Look for the image. Keep the decisions that make it right.

Turn intent into visible criteria

Start with the destination of the image: poster, product page, illustration or mood exploration. Define the subject, the framing, the palette, the light and what must remain free for a layout. Add the errors that make a proposal unusable: truncated object, illegible silhouette, reflection hiding a detail or an overloaded text area.

For a poster, separate the background exploration from the final typography. A beautiful image can be impossible to use if the title covers its focal point. Prepare a layout template and judge the candidates with it. Keep a starting reference that you have the rights to, then describe what you take from it: a palette or a rhythm of masses does not require copying the visual.

Gather the graph, the models and the inputs

The ComfyUI workflow links operations and their parameters. Save the JSON separately from the images, with the exact reference of the models and the versions of the extensions. The graph file does not contain the weights. If the project uses an input image, keep its original file and the preparation applied before loading: crop, mask or dimensions.

The Céramique folder provides an original workflow with core nodes, three lighting intents and a manifest. Its structure has been checked; no image was generated for the pack. Use its files to understand how to document a test, then verify your own run. The Environments page helps separate present dependencies from a treatment that actually succeeded.

Vary one question, then verify that it matters

On the first pass, choose one axis: lighting, framing, or material. Keep the model, dimensions, and settings comparable, then change the wording that carries that intent. This doesn't force the generator to change only that property; it makes your test more legible. Also inspect what changed unintentionally, for example the product's shape along with its lighting.

Save the seed, settings, and graph version for each result. A seed can be used to compare directions within the same combination, then multiple seeds let you examine whether the direction holds beyond a lucky accident. Don't promise an identical reproduction after changing software or platform: PyTorch documents this limitation. What you're after here is a reasoned creative decision, not statistical proof of superiority.

Hypothetical example: three lights for a poster

Let's imagine a bottle placed on the right half of a poster, with a title on the left. The common intent is a matte ceramic and a calm background. Three directions are prepared: soft frontal light, more contrasty side light and backlight. The rejection criteria are set before looking at the results: clipped silhouette, material turned metallic or contrast too strong in the title area.

A first comparison with the same seed produces three expected candidates. The team then plans three more seeds for each of the directions: nine additional candidates, so twelve in total. This calculation describes the hypothetical work plan; it reports no generation actually run. The number of outputs says nothing about their acceptance rate or GPU time.

The grid below therefore contains no invented score. If two directions succeed, the next step can test a horizontal and a vertical framing. This new need is named as such, because switching to another format can change the composition; counting two exported files is not enough.

Choosing memory for the full workflow

The requirement depends on the model, the dimensions, the number of images processed together and the steps added. A successful generation may be followed by a decoding or upscaling step that is more demanding. Note the step where memory runs short before changing cards. A model file size is not a complete measure of the requirement at runtime.

The RTX 4090, 24 GB, RTX 5090, 32 GB and RTX 6000 Ada, 48 GB pages give the nominal capacity per card. Compare them with the verified requirements of your graph; they do not establish a speed ranking. Start with a controlled batch, then check the final chain before committing to the series. A requested setup does not attest that the software is installed.

Selecting a direction before the finishing touches

First look at the candidates together, at the same size and under the same color interpretation. A contact sheet lets you compare balance, palette and consistency; a detail that looks appealing full screen must not hide a disjointed series. Keep each image's identifier visible and link it to its workflow. The guide Creating a contact sheet covers this step in detail.

Distinguish local corrections from a change of direction. A small anomaly near the edge may justify a retouch; an incorrect subject calls for another proposal. Select before upscaling all the files. For the images you keep, the guide Upscaling an image helps compare the result at its intended size: a detail added by a generative process is not automatically faithful to the expected object.

Recognizing the errors that make a series drift

Unlogged settings make a variation impossible to understand; the same file name can hide a replaced model. An update in the middle of a selection can change the results. Keep a stable version of the approved recipe and create a named copy for the next step, instead of overwriting the only copy.

Other drift shows up in production: too many directions to judge, no rejection criteria, a series decided from a single image, or retouches that destroy its coherence. Write down why a result was rejected. That reason may lead you to fix the brief rather than relaunch endlessly. If the exact form of a product is essential, decide which elements will come from a controlled source and which will remain exploratory.

Finish with a readable selection and archive

Name the selected results and state their purpose: validation, layout or final delivery. Open them in the receiving application to check dimensions, color and cropping. Keep the original files, the JSON, the authorized inputs and the model references. Metadata embedded in some media is useful, but a separate JSON is easier to find after a conversion.

The workshop deliverable is a justified selection, tied to its recipes and criteria. The export guide organizes its reception. To choose 3, 7 or 30 days, also count the revisions, the selection and the retrieval; a plan does not promise a number of images. You keep the creative decision and choose the software and processing used.