Z-Image-Turbo: understand the local workflow before evaluating speed
Z-Image-Turbo is a specific model variant, not a promise that every website using its name offers unlimited free generation. Identify the weights and execution route first, then compare the actual local workflow's output quality and completion time under recorded conditions.
Separate the model from the service
The official model card identifies Z-Image-Turbo and provides its licence and implementation examples. A hosted interface adds another layer of limits, billing and data handling. Check both when applicable rather than assuming that model availability determines the terms of a website.
Record the exact checkpoint and variant. Turbo and other members of the family can require different settings and support different tasks. Do not transfer a negative-prompt assumption, guidance value or editing capability from one variant without checking its maintained example.
Before sharing a result, open the downloaded file outside the generating interface and confirm its dimensions and format. Retain the original output separately from edited versions so an attractive post-processed image is not attributed entirely to the model.
Load a complete reference workflow
Use the official ComfyUI Z-Image-Turbo workflow documentation to identify the model components and their intended connections. Confirm that the installed ComfyUI version supports the nodes used by the example. A missing node should prompt a compatibility check, not an invented replacement based on a similar name.
Keep the first working graph as a baseline. Record model filenames, software versions, output dimensions and the initial prompt. Save the actual output and graph together before adding adapters or optional processing stages that could change the interpretation of the result.
Measure the experience you care about
Separate cold startup, first complete image and subsequent generation times. State whether a preview, download or final saved file marks completion. A publisher's demonstration on a different accelerator is not a benchmark of your workstation or a hosted site's queue.
Test a consistent set of original prompts and review object count, composition, text and visual artefacts. Keep an attempt budget and include failures. A fast incorrect image is not equivalent to a slower accepted deliverable, so report usefulness as well as elapsed time.
Keep the conclusion conditional
Repeat important comparisons in the same environment and retain the inputs. Fixed seeds help organise tests, but reproducibility has software and hardware limits. Check the documented behaviour rather than claiming that one seed is a universal quality control across models.
This revision withdraws the earlier unlimited and universal-best framing. It describes how to evaluate the model in a known workflow and how to distinguish upstream claims from locally reproduced observations. When the model, integration or service changes, rerun the relevant cases instead of carrying an old performance conclusion forward indefinitely.