2x performance boost with NVIDIA G4 instances
Model load times reduced from 1 hour to 1 minute
Out-of-memory GPU errors virtually eliminated
Serverless scaling for peak global workday loads
Enterprise-grade security and SOC 2 Type II compliance
ComfyUI uses Google Cloud G4 VMs with NVIDIA Blackwell GPUs to give global creative teams fast, serverless AI generation.
ComfyUI uses Google Cloud G4 VMs with NVIDIA Blackwell GPUs to give global creative teams fast, serverless AI generation.
Balancing precise creative control with increasing infrastructure demands
Founded in 2023 by Robin Huang and his co-founders Yolanda Yan and comfyanonymous, who goes by Comfy Anonymous, ComfyUI set out to redefine how generative AI models are deployed and customized. Built as an open-source framework, ComfyUI allows machine learning enthusiasts to run open-source visual diffusion models on their own hardware.
Unlike traditional single-prompt generators, ComfyUI exposes the inner architecture of diffusion pipelines by chaining together text encoders, diffusion models, and variational autoencoders (VAEs) to give creators granular control over every parameter.
This node-based graph approach closely mirrored industry-standard visual effects tools like Nuke, Blender, and Houdini.
As a result, technical artists from Hollywood, top advertising agencies, and global gaming studios quickly gravitated toward ComfyUI to produce high-stakes creative assets.
Transitioning to Google Cloud G4 instances powered by NVIDIA Blackwell doubled our performance while cutting model load times from an hour to one minute. The price-performance gain was an instant win for our team and our user community.
Robin Huang
Co-founder, ComfyUI
"When you talk to people creating ads for major fashion retailers or producing films for Hollywood studios, control is their first, second, and third concern,” says Huang. “They're willing to go to extreme lengths to achieve outstanding results."
However, providing this level of control created many infrastructure hurdles as more people needed to use the tools. Diffusion models require downloading asset files ranging from 50GB to 100GB, forcing artists on typical internet connections to wait over an hour just to initialize a model.
Local consumer GPUs frequently suffered out-of-memory crashes when attempting to swap or run complex fine-tuned models and custom nodes. In addition, operating a global creative platform resulted in intense, overlapping daily usage spikes during regional workdays across Europe, Asia, and North America.
To enable all users to stay focused on creativity and results, ComfyUI needed a modern cloud infrastructure offering high-bandwidth storage, top-tier GPU performance, and serverless scalability.
Architecting Comfy Cloud with GKE, NVIDIA Blackwell, and Google Cloud Storage
To solve the technical challenges, ComfyUI built Comfy Cloud on Google Cloud AI Hypercomputer to abstract backend hardware complexities while preserving deep local-level customization. AI Hypercomputer’s choice and flexibility is the reason ComfyUI selected Google Cloud to access NVIDIA GPUs for shared open-source philosophy.
Our users—from Hollywood visual effects artists to global advertising agencies—demand extreme control over every creative element. By building Comfy Cloud on Google Cloud with NVIDIA GPUs, we abstract away backend hardware complexity so loading massive 100GB models feels seamless and magical.
Robin Huang
Co-founder, ComfyUI
To power its serverless workloads, Comfy Cloud transitioned to Google Cloud G4 virtual machines powered by NVIDIA RTX PRO 6000 Blackwell GPUs. Visual diffusion models require frequent model swapping and rapid memory access rather than massive memory bounds.
The NVIDIA G4 instances provide the optimal balance of compute performance, high VRAM, and fast host system RAM to handle massive visual models cost-effectively.
"Transitioning to Google Cloud G4 instances powered by NVIDIA Blackwell doubled our performance while cutting model load times from an hour to one minute,” says Huang. “The price-performance gain was an instant win for our team and our user community."
Comfy Cloud orchestrates their clusters using Google Kubernetes Engine.
To manage spikes in global traffic, ComfyUI implements a tiered strategy that combines committed use instances with Dynamic Workload Scheduler Flex Start mode and on-demand GPUs to instantly meet peak workday demand.
Asset delivery is another critical piece of the architecture. By colocating model storage in Google Cloud Storage buckets and GPU compute within the same data center, Comfy Cloud allows users to load 100GB model weights in seconds.
"Our users—from Hollywood visual effects artists to global advertising agencies—demand extreme control over every creative element,” says Huang. “By building Comfy Cloud on Google Cloud with NVIDIA GPUs, we abstract away backend hardware complexity so loading massive 100GB models feels seamless and magical."
Additionally, ComfyUI uses Gemini to access foundation models like Nano Banana with enterprise-grade security guarantees.
Doubling performance, eliminating errors, and scaling Enterprise AI
By combining NVIDIA’s latest GPU architecture with Google Cloud’s serverless infrastructure, ComfyUI is delivering significant performance and operational improvements across its platform. The move to NVIDIA Blackwell-powered G4 instances on Google Cloud enables faster, more fluid creative collaboration, helping teams turn ideas into finished work more quickly while improving the platform’s overall price-performance.
When moving from legacy NVIDIA A100 GPUs to Google Cloud G4 instances, Comfy Cloud saw a two-fold performance boost. Model load times dropped from over an hour on local internet setups to approximately one minute through colocated Google Cloud Storage. System reliability also increased, as the vast memory capacity of the G4 instances virtually eliminated out-of-memory crashes throughout complex diffusion pipelines.
"Partnering with Google Cloud and NVIDIA gives us the enterprise-grade security, high-bandwidth storage, and cutting-edge GPU architecture to scale open-source generative AI for the world's most demanding creative studios," says Huang.
Partnering with Google Cloud and NVIDIA gives us the enterprise-grade security, high-bandwidth storage, and cutting-edge GPU architecture to scale open-source generative AI for the world's most demanding creative studios.
Robin Huang
Co-founder, ComfyUI
Looking ahead, ComfyUI is expanding its enterprise footprint on Google Cloud. Having completed SOC 2 Type II, the team is now rolling out advanced data security, admin observability tools, multi-GPU inference, and native agentic workflows to support large-scale enterprise deployments. Through its ongoing partnership with Google Cloud and NVIDIA, ComfyUI remains committed to empowering creators with open-source generative AI tools that scale effortlessly from desktop to cloud.
ComfyUI is an open-source, node-based framework that enables artists, developers, and machine learning enthusiasts to build and customize visual generative AI workflows. Founded in 2023, the platform provides deep granular control over diffusion pipelines, empowering Hollywood visual effects studios, top advertising agencies, and gaming creators worldwide.
Industry: Technology
Location: United States
Products: G4 VMs, Google Kubernetes Engine (GKE), Dynamic Workload Scheduler, Gemini for business, Google Cloud Storage, NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs (NVIDIA Architecture), NVIDIA A100 GPUs
About Google Cloud partner – NVIDIA
NVIDIA is a global leader in accelerated computing and artificial intelligence, pioneering graphics processing units and high-performance hardware architectures that power the world’s most demanding data center, creative, and enterprise workloads.
