ComfyUI and Stable Diffusion on a DGX Spark

ComfyUI runs on a DGX Spark through NVIDIA's own playbook, and one NVIDIA forum user reports 4.4 seconds for a 1024x1024 FLUX.2-klein-9B image and 11.3 seconds for SDXL 1.0 at 30 steps on a GB10 machine. Those runs used the Hugging Face Diffusers library with torch.compile, not ComfyUI, and GPUwerk has not measured image generation itself. I run GPUwerk, which rents DGX Sparks, so weigh everything below with that in mind.

Reported times, with their limits

The only per-model timings I found come from one forum thread, Image diffusion speeds, posted between May and July 2026. They are one person's measurements on one GB10 machine (the poster refers to it as a GX10, so possibly an ASUS Ascent GX10 rather than NVIDIA's own Spark). Asked about ComfyUI, the poster said they use only the Diffusers library. Treat the numbers as an order of magnitude, not as what a ComfyUI workflow will give you with its own attention and precision settings. The full table is further down.

• FLUX.2-klein-9B: 4.4 s at 4 steps (3.3 s with NVFP4).
• SDXL 1.0: 11.3 s at 30 steps.
• Qwen-Image-2512: 61.0 s at 50 steps.

In the first post, each time is the average of the last five runs, and the poster discarded the first two, which took about 30 seconds each while torch's lazy compile finished. Another user in the thread notes that ComfyUI's default sampling node cannot split one image across several Sparks, so a second Spark adds capacity for parallel jobs, not speed for a single image.

What fits

A Spark has 128 GB of unified memory, shared by the CPU and GPU. That is four times the 32 GB on an RTX 5090, so FLUX, SDXL and Qwen-Image checkpoints should load with room to spare, and room remains for an LLM next to them if you cap its memory (see the section on running both). This is an inference from the memory size, not something GPUwerk has measured. Check each model's published weight size against the 128 GB before you plan around it. The Spark's 273 GB/s of memory bandwidth is well below a discrete GPU's, so do not assume the reported times carry over to other models.

Setup on a rented Spark

GPUwerk has not yet run ComfyUI inside a renter container on a fleet node, so I am not publishing install commands for it. Until I have confirmed that the install works on ARM64 with the preinstalled CUDA stack, follow NVIDIA's own playbook: Generate Images and Videos with ComfyUI (build.nvidia.com, checked October 11, 2026). It is written for the DGX Spark, covers image and video workflows (Stable Diffusion, FLUX, Wan 2.1, HunyuanVideo, Cosmos), starts with Z-Image-Turbo, and serves the web UI on port 8188. NVIDIA estimates two hours, longer on the first run while large models download. It does not discuss ARM64 compatibility explicitly, since the hardware it targets is already ARM64.

You get root in your own container with Docker, CUDA and the NVIDIA stack preinstalled, and 1 TB of NVMe. New to renting? How to rent a Spark covers the first steps. Image models are large downloads, so budget that time against the per-minute bill.

Reaching the UI

ComfyUI listens on port 8188 on the machine, bound to localhost by default. Reach it the way GPUwerk documents for any web UI on an instance: an SSH local port forward, which needs your SSH key set up (see reach a web UI over the tunnel). With the spark host alias from that guide:

# -N means "no shell, just the tunnel"
ssh -N -L 8188:localhost:8188 spark

Then open http://localhost:8188 on your laptop. Keep ComfyUI on localhost: starting it with --listen 0.0.0.0 would expose an unauthenticated UI. The optional desktop in Connect to your Spark's VNC desktop also works if you prefer a browser on the machine itself.

Reported speeds by model

Checked October 11, 2026. All rows are from one forum user (ijontichy), 1024x1024, Diffusers with torch.compile, with the first post's method of averaging five runs after two warm-up runs. According to the thread, the Krea2-Turbo time fell from 39.3 s to 13.9 s after setting a float32 matmul precision and a cuDNN attention backend through an environment variable.

Running ComfyUI next to an LLM

CPU, GPU, the OS, the container runtime, model weights and the KV cache all draw on the same 128 GB. A vLLM server started with a high --gpu-memory-utilization reserves most of it for its cache, and ComfyUI then has little left. In the forum thread Running a large LLM on DGX Spark Cluster while keeping ComfyUI usable? (July 13, 2026), a reply from AakankshaS says separate Sparks for the LLM and ComfyUI is the practical approach. If you co-locate anyway, the reply suggests 0.7 instead of 0.9, or 0.6 for 40 percent headroom, plus disabling swap and using quantised models. Users in the same thread report mixed results: one ran both on a node at 0.8 and called the performance slow, and another found image generation five to six times slower while a llama.cpp model was under load.

For scale: GPUwerk's own measured KV-cache figures in the vLLM guide were taken at 0.85 on an otherwise empty GB10, which suits a dedicated LLM node and leaves little for an image model. The flag is a fraction of total memory, not of free memory, so if ComfyUI already holds memory, vLLM at a high fraction can refuse to start rather than shrink its cache. My suggestion, untested with ComfyUI on GPUwerk: start vLLM first, at the lower fraction. The same guide treats 0.7 as a safer ceiling than 0.85 to 0.9 when anything else shares the box.

Thermal and long-job caveats

These are reports from other people's machines, not GPUwerk findings. I cannot say whether GPUwerk's nodes are affected or immune.

• Part 48 of the ai-muninn DGX Spark series (September 14, 2026, one author, one ASUS Ascent GX10) describes two hard power-offs during sustained MiniMax-H3 video generation. A 2200 MHz clock cap stopped them in a deliberate reproduction, at a cost of about 9 percent throughput on sustained jobs, and the author has since removed the cap on that machine.
• A petronellatech.com write-up (September 26, 2026, eight GB10 units) reports one unit on older firmware going dark during a prefill of about 131,000 tokens, with no shutdown sequence or thermal message in the journal. The authors say they do not know the root cause.
• Part 47 of the same series took all its Spark video timings with the GPU capped at 2200 MHz because of those power-offs, so its absolute times are slower than an uncapped machine would give. Expect a long video job to be slower if your machine needs the same cap.

If you plan a long image batch or a video job, copy outputs off the machine as you go. Terminating an instance deletes the workspace.

Cost

A Spark is $0.79 per hour, billed by the minute, with no egress fee. Compile time and model downloads count as running time. For the text-and-image side of multimodal work, see vision and multimodal models on a Spark.

FAQ