Seeds and Reproducibility: Making Workflows Deterministic

Generative AI models sample from a probability distribution over possible outputs. A seed initializes the random state that controls this sampling, making the sampling process repeatable. Given the same seed, the same model, the same prompt, and the same parameter values, the model will produce the same — or nearly the same — output. This is the definition of determinism in the context of generative models.

In a single-node workflow, seed management is straightforward: you set the seed on the node, save it with your output file, and use it to regenerate if needed. In a multi-step workflow, each node that performs generation has its own random state and therefore its own seed. If any intermediate node runs with a random (unspecified) seed, its output varies on each run, which cascades through all downstream nodes and makes the final output non-reproducible even if the terminal nodes have fixed seeds. Full workflow reproducibility requires every generation node to carry an explicit, recorded seed.

In the Floniks editor, every generation node exposes a Seed field in its configuration panel. The default value is typically zero or empty, which is interpreted as "use a random seed." To make a node deterministic, enter a specific integer in the Seed field. Any integer in the range the model supports will work — a common practice is to use a timestamp-derived integer (like 20260619) or a sequential identifier (1001, 1002) to make seeds human-readable in output logs.

When you first generate an output you want to reproduce, note the seed value reported in the node's execution log. If you used a random seed, the executed seed is recorded in the task record alongside the output. Copy that executed seed back into the node's Seed field before saving the workflow template, locking it for future runs. Make this a post-generation habit: inspect the task record, extract all executed seeds, and update the workflow template with the locked seeds before archiving it.

In a chained workflow where the output image of node A becomes the input to node B (for example, an image generation node feeding an upscaling and refinement node), the seed of node B controls the variation introduced by the refinement step. If node A produces image X with seed 42, and node B refines image X with seed 99, the final output is determined by both seeds together. To reproduce the final output, you must record and re-apply both seeds.

The Floniks task record captures the executed seed for every node in the execution graph, associated with the task ID. To reproduce a previous output, retrieve the task record, extract the per-node seed list, apply them to the corresponding nodes in the workflow, and re-run. This workflow-level seed retrieval is more reliable than trying to remember or reconstruct seeds from file metadata, which is why recording seeds in the task record (not just in file names) is critical for reproducibility workflows.

Not all AI models guarantee bit-exact reproducibility even with the same seed. Three conditions can break determinism. The first is model updates: if the model weights change between runs — even a minor update to a hosted model — the same seed will produce a different output. Always record the model version alongside the seed, not just the seed alone. The second is hardware variation: some inference backends produce slightly different floating-point results on different hardware configurations, particularly when run on different GPU types or with different batch sizes. The outputs will be visually similar but not pixel-identical. The third is non-deterministic post-processing: if a node in the pipeline applies a random-seed-driven upscaler or a stochastic color grading step, that step's randomness is independent of the generation seed and must be seeded separately.

For workflows where exact bit-level reproducibility is required — product photography archive, legal evidence, scientific visualization — test and document the reproducibility envelope of each model in your pipeline before relying on it in production.

Seed locking is especially important in A/B testing and localization variant workflows. In these contexts, the goal is to hold the underlying generation constant while varying one parameter. If any generation node uses a random seed, the generation itself varies between branches, contaminating the test. All generation nodes that are shared (upstream of the branch point) and all generation nodes within branches that should produce structurally identical compositions must carry the same explicit seed value. The only nodes that should differ between branch paths are those intentionally applying the test variable.

Document the seed choices for each variant batch in the task record and the campaign brief. When test results come back from the ad platform, you will need to re-run the winning variant at high volume — seed documentation is what makes high-volume reproduction of the winner possible without quality drift.

The Floniks task record stores all per-node executed seeds automatically. Retrieve them via the task history view or the task API using the task ID. For long-running projects where you need to reproduce outputs months later, export the seed record as a structured document alongside your output files — a simple JSON file mapping node ID to seed value and model version is sufficient. Store this alongside your generated asset files in your content archive. When a client requests a re-run of a specific campaign image six months later, the seed record combined with the workflow template is everything you need to regenerate an identical (or near-identical) output from the same model version.