GuidePublished by EverXCloud Team4 sections2 min read
Save a Reproducible Workflow
Start with an image-generation workflow you can already run in an authorized ComfyUI environment. Export the workflow JSON and record the ComfyUI revision, custom-node revisions, checkpoint, VAE, text encoders and model licenses. For each model file, keep a revision or checksum so a later comparison uses the same weights.
Use the official first-generation walkthrough if you need to learn how to load a workflow and install its required models. This guide assumes the environment and licensed files are already available. It does not establish that an EverXCloud template contains a particular model or custom node.
Hold the Generation Settings Constant
Fix the prompt, negative prompt, dimensions, batch size, sampler, scheduler, steps and guidance settings. Start with one image per batch. For video, also record frames, frame rate and any interpolation or upscaling stage. A change in any of these settings changes the work being compared.
Choose a short, fixed list of different seeds and use that same list on each GPU. Changing seeds between measured runs prevents an unchanged graph from reusing cached output. Record execution-cache settings as well; a cached completion is not a new image-generation benchmark.
Separate Loading from Measured Execution
Run one warm-up separately to load the model. Then queue one generation at a time for each seed and retain the history or logs and output images. Measure the same execution boundary on each machine. Keep checkpoint loading, queue waiting and warm execution separate, and record failures and out-of-memory errors.
Keep CPU offload, precision and attention settings unchanged when comparing like for like. If a configuration requires offload, report it as a different configuration rather than describing the difference as a GPU-only speed comparison. Export an API-format workflow if you automate the process, following the official API example.
Calculate Cost per Successful Image
Use the complete configuration rate and total billable session duration, including preparation and failed runs. Divide total session charges by the number of usable outputs. An hourly rate alone does not establish which GPU is cheapest for your workflow; output quality and successful completion matter too.
Publish the workflow, software versions, settings and timing boundary alongside any measured result you choose to share. Remove secrets and private prompts first. Save outputs before releasing compute and review retained storage. This is a test protocol, not a measured EverXCloud performance claim.
Record the same fields on every GPU
| Record | Keep with your result |
|---|---|
| Environment | GPU and count, VRAM, driver, CPU, RAM, ComfyUI and node revisions |
| Workload | Workflow JSON, model checksums, dimensions, batch, sampler, steps and seed list |
| Execution | Loading time, warm-up, per-run execution time, cache/offload settings and failures |
| Outcome | Usable images, total session charges, currency and cost per successful image |
Sources
- ComfyUI first image-generation walkthrough (checked October 28, 2026)
- ComfyUI API workflow example (checked October 28, 2026)
