# Difference Engine — research brief David Bryan Stevens · After-Image Engine · October 10, 2026 Contact: info@afterimageengine.com ## Scope An artist-led exploratory archive of iterative image generation. Each accepted image becomes the next request's visual input, accompanied by a fixed prompt within that run. The archive connects long lineages, source and prompt variations, continuations, and a daily experiment across a year. DEAD END is a resulting film; its selection, motion, editing and score are distinct from the underlying image-generation record. The index describes 18 branches: four completed 365-generation batches; nine completed 100-generation experiments; three 100-generation continuations; a daily lineage with 56 accepted generations as of October 10, 2026; and a planned final 365-generation batch. All sixteen completed branches are public in fifteen galleries, including Rainbow Escape generations 0–200 in one viewer. Sunset and Lush continuations retain separate 0–100 galleries linked to their parent endpoints. The Rainbow Escape continuation completed on August 27; the saved completion flag was reconciled on September 19 after verification, with the original state retained. ## Research questions - How long do source elements persist, and what forms recur or replace them? - How do trajectories vary with starting images and degrees of written direction? - Where do viewers perceive narrative continuity or emotional change? - How do batch and daily lineages develop across calendar time? ## Methods and available records The recorded model alias is gpt-image-2; no dated snapshot is recorded. Output canvases vary by run and quality is recorded as medium. The index includes exact prompts, source and prompt hashes, output settings, generation counts and dates. Batch manifests record parent/output hashes and timestamps. The daily branch has images, state and an event log, without an equivalent per-generation manifest. The September 19 index checks archive counts and numbering, seed and prompt hashes, and continuation seeds against parent endpoints. The V05 and V06 revisions additionally verify all archive images and 1,000 parent/output records across both Rainbow Escape branches and the eight V06 branches. No fresh full-lineage audit is claimed for the six earlier public branches. The raw records preserve failures, retries and documented repairs where available. ## The yearlong comparison START-A completed 365 generations on August 16, 2026. A separate lineage is intended to advance once a day across the year. A final batch is configured to begin no earlier than August 15, 2027, from the same original seed and fixed prompt; it has not started. The moving model alias means differences can reflect random variation, accumulated image changes and service changes. The design alone cannot attribute differences to model updates. ## Limits and prospective analysis Images within a chain are dependent observations, not independent trials. The long prompt deliberately invites transformation and dream associations. Some comparisons change several conditions, including source and canvas. Replicated runs and a defined analysis plan would be needed for broader causal claims. Visual themes are interpretations, not evidence of model intent. No peer-reviewed findings or established novelty are claimed. Potential collaborations include annotation of object persistence, repeated motifs, prompt sensitivity and viewer interpretations of narrative. A first enquiry should identify the question, required runs and proposed method. ## Access and citation Public images are watermarked derivatives. Clean originals remain private, with research access and reuse terms agreed individually. No dataset DOI or general reuse licence is assigned. Suggested descriptive reference: David Bryan Stevens, Difference Engine experiment index, snapshot October 10, 2026, V09 revision, After-Image Engine. The accompanying CSV and JSON identify the snapshot; retain its date when referring to it. Related work: Aditi Ramaswamy, Hana Chockler and Melane Navaratnarajah, “Defining and Quantifying Creative Behavior in Popular Image Generators,” ICCC 2025. Their short iterative chains use regenerated captions, providing a methodological connection to examine rather than proof of this archive's novelty. https://computationalcreativity.net/iccc25/papers/iccc25-ramaswamy2025defining.pdf Weekly update, September 26, 2026: seven new daily images (36–42) and their success-log dates were checked, along with seed/prompt continuity and the current image matching archive generation 42. No dates are missing this week. This is a weekly website snapshot of the daily experiment. It adds no batch-style lineage-hash claim. Weekly update, October 3, 2026: seven new daily images (43–49) and their success-log dates were checked, along with seed/prompt continuity and the current image matching archive generation 49. No dates are missing this week. This is a weekly website snapshot of the daily experiment. It adds no batch-style lineage-hash claim. Weekly update, October 10, 2026: seven new daily images (50–56) and their success-log dates were checked, along with seed/prompt continuity and the current image matching archive generation 56. No dates are missing this week. This is a weekly website snapshot of the daily experiment. It adds no batch-style lineage-hash claim.