How SPREEAI cut storage costs by more than 30x and doubled model FLOPs utilization with Lambda

When most people imagine virtual try-on, they picture Photoshop. A garment cut out and layered over a photo.

Commercial-grade try-on is a different discipline. SPREEAI's unified diffusion model has to solve four interdependent problems at once.

Understand the human subject

Before generating a pixel, the model has to read how a person stands, since posture and balance decide where the fabric falls. And the input is unpredictable. A brand shoots its catalog in one controlled studio pose. A shopper sends whatever photo they have. The model has to place the garment right every time, and keep the person looking like themselves.

Encode the garment

Fabric behaves. It stretches, drapes, stiffens, catches the light differently by weight and weave. The model learns how a structured blazer holds its shape, and how a soft knit gives and clings instead.

Transfer the pose and style

The garment was photographed on one body in one pose. The model has to re-render it onto the shopper's stance, re-draping it to fit how they're standing.

Preserve the scene

The model changes only what needs to change. If something hides part of the garment, it reconstructs what's behind it. If something belongs in front, like a handbag strap over the shoulder, it stays. The original lighting and setting carry through, so the result looks like the shopper's own photo.

Fidelity at that level takes work at every stage: high-resolution training, targeted upscaling, and extra headroom where fine detail is most at risk. Prints, patterns, and typography have to survive generation intact, and every output clears a set of quantitative quality gates before it reaches a shopper.

Solving all four together meant consolidating a multi-module architecture into one unified model. That speeds up development and simplifies validation, but demands the full memory footprint of a top-end NVIDIA H100 GPU. Around it sits a wider stack: vision-language models grade output quality at scale, flagging distortion and drift, while large language models automate configuration and orchestration across the training workflow.

SPREEAI trains with distributed data parallelism across multi-node NVIDIA HGX H100 systems on Lambda 1-Click Clusters, so every optimizer step needs a full gradient all-reduce, and interconnect bandwidth becomes the binding constraint.

Diffusion models stress the interconnect differently than large language models (LLMs) do. Attention cost in an LLM is dominated by a KV cache that grows with sequence length, while a diffusion transformer recomputes attention across the image at every denoising step. Multiply that across a full pass and any latency in the fabric compounds through the run, surfacing as GPU idle time.

Lambda's 1-Click Clusters run on an NVIDIA Quantum-2 400 Gb/s InfiniBand non-blocking fabric in a rail-optimized topology, delivering up to 3,200 Gb/s of peer-to-peer GPUDirect RDMA between nodes. Gradient synchronization stayed off the critical path.

Before the rebuild, model FLOPs utilization (MFU) sat in the low twenties even during active training windows. Workloads had been fragmented across features, some needing the HGX H100 system and some not, but all of them scheduled onto the same high-end hardware. GPUs also spent real time waiting on data that hadn't been pre-staged.

Lambda's ML engineers started with systematic workload profiling, tracing GPU idle time to three root causes: storage I/O stalls, an HGX H100 system allocated to jobs that didn't need them, and conservative numerical precision defaults.

That last one was the least obvious. Large parts of the training graph turned out to be stable at lower precision than they were running at, and right-sizing it reclaimed memory headroom and lifted per-GPU throughput.

SPREEAI then built experiment orchestration directly on Lambda's stack: H100 1-Click Clusters, Quantum-2 InfiniBand, and a persistent shared filesystem. Lambda supplied documentation, guidance, and working examples covering data pipelines and GPU observability, and SPREEAI wired Lambda's metrics APIs into their own dashboards.

MFU more than doubled, from roughly 20% to 43%. Queue starvation, meaning experiments blocked waiting on a GPU allocation, fell 74%.

SPREEAI handles static try-on today, for single garments and complete outfits. Next comes full layering, a shirt under a jacket or a dress over denim, and complete looks with accessorizing. Size-aware try-on is being built, so shoppers get a fit recommendation based on their photo instead of a static size chart.

The frontier is video.

Temporal consistency changes the problem. A model that produces a flawless single frame will still show artifacts once that frame sits in a sequence. Fixing it requires temporally consistent video diffusion, far more compute-intensive than image generation. The roadmap extends beyond a single clip to 360-degree video and custom sequences with distinct camera angles, so a shopper can check a garment from every angle, as they would in a mirror.

Try-on is just the entry point. The same platform extends across the commerce chain and into agent-to-agent ecosystems: wholesale procurement, where virtual visualization replaces physical sample shipments, and AI-generated fashion content that cuts catalog production costs while widening representation across body types, skin tones, and ethnicities.

Visit demo.spreeai.com, upload a photo, and see full outfits on yourself. Or pick from a library of virtual twins.

Read more about how SPREEAI trains their virtual try-on model on Lambda, and explore Lambda's 1-Click Clusters for your own workloads.