A green queue status tells you the graph finished. Your editor still needs a file that opens, fits the shot, and leaves room for the subject. If you're using ComfyUI to make background plates for a short drama, test that handoff with one frame before asking for four.

A historical ComfyUI issue report, now closed, describes a different workflow that completed while producing wrong or black images. The two runs below are a way to check your own files. The workflow structure was checked offline. The settings below are a test plan, with GPU output and timings still to be measured.

Build one repeatable run

Start from the official ComfyUI text-to-image example. That introductory example uses SD1.5. For the test plan here, replace its checkpoint with SDXL Base 1.0, connect its CLIP output to both prompt encoders, pass an empty latent through KSampler, and decode with the checkpoint's VAE. Connect the decoded IMAGE output to Save Image. Record the ComfyUI version, checkpoint filename and SHA-256, VAE choice, GPU, and any offload settings alongside the exported workflow JSON.

Setting Example value
Canvas 1024 × 1024
Seed 42117, fixed
Sampler and scheduler Euler / normal
Steps, CFG, denoise 20, 7, 1
First run Batch size 1, filename prefix plate-single
Second run Batch size 4, filename prefix plate-batch

Try an empty hallway with soft light from the upper left and a clear center for a character. In the negative prompt, ask to exclude people, lettering, logos, and watermarks. Keep the prompts and other generation settings fixed across runs. Use a fresh output folder or a unique prefix per attempt so the file counts come from the current job.

Inspect the file your editor will receive

  1. Queue the single-image run. Open its new PNG from disk in an image viewer, check that it decodes at 1024 × 1024, and inspect it for an unexpectedly black frame or other visible corruption. If it fails, keep the job log, workflow and failing file. Record any settings change in a new attempt.
  2. Put a rough character cutout over the plate in your editor. Check that the framing and lighting work for the shot and that the intended subject area stays clear. If the delivery is vertical, check the actual 9:16 crop in the editor. The square test image is the source plate.
  3. Once the first plate passes, try the four-image batch. Open every new PNG and apply the same checks. If the batch fails on memory, record the failure and try four single-image runs as a separate case. Record the elapsed time for the entire four-image job so you can compare the two approaches.
  4. Copy the accepted PNGs and workflow to storage outside the render machine. Open the copies in the destination environment and count the files that pass there. Keep that count with the run settings so the next person can tell which files cleared the handoff.
Run Expected files New PNGs saved Passed after export Queue-to-save time Failure or framing note
Single 1
Batch 4

Copy the two rows into your project notes and fill in the single-image row first. If you're handing off a short-drama plate, include one screenshot showing where the character will sit in the final crop. If you're debugging the graph, save the workflow JSON and the failing PNG together before changing settings.

Prepared by the CloudGPU team with AI assistance. This is an unrun test plan, not a GPU benchmark.