Use cases · image and video
Generate images and video on a GPU of your own.
Run open image and video models in ComfyUI on a dedicated machine, billed per minute. Below: the models people use now, the tools, demos, a one-click deploy, and everything we have written on the subject.
Updated 2026-11-21. Model facts come from the articles linked under each list.
Models people use now
Video models
Two open video families are licensed for use in the EU. HunyuanVideo 1.5 and MiniMax H3 are not: their licenses exclude the EU, so they are left out here.
Wan 2.2
T2V and I2V at 14B active (27B total), a 5B TI2V model, and Animate for character motion.
LTX-2.5
A 22B DiT with a Gemma 4 12B text encoder that generates sound with the clip. One owner reported about 70 seconds for a 5-second 768x512 clip on a DGX Spark.
Source: Video generation guide, 7 November 2026, with sizes, licenses and links to each model card.
Models people use now
Image models
The top of Artificial Analysis's open-weights image board on 7 November 2026, plus the strongest models you can use commercially at any company size.
Qwen-Image-2.1
7B DiT, ranked first among open-weight image models (Elo 1045).
FLUX 3 Image Klein
5.6B with a Qwen3-VL-4B encoder; about 4 seconds per 1 MP image on an RTX 5090, per BFL.
Z-Image-Turbo
6B and fast; the pick when you need an Apache 2.0 license at any company size.
Source: Local image models compared, 7 November 2026, with sizes, licenses and links to each model card.
Tools
What runs them.
ComfyUI
The node editor most image and video workflows are shared in. Checkpoints, custom nodes and outputs live in /workspace/comfy and survive a stop.
Diffusers
Hugging Face's library for running the same models from a script or a notebook. Install it with pip in the JupyterLab preset.
NVIDIA's ComfyUI playbook
NVIDIA's own steps for installing ComfyUI on a DGX Spark.
See it running
Demos and walkthroughs.
From the model makers and the ComfyUI team. A video loads from YouTube (youtube-nocookie.com) only when you press play.
Deploy
Start in one click.
A whole machine for yourself, billed per minute from prepaid credit. Terminate when you are done and billing ends.
ComfyUI on a DGX Spark
128 GB of memory, so pipelines that peak above 100 GB fit. Opens in your browser through an SSH tunnel.
JupyterLab on a DGX Spark
For Diffusers scripts and notebooks on the same 128 GB machine.
RTX 5090, Ubuntu 24.04
Faster per clip: reports put the RTX 5090 2.4 to 5.5 times ahead of a Spark on the same video job, within 32 GB. You install ComfyUI yourself.
Guides and articles
Go deeper.
DGX Spark video generation: seconds per clip, November 2026
DGX Spark video generation, November 2026: LTX-2.5 5-second clips in about 70 s, Wan 2.2 5B 720p in 14 min, and why an RTX 5090 runs 2.4 to 5.5x faster.
Read →Best local image generation models 2026: VRAM and licenses
Best local image generation models on 7 November 2026: arena ranks, VRAM for 14 open models, RTX 5090 seconds per image and which licenses allow commercial use.
Read →ComfyUI and Stable Diffusion on a DGX Spark: setup and speeds
Run ComfyUI, FLUX and SDXL on a DGX Spark: setup on ARM64, reported image times per model, unified-memory contention with LLMs, and thermal caveats.
Read →