NovaPodWorkload Playbook
ComfyUI cloud GPU for image and video
Use this flow for a ComfyUI cloud GPU paid with USDT or USDC. Start from the template, validate one workflow, then scale the batch and stop the pod when it is done.
Before you launch
- - Workflow JSON and model checkpoints are version-aligned.
- - Disk budget covers model downloads and intermediate outputs.
- - Output resolution and batch settings are capped for the first run.
- - Failure handling is defined for node or sampler errors.
Execution steps
STEP 01
Deploy ComfyUI template
Use a prebuilt template with CUDA, common nodes, and web UI route already configured.
STEP 02
Load one reference workflow
Run a single prompt with expected node graph to verify model paths, samplers, and outputs.
STEP 03
Tune quality versus speed
Adjust sampler steps, resolution, and batch size until turnaround time meets your production target.
STEP 04
Scale generation volume
Move high-volume jobs to better ask/median combinations and keep queue drains observable.
Boot ComfyUI from a template, then one workflow
ComfyUI on NovaPod is a template launch, not a from-scratch CUDA install. Open Templates, start the ComfyUI image, and wait until the web UI route in the template notes responds. Load one workflow you already know, with one checkpoint that is actually on disk. The first render should be a single image at a modest resolution. If a node errors, it is usually a missing model path or a custom node that the template does not include.
Disk fills quickly once you add video checkpoints and intermediate frames. Budget the volume before the download, not after the pod reports no space left. The marketplace ask is the runtime price. Storage does not make a stopped pod free of the download you already did, so keep the models you will reuse and delete the rest before you stop.
Tune one setting at a time
Change sampler steps, resolution, or batch size one at a time and time a representative prompt. A jump from 1024 to a much larger frame can multiply the render. A 24 GB consumer card is a reasonable place to debug the graph. Move the same workflow to a larger card only when the graph is right and you need throughput. Compare that card's ask with its median on the price page before a long queue.
Keep the workflow JSON next to the checkpoint names. When you relaunch, the template gives you the app, not your graph. If the UI port is closed, check that the pod is still running and that you are using the host from the rental, not a bookmark from yesterday.
Scale the queue, then stop the idle pod
After one image looks right, raise the batch. Watch the GPU memory and the time per image. A cheaper ask that crashes every tenth frame costs more than a stable ask. Write down the GPU name, resolution, and steps that worked. That line is what you reuse, and what you send if a later boot cannot find the model file.
Export the outputs, then stop the pod. Per-second billing continues while ComfyUI sits open with no jobs. Relaunch from Templates when the next batch is ready. Use the GPU list to see which cards have the VRAM your checkpoint needs, and the RTX 4090 price page when that is the card you actually tested.
Pre-warmed Container Cache & 1s Crypto Settlement
Base CUDA runtimes and foundational weights are distributed across verified high-bandwidth nodes. Boot instances in under 15 seconds with zero KYC and pay with USDT, USDC, SOL, or ETH.
Budget and reliability tips
- - Large resolution jumps can multiply runtime; profile representative jobs before full batches.
- - Use lower-cost cards for queue warmup and premium cards for final renders.
- - Archive outputs then stop idle instances; storage and runtime should be managed separately.