Persistence and loss
How long do people, objects, lettering and spatial relationships remain recognizable? What replaces them, and do particular forms recur?
DIFFERENCE ENGINE / FOR RESEARCHERS
Follow the image.
Examine the process.
An artist-led archive of iterative image generation.
Difference Engine records what happens when a generated image becomes the input for the next. It brings together long image lineages, variations in source and instruction, and an ongoing yearlong experiment. We welcome conversations about analysis, interpretation and carefully designed follow-up studies.
QUESTIONS THE ARCHIVE OPENS
What survives? What appears? Who sees a story?
Prompt and starting conditions
How do trajectories differ across longer, neutral and minimal instructions, or different source images? Which comparisons need additional controls?
From sequence to narrative
Where do viewers perceive continuity, mood or a story? How does that reading change with selection, editing, motion and sound?
Generation across a year
How does a daily lineage develop alongside beginning and ending batches? How should variation be interpreted when the service may change?
What is available now?
All sixteen completed branches are public: four 365-generation runs, nine 100-generation experiments and three 100-generation continuations. They are presented in fifteen galleries, including the combined Rainbow Escape viewer. Sunset and Lush continuations keep separate galleries and their original 0–100 numbering. Every gallery includes its starting image. The experiment index describes all 18 recorded branches, including additional finite runs, continuations, the daily lineage and the planned final batch.
Exact prompts, recorded model aliases, output settings and dates are available for the public galleries. The new index provides a downloadable overview and exact prompts for every listed branch.
The daily lineage has 56 accepted generations recorded as of October 10, 2026. Selected watermarked milestones appear on the 365/365 page.
How the records are kept
Each generation uses the immediately preceding image and the run’s fixed instruction. Available records include ordered source images, seed and prompt hashes, batch manifests, timestamps, settings, and logs of retries or repairs.
Batch manifests record parent and output hashes. The daily branch currently has archived images, state and an event log; it does not have the same per-generation manifest as the batches.
This index checks file counts, numbering, seed and prompt hashes, and the continuation starting points. Both Rainbow Escape branches and the eight branches added in V06 have additionally passed checks of every archived image and all their parent/output records: 1,000 generation records across those ten branches. This is an integrity check, not a quantitative analysis of the images.
METHODS & INTERPRETIVE LIMITS
The conditions belong with the results.
A chain is a connected trajectory.
Its images are dependent observations. A 365-generation run is not 365 independent trials. Broader claims require repeated runs and an explicit analysis plan.
The instruction is an experimental condition.
The long prompt actively invites transformation, dream associations and changes in emotional tone. Its outcomes must be read alongside that direction. The shorter-prompt experiments help frame further comparisons.
Several variables can change together.
Some runs differ in source image, lettering, canvas or date. HOME changes both the photograph and its word. START-A and the square Neutral/Minimal runs also differ in canvas. These comparisons do not isolate a single cause.
A model alias does not fix a model version.
The records name gpt-image-2. They do not identify a dated model snapshot or establish whether the underlying service stayed unchanged. A fixed prompt alone cannot guarantee an identical process over time.
Retries and repairs affect the record.
Moderation-blocked attempts are retried from the same parent under the recorded scripts. START-B has a documented repair. These details belong in any analysis of how the accepted lineage arose.
The film is an artistic interpretation.
DEAD END uses selected early images as motion-source material, with human direction, editing and an original score. Its emotional effect should be distinguished from measurements or viewer studies of the underlying still-image sequences.
Related work
Aditi Ramaswamy, Hana Chockler and Melane Navaratnarajah’s Defining and Quantifying Creative Behavior in Popular Image Generators (ICCC 2025) examines iterative image chains through measures of prompt satisfaction, cohesion and novelty.
Their reported setup uses short chains with regenerated captions. Difference Engine includes longer fixed-prompt trajectories. That is a methodological connection to explore, not an established claim of novelty or superiority.
The computational-creativity community brings together researchers and artists working across these questions.
Research access
Public images are watermarked derivatives. Clean originals and working archives remain private. Researchers can enquire about a defined set of runs, their records, and the intended analysis; access and reuse terms will be agreed individually.
No public download of the clean archive, dataset DOI or general reuse licence is offered at this stage. A versioned, citable release can follow once its scope and access conditions are agreed.
AN INVITATION TO COLLABORATE
Bring a question to the archive.
We are interested in collaborations on visual persistence, recurring forms, prompt sensitivity, human readings of narrative, and methods for comparing long image lineages. A useful starting point is the question you would investigate, the runs you would need, and the approach you would use.
Contact David Bryan Stevens ↗Exploratory work in progress · No peer-reviewed findings are claimed here.