AI Glossary / MLOps and deployment / Intermediate

Generation loss

Definition

Generation loss is degradation that can accumulate when media are repeatedly copied through a process that does not preserve all information. In digital workflows, it commonly occurs when an image, audio recording or video is decoded and then recompressed with a lossy codec. A bit-for-bit digital copy does not itself introduce this loss.

What it means

A first compressed export may discard subtle information while remaining acceptable for its intended use. If that export becomes the input to another lossy export, the next stage works from an already altered version. It can retain existing artifacts and add new changes. Repeated editing, resizing, color conversion and changing codec settings can complicate the process further. The amount of degradation is not a fixed percentage per generation and need not increase equally on every save.

How it works

Lossy encoders approximate or quantize information to reduce the required data. Decoding reconstructs a usable signal from that reduced representation, but does not recover everything discarded. A later encoder sees this reconstruction rather than the original. If its transform, quantization, alignment or format differs, it can introduce additional error. Some repeated operations may stabilize under particular conditions, so avoid claiming that every re-encode necessarily causes the same visible change. A lossless intermediate preserves its input exactly but cannot undo earlier loss.

Three practical examples

• A marketing team edits a previously exported JPEG and saves another JPEG, gradually affecting fine text and edges across revisions.
• A video moves through several production systems that each transcode it to a delivery format, potentially compounding artifacts.
• An AI team gathers images from several republishing sources and discovers that compression history varies, creating differences unrelated to the labels it wants the model to learn.

Think of it this way

Think of making a photocopy of a photocopy: each imperfect copying stage works from the previous result. The analogy explains accumulation, but ordinary digital file duplication is different. Copying the exact bytes of a JPEG preserves that JPEG exactly; a new generation of coding loss requires a process such as lossy recompression or another information-reducing transformation.

Worked example

A toy scalar quantization example separates the stages. Start with value 23. Quantizing to the nearest multiple of 10 gives 20, an error of −3. If a later stage quantizes that result to the nearest multiple of 6, it becomes 18, an overall error of −5 relative to the original. In contrast, applying the same multiple-of-10 quantizer again to 20 leaves it at 20. This simplified example does not model all JPEG or video behavior, but it shows why different quantizers can compound loss and why an identical repeated operation need not add the same amount every time.

Illustrative case study

Fictional case: Laurel Media receives several versions of a product photograph and selects the smallest file for a new campaign. After adding labels and exporting again, fine surface detail looks worse than in an earlier version. The team locates the original master, performs the edits there and produces the required delivery copies directly from that master. It records export settings and avoids using a delivery copy as the next editing source. This fictional workflow illustrates prevention rather than a promised restoration of missing information.

Common misconceptions and mistakes

Do not assume that opening or copying a file automatically degrades it. Do not expect a larger file or a higher quality setting to restore details already discarded. Saving a damaged image in a lossless format preserves its current state, including existing artifacts. Some tools perform certain transformations without recompressing the underlying data, but that capability must be checked for the specific operation and format. Visual acceptability also depends on content: fine text may reveal damage that a smooth photograph hides.

Business use

Businesses can reduce generation loss by preserving a high-quality master, keeping suitable working files and creating distribution versions from the best available source. Plan the complete handoff chain so each system's export settings are understood. For AI data, record source history when available and inspect whether compression correlates with labels or data splits. Quality checks should include the features important to the task, not only overall appearance or file size.

People also ask

Does copying a JPEG to another folder reduce quality?

No, if the bytes are copied unchanged.

Can conversion to PNG restore a JPEG's lost detail?

No. It can avoid another lossy encoding stage for the current pixels, but it cannot recover discarded information.

Is generation loss always visible immediately?

No. Its impact depends on the content, operations, settings and viewing or analytical task.

Check your understanding

A compressed video is transferred unchanged through three storage services. Has it necessarily undergone three lossy generations?

Answer: No. If the services preserve the same bytes, no new codec generation is introduced. Transcoding or another information-reducing processing step would change that assessment.

Related terms

• Lossy compression
• Compression artifact
• transcoding
• quantization
• master file
• lossless intermediate
• recompression

Learn next

Compare an untouched digital copy with an image that has been decoded and re-exported at several settings. Use the original as a reference and inspect fine edges and text. Document the operations between versions so the cause of any difference is clear rather than inferred from filenames.

More in MLOps and deployment

• Fibonacci coding - Fibonacci coding is a self-delimiting representation of positive integers based on Fibonacci numbers.
• FM-index - An FM-index is a compressed full-text index developed by Paolo Ferragina and Giovanni Manzini.
• Fractal compression - Fractal compression is a family of compression methods that represents data, especially images, through transformations whose repeated application...
• Goal monitoring - Goal monitoring is the process of checking whether the goals an agent is currently pursuing remain the most appropriate goals to pursue.
• Golomb coding - Golomb coding is a family of lossless variable-length codes for nonnegative integers.
• Gradient Checkpointing - Gradient checkpointing, also commonly called activation checkpointing, is a training technique that reduces stored intermediate activations by recomputing...

Sources

Primary product documentation: Adobe Media Encoder compression guidance, https://helpx.adobe.com/media-encoder/desktop/encoding-quick-start-and-basics/compression-tips.html ; Adobe image-quality workflow guidance, https://experienceleague.adobe.com/en/docs/experience-manager-cloud-service/content/assets/dynamicmedia/best-practices-for-optimizing-the-quality-of-your-images . Official documentation search excerpts consulted; full workflow-specific and editorial review remain pending.

Previous: Generated text Next: Generative adversarial imitation learning