Use one rented NVIDIA GPU and ComfyUI’s official MiniMax H3 text-to-video graph to make and download a short clip with synchronized audio.
Rented Linux GPU recommendations for the complete 42.5 GB optimized H3 text-to-video graph at 0.2 MP and shortest duration. Your own Windows, Mac or Linux computer only controls the browser.
| Machine | MiniMax H3 · pruned INT8 + 8-step LoRA |
|---|---|
| Rented NVIDIA H100 · 80 GB VRAM | Estimated short-clip fit; verify the full graph and watch peak use. |
| Rented NVIDIA H200 · 141 GB VRAM | Recommended rental for generous first-render memory headroom. |
| Rented NVIDIA B200 · 180 GB VRAM | Ample capacity if available; compare the current rental price. |
Capacity estimates, not timed benchmarks. Confirm that the selected Pod template supports the current H3 INT8 ConvRot package and driver.
Rent one compatible Runpod GPU, open ComfyUI privately from any desktop browser, load the five official files and text-to-video template, render one short clip with sound, download it and stop billing.
This setup uses the package documented for this task. Review its source and supported platforms before starting.
Save your machine in My Hardware to get an automatic starting choice. You can always choose any package yourself.
Choose an 80 GB H100 or preferably 141 GB H200 for a short 0.2 MP first render; H3 optimized files occupy roughly 42.5 GB before frame activations. This is a conservative estimate, not a measured peak. Local browser hardware does not affect model fit.
Choose where you want to run it. Follow the visual setup and import the linked workflow.
ComfyUI runs on the rented Linux NVIDIA Pod. From Windows use PowerShell; from Mac/Linux use Terminal for SSH and any browser for the canvas.
In Runpod choose an on-demand NVIDIA H200 with 141 GB VRAM and at least 64 GB system RAM. An H100 80 GB is a smaller estimated starting point for the short graph; avoid reserving a small GPU based only on the checkpoint size. Set 50 GB container disk and at least 100 GB persistent volume at /workspace. Check the current per-hour rate and storage charges before deploying.
Runpod · affiliate Runpod pricing and storageSelect Runpod’s official ComfyUI template, not a bare Python image. Choose one H200 or H100, On-Demand and a direct-SSH public IP. In port settings expose TCP 22 only; do not publicly expose the stock ComfyUI HTTP port because it has no login. Deploy, open Pod Logs and wait until ComfyUI is ready. The template provides a Linux Web Terminal and /workspace/runpod-slim/ComfyUI.
Official Runpod ComfyUI template ComfyUI official installer and updates Template source and folder layout Runpod port safetyOn your own computer, create or reuse an SSH key and add its public half to Runpod. In the Pod Connect panel copy the direct SSH command containing the current IP, SSH port and your key path. Append -N -L 127.0.0.1:8188:127.0.0.1:8188 and run it in PowerShell on Windows or Terminal on Mac/Linux. Keep the tunnel open, then browse http://127.0.0.1:8188 locally. If local 8188 is occupied, change only the first 8188 to 8190 and browse 8190.
Runpod SSH key and direct connection Windows OpenSSH clientIn the Runpod Web Terminal, download all five source-linked files to the exact folders below. The diffusion model alone is not enough. Let every transfer finish, then restart ComfyUI or refresh the model list. These files are stored on the rented /workspace volume, not your laptop.
diffusion_models/cd /workspace/runpod-slim/ComfyUI
mkdir -p models/diffusion_models
mkdir -p models/text_encoders
mkdir -p models/vae
mkdir -p models/loras
curl -fL --retry 3 'https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors' -o 'models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors'
curl -fL --retry 3 'https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors' -o 'models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors'
curl -fL --retry 3 'https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_int8_convrot.safetensors' -o 'models/vae/minimax_h3_video_vae_int8_convrot.safetensors'
curl -fL --retry 3 'https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors' -o 'models/vae/minimax_h3_audio_vae_fp32.safetensors'
curl -fL --retry 3 'https://huggingface.co/lightx2v/Minimax-h3-Turbo/resolve/main/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors' -o 'models/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors'
test -s 'models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors'
test -s 'models/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors'
test -s 'models/vae/minimax_h3_video_vae_int8_convrot.safetensors'
test -s 'models/vae/minimax_h3_audio_vae_fp32.safetensors'
test -s 'models/loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors'In the private ComfyUI browser canvas, choose Template Library → Video → MiniMax H3 Text to Video. Alternatively download the official JSON and drag it into the canvas. Select the five files installed above. Update ComfyUI if a core node is red; do not replace the graph with an unrelated video template.
Download the official H3 text-to-video workflow Read the official node walkthroughIn the official H3 text-to-video graph, select the pruned INT8 diffusion model and eight-step LoRA. For the first render choose 9:16, 0.2 megapixels, the shortest available duration, turbo_mode enabled and turbo_steps 8 (8 steps). Enter one subject, one action and one sound. These settings keep the first GPU rental small enough to check before a long render.
Click Queue Prompt once. Wait for the graph to finish, then play the result in ComfyUI and listen for the synchronized audio. If it fails, reduce size or duration before renting more time. This guide does not promise a particular render speed on an untested rental.
Download the finished clip to your own computer and open the local file to confirm it plays with sound. Then go back to Runpod Pods and click Stop; closing your browser or SSH tunnel does not stop compute billing. A stopped Pod can continue to incur storage charges. If you will not return, terminate the Pod only after saving the clip and any workflow JSON locally.
Stop or terminate a Pod Persistent storage and chargesComfyUI completes the short video-with-audio graph on the rented GPU. Download the result to your computer, play it to verify motion and synchronized sound, then stop or terminate the Pod.
The GPU rental executes Linux even when your browser runs on Windows or Mac. The 42.5 GB package footprint and 80/141 GB suggestions are capacity estimates, not a published H3 speed test. Review live Runpod rates and storage before deploying.
Check the Pod log and open a private SSH tunnel to 127.0.0.1:8188. Do not expose the stock unauthenticated ComfyUI port publicly.
Check the five exact files and their folders under /workspace/runpod-slim/ComfyUI/models, refresh loaders and restart ComfyUI after the download completes.
Use the official H3 text-to-video graph, 0.2 MP and shortest duration with turbo_mode 8. Close other Pod GPU tasks or choose a larger GPU; do not assume weight size equals peak memory.
Download and open the video on your computer, then stop or terminate the Runpod Pod. Closing the browser or tunnel does not stop GPU billing.
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.