with a cinematic promo video, inside NVIDIA Omniverse — and I wrote almost none of the code by hand. This is the full workflow: concept, procedural 3D modeling, an AI-assisted debugging loop with Claude, the problems we hit, and how a Dell Pro Max 16 Plus with an NVIDIA RTX PRO 5000 Blackwell handled the rendering.

Project introduction

The goal was a luxury handbag that a client can configure live in NVIDIA Omniverse and then see in a polished promo video — all from one tool.

I wanted to explore a practical question: how far can one designer go in NVIDIA Omniverse with an AI assistant writing and debugging the code? Although I’m not a USD or Python specialist, I understand what makes a premium bag look right and can identify issues directly in the viewport. Claude handled the entire technical side, from writing and testing the code to debugging and refining the final result.

The project had four objectives:

A real 3D model, not a mock-up. Every part is real USD geometry: body, flap, handles, chain strap, hardware, lining and an interior pocket.
A configurator. Model, size, leather type, print, colours, hardware metal, text, logos and clasp style all change live in a panel inside Omniverse.
A promo animation. A camera move, a working turn-lock clasp, a flap that opens, a product reveal and a chain that falls with real physics.
Production-quality output. Path-traced video at up to 4K, rendered locally on a mobile workstation.

Everything runs as a custom Kit extension (handbag.configurator) inside a USD Composer-based app built from NVIDIA’s Kit App Template.

Figure 1: Hero render of the finished bag, e.g. Maison Aurèle in black quilted leather with the gold chain

Figure 2: Screenshot of the Omniverse window with the Handbag Configurator panel open beside the viewport

3D modeling workflow

The bag is fully procedural: Python generates every mesh from a handful of parameters, so changing the width or the flap length rebuilds the whole bag in under a second.

I chose procedural over sculpted geometry because a configurator needs parts that adapt to each other. When the body gets wider, the flap, stitching, handles, strap rings and even the graphics card inside must follow. Each part is computed from the same body description.

The body

The body is built from horizontal cross-sections (rings) stacked from the base to the top. Each ring is a superellipse, so one “corner roundness” value moves the shape from boxy to soft.

On top of that base shape, three effects make it read as leather instead of plastic:

Panel bulge: the front and back panels swell slightly in the middle.
Gusset pinch: the side panels fold inward towards the top, like a real gusset.
Wrinkles: a low, smooth displacement so no surface is perfectly flat.

The base has a rounded bevel, and four metal feet sit underneath.

The main compartment

The bag is a hollow shell, not a solid block. The outer wall rolls over a rim at the top and continues down the inside as a separate lining surface, with its own fabric material.

Inside, I added:

a slip pocket on the back wall, with a leather binding on top;
saddle stitching around the pocket and a centre seam that splits it into two slots;
stitching around the rim of the opening.

The flap

The flap is the hardest part, and it was rebuilt several times (section 4). The final version is a solid 3 mm leather piece:

it starts on the back panel, where it is stitched on, then follows the body’s real cross-section over the top and down the front;
it has an outer face, an inner face and a painted edge band, like real edge paint;
its side edges behave like stiff leather near the rounded top corners, so they stay smooth instead of creasing;
a slot is cut through it for the turn-lock knob, finished with a metal eyelet and a back plate.

Hardware and straps

Turn-lock clasp: an oval domed plate on the flap and a pill-shaped knob fixed to the body. An alternative “interlocking rings” design is available.
Top handles: rolled or flat leather, or chain, attached with metal rings.
Shoulder strap: a gold chain or a leather strap, hanging from rings held by leather tabs and rivets on the side seams.
Details: contrast piping on the seams, saddle stitching everywhere leather is joined.

Materials

All surfaces use physically based materials. Claude generated nine tileable leather texture sets with normal and roughness maps: nappa, pebbled calf, saffiano, suede, patent, croc-embossed, snake-embossed, quilted and woven. Printed patterns, such as florals, tartan, leopard and a monogram of your own initials, are generated in the chosen colours and blended with the leather grain.

Figure 3: The bag closed, three-quarter view, showing the quilted texture, piping and stitching

AI-assisted design with Claude

Claude acted as my technical partner: I directed the design and judged the results in the viewport, and Claude wrote, tested and fixed the code.

I worked with Claude in the Claude desktop app, connected to my project folder and to the running Omniverse app. That setup mattered. Claude could read the Kit App Template, write the extension’s Python files, run tests, and read Omniverse’s own log after every change.

How a typical change worked

Most of my requests were short and visual. I would take a screenshot, circle the problem in red, and write one line, such as “the strap rings don’t touch the bag” or “the flap only opens halfway”. Claude then:

Diagnosed the cause from the geometry, not just the symptom. For the floating rings, it found that the rings used the bag’s nominal width, while the side gussets fold inward near the top, leaving a 1 cm gap.
Changed the code in the extension.
Tested offline before touching my scene. Claude built a test harness that runs the extension with Omniverse’s own USD libraries outside the app, rebuilds every bag model and preset, and renders quick preview images.
Measured, not guessed. Many checks were numeric: the gap between two surfaces in centimetres, whether every corner of a bag stays inside the camera frame, how many times the falling chain touches the bag.
Let Omniverse hot-reload the extension and checked the Kit log for errors.
Reported back with what changed, what was verified, and what it had not seen yet, usually the final RTX render, which only I could look at.

I then looked at the result in the viewport and either moved on or sent the next screenshot.

Figure 4: The Iterations

The iteration loop · 6 steps

A failed offline check sent the change straight back to the code, so most problems never reached my viewport.

What I directed, and what Claude decided

I made the design calls: the look of the clasp, how far the flap opens, which camera moves to use, the product reveal. Claude made the engineering calls and explained them. Examples:

A soft hinge instead of a rigid one. When the flap opened as a stiff “L” shape, Claude replaced the hinge with a bend model, so the leather’s curve relaxes as it lifts.
Physics instead of keyframes for the chain. The chain drop is a real rope simulation with gravity, friction and collisions, baked into the animation.
An exact camera fit. Instead of guessing camera distances, the camera solves for the distance at which every corner of the bag fits in the frame.

Where Claude pushed back

Twice I asked for things Claude would not build as asked, and in both cases it proposed an alternative:

I asked for Louis Vuitton, Gucci, Dior, Fendi and Saint Laurent logos. Claude explained these are registered trademarks, so recreating them would produce counterfeit designs. Instead it built original fictional collections and an option to load logos I am licensed to use.
For the clasp, my reference was Chanel’s interlocking “CC”. Claude built an original “interlocking rings” clasp with the same intertwined feel.

For the NVIDIA graphics card inside the bag, I load the official logo from my sponsorship kit as a decal, rather than having it redrawn.

Figure 5: One of my annotated screenshots (a problem circled in red) next to the fixed result

Design challenges and solutions

Almost every fix came from the same pattern: a problem I could see in a screenshot traced back to a geometric cause Claude could measure and correct.

Modeling and Geometry Challenges

Challenge	Root cause	Solution
The first bag looked basic and plastic	A single flat material on a smooth loft	Nine procedural PBR leather sets, coloured prints, soft bulging panels, gusset pinch, piping and stitching
Strap rings floated about 1 cm from the bag	Rings placed at nominal width; gussets fold inward near the top	Rings anchored to the real side surface on stitched leather tabs with rivets
The flap looked like a paper-thin blade	A zero-thickness sheet starting in mid-air above the top	A 3 mm leather shell traced from the body’s own cross-sections, with a painted edge
No inside to the bag	The body was a closed solid under the flap	Hollow shell with fabric lining, rolled rim, slip pocket and divider
The clasp looked like a generic fastener	A cylinder and a capsule	Domed oval turn-lock; slot cut through the flap with eyelet and back plate
A dark slit appeared at the sides	The flap floated above the rim and was narrower than the mouth	Flap rests on the rim and spans the full opening
Creases appeared at the flap’s top corners	The edge hugged the rounded corner too closely	“Stiff leather” edges near the corners only
The flap looked laid on, not sewn	Stitching only ran around the front	Saddle stitching around the entire flap, including where it joins the back

Animation and Camera Challenges

Challenge	Root cause	Solution
The promo opened zoomed in	Camera distance used only the body, ignoring the strap and handles	Camera fit to the full bounding box; later an exact fit solved per keyframe
A bump in the camera speed	The camera took the larger of two distances, creating a kink	One smooth spline through log-distances: under 5% speed change per frame
The flap opened as a stiff “L”	A rigid rotation about one hinge	A bend model: the flap folds at the hinge and its curvature relaxes as it lifts
Steps appeared along the flap’s edges while opening	Each column folded at a whole row, not a straight line	Fractional hinge position per column, so the fold runs straight
Top handles blocked the flap	Handles stood in the flap’s path	Handles moved behind the hinge and tilt back 82° as the flap opens to 150°
The chain fell too fast and barely touched the bag	Real gravity is fast; handles and the open flap weren’t obstacles	Slow-motion speed ramp, collisions with handles and flap, friction measured in time, not steps
The graphics card was hard to see inside the bag	Camera looked in at 58°, so the front wall hid the card	Interior shot at 72°, aimed at the card’s face
The card reveal cut the ending short	The video length was fixed	Video length adapts: reveal plus enough time for the hero shot and chain drop

Two lessons stood out. First, a precise screenshot beats a long description: circling the exact spot got the right fix almost every time. Second, numeric checks caught problems before I saw them, such as the camera clipping a tall strap during the descent from the bird’s-eye view.

Figure 6: before/after pair — the rigid flap opening vs. the soft-bend opening

NVIDIA Omniverse visualization

The final deliverable is a 25-second path-traced promo film, rendered inside Omniverse with Movie Capture, plus live configuration in the RTX viewport.

Live configuration in the RTX viewport

Every change in the panel rebuilds the USD geometry and materials directly in the open stage. In RTX Real-Time 2.0 the viewport stays interactive while I change leather, prints, colours or hardware, so I can show clients variations in real time.

The extension writes standard USD: meshes with UVs, UsdPreviewSurface materials, lights and an animated camera. The bag can be exported as a self-contained .usda file with its textures, so it opens in any USD tool.

The promo film

The promo is fully generated from the panel: one click builds the camera path, the clasp and flap animation, the card reveal and the chain physics as USD time samples.

Time (s)	Shot
0–7	Bird’s-eye view of the bag, a slow turn, a spiral down, and a close-up on the clasp
7–10.5	The knob turns 90°, the flap folds back, the top handles lie back
~11	The camera looks down into the lined compartment
13–15	The graphics card rises out of the bag and turns to face the camera
15–18	Showcase: a slow orbit around the floating card, with a light sweep across it
18–20	The card sinks back into the bag
20–25	Hero shot; the gold chain falls onto the bag and settles on the floor

A few techniques made it feel like a real product film:

One continuous camera spline. No cuts between phases; the camera eases in, flows through every shot and eases out.
Exact framing. For each wide keyframe, the camera distance is solved so the whole bag, strap included, fits the frame for that angle and lens.
Physics for the chain. An 80-point rope with gravity, friction and collisions against the bag, handles, open flap and floor, played in slow motion.
Light that moves. A narrow softbox sweeps a highlight across the leather and the card, and a soft light fades in inside the bag as it opens.

Effects and rendering

The panel exposes RTX settings that matter for a premium look: depth of field that follows the bag, bloom, motion blur, film grain with vignette, colour grades, a seamless studio backdrop and a glossy reflective floor.

For the final render, I used Movie Capture’s path tracer:

Quality: Path traced, 128 samples per pixel for tests, 256 for the final.
Resolution: 1920×1080 for tests, 3840×2160 for the final.
Output: PNG sequence for the final, so I could grade and edit without compression. MP4 for quick reviews.
Motion blur: 6 sub-frames per frame, so the falling chain blurs naturally.

Figure 7: 3–4 stills from the promo: bird’s-eye, clasp close-up, card showcase, chain on the floor

Hardware performance

The whole project — modeling, live configuration and final rendering — ran locally on one mobile workstation, with the RTX viewport staying above 140 FPS while I worked.

Disclosure: I am a Dell Ambassador, and Dell provided the Dell Pro Max 16 Plus used for this project. The NVIDIA graphics card shown in the promo appears as part of a brand sponsorship.

The machine

Component	Specification
Laptop	Dell Pro Max 16 Plus
GPU	NVIDIA RTX PRO 5000 Blackwell (laptop), 24 GB VRAM
System memory	64 GB RAM
CPU	Intel® Core™ Ultra 9 Processor 285HX
Software	NVIDIA Omniverse Kit 110.4 (USD Composer-based app from the Kit App Template), RTX Real-Time 2.0 and path tracing

Rendering the promo

My first test render captured 241 frames and encoded an 11.4 MB MP4 in about 10 minutes, roughly 2.5 seconds per frame.

For the final film, these numbers from my own render:

Measurement	Result
Final render settings	3840×2160, path traced, 256 spp, Mp4 Video
Frames / video length	599 frames / 24.9 s
Total render time	5h55m
Average time per frame	35.86s
GPU utilization during the render	94%
Dedicated GPU memory used	9GB
System RAM used	31.8 GB
CPU utilization	12%
Power source	On AC power (Dell Pro Thunderbolt 5 Smart Dock)

Figure 8: Task Manager, Performance tab, GPU page during the path-traced render]

The final 8K render images are also available as a ZIP download. The archive is over 39 MB, so I’ve provided it separately rather than embedding the full-resolution images in the article. 8K Render

The combination worked because each part did what it is best at: I judged the design, Claude engineered and verified it, Omniverse rendered it, and the workstation kept every step local and interactive.

Show, don’t describe. An annotated screenshot was the fastest way to get the right fix.
Ask for verification, not just code. The offline tests and numeric checks caught framing, gap and timing problems before they reached my viewport.
Physical thinking beats tweaking. The biggest improvements came from modeling how real leather, hinges and chains behave, not from adjusting numbers by eye.
Respect brands. Original designs and licensed logos only; Claude flagged the trademark issues before I had to.
Local hardware changes the workflow. With fast viewport performance and 24 GB of VRAM, I could iterate on a path-traced look without sending anything to a render farm.

The Workflow at a Glance

The project came together through a combination of AI-assisted development, procedural modeling, and real-time visualization. Claude helped me write and refine the Python code, while NVIDIA Omniverse brought the handbag configurator and animation to life. The final step was rendering the promotional video using the RTX PRO 5000 Blackwell GPU in the Dell Pro Max 16 Plus.

Figure 9: AI-Assisted Workflow from Concept Design to Cinematic Rendering in NVIDIA Omniverse.

Why the workstation matters

The project was developed and rendered on the Dell Pro Max 16 Plus, powered by the NVIDIA RTX PRO 5000 Blackwell Laptop GPU with 24 GB VRAM, 64 GB RAM, and an Intel Core Ultra 9 285HX, paired with a Dell UltraSharp Curved 49-inch QHD Monitor. This setup handled the procedural modeling, Omniverse configurator, and final 4K path-traced rendering.

Figure 10: My Workstation

This project showed how AI-assisted development and real-time 3D visualization can come together to turn an idea into a working experience. It was a great opportunity to explore new creative and technical possibilities with NVIDIA Omniverse.

Author
Recent Posts

Hanen Bdioui

Sales Development Representative at ChampionXperience

Hanen Bdioui is the Editor-in-Chief at ChampionXperience, where she leads content on CAD, VR, and emerging engineering technologies. She works with Dassault Systèmes as a Content Creator and SOLIDWORKS Application Engineer, and contributes as a technical writer for Engineering.com and EngineeRules, covering topics in CAD, PLM, and simulation. Hanen also creates specialized content for 3DEXCITE, the 3DEXPERIENCE platform, and Onshape, supporting engineers and companies in adopting modern digital workflows.