Generate a structured robotics concept image using NVIDIA’s publisher-tested GB200 Diffusers route, then inspect its physical details rather than treating it as simulation data.
Publisher-tested GPU routes, not a claim that GB200 is the lowest-cost working configuration.
| Machine | Cosmos3 Super · 65B |
|---|---|
| NVIDIA RTX 5090 · 32 GB VRAM | No verified full-pipeline route for this exact checkpoint; choose a documented host or Qwen Image for ordinary art. |
| NVIDIA GB200 · 192 GB VRAM | Publisher-tested Diffusers example; confirm package versions and capacity. |
| NVIDIA 8× H100 · 640 GB aggregate VRAM | Publisher vLLM-Omni alternative; this guide follows the simpler GB200 Diffusers route. |
| AMD MI355X · 288 GB VRAM | No publisher-tested ROCm Diffusers path here; choose a documented NVIDIA route. |
| Apple Silicon Mac · 192 GB unified memory | No verified MPS path; use Mac as a client to a Linux GPU. |
| RunPod · NVIDIA GB200 · 192 GB VRAM | Publisher-tested GPU class if available; compare live price with multi-H100 before renting. |
| Intel Arc · 32 GB VRAM | No documented Intel GPU pipeline for this checkpoint; use the linked Linux NVIDIA route or a smaller compatible model. |
| CPU only · 128 GB RAM | No practical documented CPU inference route for this full pipeline; use an owned Linux NVIDIA host or a rental. |
No ComfyUI JSON was verified for this exact 65B checkpoint. Do not substitute an unrelated Cosmos-family template.
Use Cosmos for structured physical-AI scene concepts when its heavy GPU path is justified. Choose an owned Linux GB200 or a rental; for ordinary images start with the simpler Qwen Image guide.
This setup uses the package documented for this task. Review its source and supported platforms before starting.
The publisher has not provided a complete memory target for this task. Check its hardware guidance before starting.
NVIDIA tested Diffusers on GB200. Start at 1024×1024 and 50 steps; this is a tested host, not a minimum. Do not extrapolate a 5090 fit.
Choose your computer first, then its GPU path. The instructions below update to match your selection.
Windows: The documented Diffusers route was tested on Linux GB200; Windows is a client to the owned or rented Linux instance, not a validated native host. Compatibility source
Mac: The documented Cosmos3 Diffusers example uses CUDA on GB200; no source-backed Apple Silicon route is offered here. Compatibility source
Use a GB200-class Linux CUDA server. NVIDIA tested the Diffusers path on GB200. For H100 hardware, follow its separate eight-H100 vLLM-Omni recipe instead of assuming the same single-host command works.
NVIDIA documented deploymentIn Linux Terminal, create a Python 3.13 environment, install current Diffusers from source plus the publisher dependency list, and confirm CUDA PyTorch matches the host driver. Keep the inference endpoint private; this guide runs as a local script.
Install Diffusers from the official repository NVIDIA exact dependency list Install uvIn Linux Terminal, download and run the complete source-derived script. Select nvidia/Cosmos3-Super-Text2Image with a structured warehouse robot and cloth prompt. Set one 1024×1024 frame at 50 steps, guidance 4.0, flow shift 3.0 and seed 1143; save the output PNG in the working folder.
Download the complete scene script NVIDIA publisher generation exampleOpen the saved output cosmos3-physical-ai-scene.png and check that the robot, cloth, workbench and camera framing follow the prompt. Inspect hand/object geometry for errors. This is a visual concept, not a physics simulator or a safety-certified robotics scene. For ordinary image art, the Qwen-Image-2.1 guide has a simpler local route.
Simpler Qwen image guide NVIDIA output limitationsOpen cosmos3-physical-ai-scene.png and check the requested robot, blue cloth, workbench and framing. Review object geometry; do not treat the image as physically accurate simulation.
Source-verified, not a YouRunAI render or cost comparison. Keep this specialist option separate from ordinary image recommendations.
Install current Diffusers from the linked Git repository with the publisher dependency list, then check the Python 3.13 environment is active.
Use a documented publisher host, confirm full disk space and CUDA driver support, and do not extrapolate a 32 GB consumer-GPU route.
Treat it as a concept image, refine the structured prompt and inspect geometry manually. Do not use it as ground truth for safety or robotics control.
Original instructions, model files, and compatibility notes behind this setup.
Source-backed package and platform-route plan for private local chat with Qwen3.5 · 9B. Preserves the official Ollama 9B package, a smaller 4B family fallback, and a clearly labeled community uncensored GGUF option.
Run Qwen3.8 27B with Ollama, connect it to Cline in VS Code, and make one reviewable code change in a tiny local project.
Generate images from text prompts in ComfyUI with the fast distilled or more adjustable base 4B model. Save your results and reuse the visual workflow.