Photoroom uses Nebius for model training and inference behind its image-editing products. In its technical presentation, the team describes training with 128 H200 GPUs linked by InfiniBand and calls Nebius its primary training partner. Its July 2026 engineering article describes testing memory checkpoints on a Nebius GPU cluster. Two models produced checkpoints of around 60 GB, restoring in 35–55 seconds. Warm-image startup fell from around 220 seconds to 35–45 seconds. Fully cold-node startup fell from around 430 seconds to 195 seconds, with image pulling accounting for most remaining time. A Qwen image-edit model restored roughly four to five times faster than normal loading and compilation. An AOT-compiled model scaled to four fresh replicas on cold nodes, each ready in about 195 seconds. Photoroom limits this checkpointing approach to models that fit on one GPU. Checkpoint production uses an idempotent job and a shared RWX volume.
Nebius describes large diffusion-model training with Slurm and Kubernetes orchestration, uninterrupted multi-month training runs, improved monitoring, FP8 model compilation and 300 TB of storage usage. Photoroom co-founder and CTO Eliot Andres credits Nebius infrastructure and expertise with helping scale training and saving engineering time. An April 2026 Nebius employee post reports threefold inference speed for Photoroom’s AI photo studio.
In October 2026, Photoroom describes its Vestiaire Collective pipeline running on H100 and H200 GPUs with capacity from Nebius. It combines its own models with FLUX.2 [klein] and DINOv2, using TensorRT and Triton Inference Server to optimize inference. The API removes backgrounds, adds shadows, centers items and corrects lighting before listings go live. Vestiaire Collective adds around 40,000 listings a day. The same processed photos appeared in its performance marketing ads, which became 2% more efficient, while digital authentication was unaffected. Some models receive only a few requests an hour and shut down when idle, restarting when traffic arrives. Photoroom says it worked with NVIDIA to put Dynamo Snapshot into production, bringing a new GPU server online in about a minute rather than around three and a half minutes.
Vestiaire Collective lifts liquidity 5% with Photoroom AI | Photoroom
13 Oct 2026 · www.photoroom.com · Web page
Photoroom says its Vestiaire Collective pipeline runs on NVIDIA H100 and H200 GPUs with capacity from Nebius. Vestiaire Collective adds around 40,000 listings daily. The same processed photos appeared in performance marketing ads, which became 2% more efficient, while digital authentication was unaffected. Photoroom reports using NVIDIA Dynamo Snapshot in production to bring new GPU servers online in about a minute rather than around three and a half minutes.
Vestiaire Collective lifts liquidity 5% with Photoroom AI | Photoroom13 Oct 2026 · www.photoroom.com · Web pagePhotoroom says its Vestiaire Collective pipeline runs on NVIDIA H100 and H200 GPUs with capacity from Nebius. Vestiaire Collective adds around 40,000 listings daily. The same processed photos appeared in performance marketing ads, which became 2% more efficient, while digital authentication was unaffected. Photoroom reports using NVIDIA Dynamo Snapshot in production to bring new GPU servers online in about a minute rather than around three and a half minutes.
The Ultimate AI Cloud
4 Oct 2026 · nebius.com · Web page
Photoroom co-founder and CTO Eliot Andres credits Nebius infrastructure, expertise and Slurm clusters with scaling training and saving engineering time. Nebius reports uninterrupted multi-month training runs, FP8 model compilation and 300 TB of storage usage for Photoroom.
The Ultimate AI Cloud4 Oct 2026 · nebius.com · Web pagePhotoroom co-founder and CTO Eliot Andres credits Nebius infrastructure, expertise and Slurm clusters with scaling training and saving engineering time. Nebius reports uninterrupted multi-month training runs, FP8 model compilation and 300 TB of storage usage for Photoroom.
How Photoroom Cut Inference Cold Start to Seconds - leads with the entity + outcome
23 Jul 2026 · Photoroom · Web page
Photoroom reports checkpointing tests on its Nebius GPU cluster, with roughly 60 GB checkpoints restoring in 35–55 seconds. Its deployment uses an idempotent checkpoint-producing job and a shared RWX volume to restore serving pods. Single-threaded torch.compile compilation avoids memory that the checkpoint tooling cannot dump. Shared-filesystem throughput constrains restore speed, and the approach is limited to single-GPU replicas.
How Photoroom Cut Inference Cold Start to Seconds - leads with the entity + outcome23 Jul 2026 · Photoroom · Web pagePhotoroom reports checkpointing tests on its Nebius GPU cluster, with roughly 60 GB checkpoints restoring in 35–55 seconds. Its deployment uses an idempotent checkpoint-producing job and a shared RWX volume to restore serving pods. Single-threaded torch.compile compilation avoids memory that the checkpoint tooling cannot dump. Shared-filesystem throughput constrains restore speed, and the approach is limited to single-GPU replicas.
Nebius Photoroom: Scaling training and inference for an AI visual creation application built for e-commerce
15 Apr 2026 · Nebius · Video
11:03–11:22: Photoroom describes training with 128 H200 GPUs on Nebius, connected with InfiniBand, and calls Nebius its primary training partner.
Nebius Photoroom: Scaling training and inference for an AI visual creation application built for e-commerce15 Apr 2026 · Nebius · Video · Starts at 11:0311:03–11:22: Photoroom describes training with 128 H200 GPUs on Nebius, connected with InfiniBand, and calls Nebius its primary training partner.
ChannelX reports that Photoroom’s Vestiaire Collective pipeline uses Nebius GPU capacity on NVIDIA H100 and H200 GPUs, with TensorRT and Triton optimizing inference. Vestiaire Collective adds around 40,000 listings daily. The processed photos also appeared in performance marketing ads, which became 2% more efficient, while digital authentication was unaffected.
Vestiaire Collective boosts sales 5% with Photoroom standardised product shots - ChannelX13 Oct 2026 · channelx.world · Web pageChannelX reports that Photoroom’s Vestiaire Collective pipeline uses Nebius GPU capacity on NVIDIA H100 and H200 GPUs, with TensorRT and Triton optimizing inference. Vestiaire Collective adds around 40,000 listings daily. The processed photos also appeared in performance marketing ads, which became 2% more efficient, while digital authentication was unaffected.
Photoroom develops AI software for creating and editing product images for e-commerce. Its web and mobile apps serve individual sellers, growing brands and enterprise teams, while its Image API lets marketplaces and retailers integrate image processing into their own workflows. Main tools include background removal and generation, product staging, AI shadows, virtual fashion models, product video generation and batch editing. Businesses use these tools to prepare product listings, maintain consistent catalog imagery and produce marketing visuals across sales channels.
Documented financing · private company
Amount$43MSeries B
Valuation at announcementValuation$500M, basis undisclosed$500M, basis undisclosed
Vestiaire Collective adopts Photoroom for all new listing photos
Vestiaire Collective now uses Photoroom to process photos for roughly 40,000 new listings daily. Photoroom reported that a randomized test increased listing liquidity by 5% and product-page views by 3–7%, while leaving digital authentication unaffected.
Photoroom deploys a new fashion fidelity model in its API
Photoroom's enterprise Image Editing API now uses its v2 product fidelity model for fashion image checks. The company says the model detects defects missed by its predecessor while retaining the existing endpoint and request format.
Photoroom co-founder Eliot Andres leaves his executive role
Eliot Andres announced his departure from Photoroom's executive team after seven years while remaining a board member. Tech.eu reported that former Groupon executive Javier Sotelo replaces him as CTO.
Photoroom uses Nebius for model training and inference behind its image-editing products. In its technical presentation, the team describes training with 128 H200 GPUs linked by InfiniBand and calls Nebius its primary training partner. Its July 2026 engineering article describes testing memory checkpoints on a Nebius GPU cluster. Two models produced checkpoints of around 60 GB, restoring in 35–55 seconds. Warm-image startup fell from around 220 seconds to 35–45 seconds. Fully cold-node startup fell from around 430 seconds to 195 seconds, with image pulling accounting for most remaining time. A Qwen image-edit model restored roughly four to five times faster than normal loading and compilation. An AOT-compiled model scaled to four fresh replicas on cold nodes, each ready in about 195 seconds. Photoroom limits this checkpointing approach to models that fit on one GPU. Checkpoint production uses an idempotent job and a shared RWX volume.
Nebius describes large diffusion-model training with Slurm and Kubernetes orchestration, uninterrupted multi-month training runs, improved monitoring, FP8 model compilation and 300 TB of storage usage. Photoroom co-founder and CTO Eliot Andres credits Nebius infrastructure and expertise with helping scale training and saving engineering time. An April 2026 Nebius employee post reports threefold inference speed for Photoroom’s AI photo studio.
In October 2026, Photoroom describes its Vestiaire Collective pipeline running on H100 and H200 GPUs with capacity from Nebius. It combines its own models with FLUX.2 [klein] and DINOv2, using TensorRT and Triton Inference Server to optimize inference. The API removes backgrounds, adds shadows, centers items and corrects lighting before listings go live. Vestiaire Collective adds around 40,000 listings a day. The same processed photos appeared in its performance marketing ads, which became 2% more efficient, while digital authentication was unaffected. Some models receive only a few requests an hour and shut down when idle, restarting when traffic arrives. Photoroom says it worked with NVIDIA to put Dynamo Snapshot into production, bringing a new GPU server online in about a minute rather than around three and a half minutes.
Vestiaire Collective lifts liquidity 5% with Photoroom AI | Photoroom
13 Oct 2026 · www.photoroom.com · Web page
Photoroom says its Vestiaire Collective pipeline runs on NVIDIA H100 and H200 GPUs with capacity from Nebius. Vestiaire Collective adds around 40,000 listings daily. The same processed photos appeared in performance marketing ads, which became 2% more efficient, while digital authentication was unaffected. Photoroom reports using NVIDIA Dynamo Snapshot in production to bring new GPU servers online in about a minute rather than around three and a half minutes.
Vestiaire Collective lifts liquidity 5% with Photoroom AI | Photoroom13 Oct 2026 · www.photoroom.com · Web pagePhotoroom says its Vestiaire Collective pipeline runs on NVIDIA H100 and H200 GPUs with capacity from Nebius. Vestiaire Collective adds around 40,000 listings daily. The same processed photos appeared in performance marketing ads, which became 2% more efficient, while digital authentication was unaffected. Photoroom reports using NVIDIA Dynamo Snapshot in production to bring new GPU servers online in about a minute rather than around three and a half minutes.
The Ultimate AI Cloud
4 Oct 2026 · nebius.com · Web page
Photoroom co-founder and CTO Eliot Andres credits Nebius infrastructure, expertise and Slurm clusters with scaling training and saving engineering time. Nebius reports uninterrupted multi-month training runs, FP8 model compilation and 300 TB of storage usage for Photoroom.
The Ultimate AI Cloud4 Oct 2026 · nebius.com · Web pagePhotoroom co-founder and CTO Eliot Andres credits Nebius infrastructure, expertise and Slurm clusters with scaling training and saving engineering time. Nebius reports uninterrupted multi-month training runs, FP8 model compilation and 300 TB of storage usage for Photoroom.
How Photoroom Cut Inference Cold Start to Seconds - leads with the entity + outcome
23 Jul 2026 · Photoroom · Web page
Photoroom reports checkpointing tests on its Nebius GPU cluster, with roughly 60 GB checkpoints restoring in 35–55 seconds. Its deployment uses an idempotent checkpoint-producing job and a shared RWX volume to restore serving pods. Single-threaded torch.compile compilation avoids memory that the checkpoint tooling cannot dump. Shared-filesystem throughput constrains restore speed, and the approach is limited to single-GPU replicas.
How Photoroom Cut Inference Cold Start to Seconds - leads with the entity + outcome23 Jul 2026 · Photoroom · Web pagePhotoroom reports checkpointing tests on its Nebius GPU cluster, with roughly 60 GB checkpoints restoring in 35–55 seconds. Its deployment uses an idempotent checkpoint-producing job and a shared RWX volume to restore serving pods. Single-threaded torch.compile compilation avoids memory that the checkpoint tooling cannot dump. Shared-filesystem throughput constrains restore speed, and the approach is limited to single-GPU replicas.
Nebius Photoroom: Scaling training and inference for an AI visual creation application built for e-commerce
15 Apr 2026 · Nebius · Video
11:03–11:22: Photoroom describes training with 128 H200 GPUs on Nebius, connected with InfiniBand, and calls Nebius its primary training partner.
Nebius Photoroom: Scaling training and inference for an AI visual creation application built for e-commerce15 Apr 2026 · Nebius · Video · Starts at 11:0311:03–11:22: Photoroom describes training with 128 H200 GPUs on Nebius, connected with InfiniBand, and calls Nebius its primary training partner.
ChannelX reports that Photoroom’s Vestiaire Collective pipeline uses Nebius GPU capacity on NVIDIA H100 and H200 GPUs, with TensorRT and Triton optimizing inference. Vestiaire Collective adds around 40,000 listings daily. The processed photos also appeared in performance marketing ads, which became 2% more efficient, while digital authentication was unaffected.
Vestiaire Collective boosts sales 5% with Photoroom standardised product shots - ChannelX13 Oct 2026 · channelx.world · Web pageChannelX reports that Photoroom’s Vestiaire Collective pipeline uses Nebius GPU capacity on NVIDIA H100 and H200 GPUs, with TensorRT and Triton optimizing inference. Vestiaire Collective adds around 40,000 listings daily. The processed photos also appeared in performance marketing ads, which became 2% more efficient, while digital authentication was unaffected.
Photoroom develops AI software for creating and editing product images for e-commerce. Its web and mobile apps serve individual sellers, growing brands and enterprise teams, while its Image API lets marketplaces and retailers integrate image processing into their own workflows. Main tools include background removal and generation, product staging, AI shadows, virtual fashion models, product video generation and batch editing. Businesses use these tools to prepare product listings, maintain consistent catalog imagery and produce marketing visuals across sales channels.
Documented financing · private company
Amount$43MSeries B
Valuation at announcementValuation$500M, basis undisclosed$500M, basis undisclosed
Vestiaire Collective adopts Photoroom for all new listing photos
Vestiaire Collective now uses Photoroom to process photos for roughly 40,000 new listings daily. Photoroom reported that a randomized test increased listing liquidity by 5% and product-page views by 3–7%, while leaving digital authentication unaffected.
Photoroom deploys a new fashion fidelity model in its API
Photoroom's enterprise Image Editing API now uses its v2 product fidelity model for fashion image checks. The company says the model detects defects missed by its predecessor while retaining the existing endpoint and request format.
Photoroom co-founder Eliot Andres leaves his executive role
Eliot Andres announced his departure from Photoroom's executive team after seven years while remaining a board member. Tech.eu reported that former Groupon executive Javier Sotelo replaces him as CTO.
