DeepDream vs Generative Art
ディープドリーム / ジェネラティブアート
Both sit in Digital Art, so the question is not what they belong to but how they behave. They are reached for with some of the same intent, which is where they get confused.
DeepDream
Google's visualization technique repeatedly amplifies features learned by an image-recognition network, making eyes, animals and towers emerge excessively from clouds, rocks and leaves until misrecognition becomes a recursive hallucinatory surface.
Generative Art
A way of working in which the maker writes only the rule and the randomness, leaving the drawing to the machine. As in Nees's plotter work, the result is shown as several variants born of one rule.
| DeepDream | Generative Art | |
|---|---|---|
| Era | 2015– | 1965- |
| Family | Digital Art | Digital Art |
| Kind | Technique | Technique |
| Cues | Eyes and animal faces repeated across the frame / Swirling texture multiplying along source contours / Multi-scale repetition that reveals new figures when viewed closer / Saturated blue, yellow and violet rolling continuously | One simple element repeated to fill the field / Variants of a single rule shown side by side / Repetition drifting in angle and position / A density and precision no hand could reach |
| Best used for | Exhibits and critical images that make machine misrecognition immediately visible · Music video and cover imagery that keeps a source landscape beneath a multi-scale hallucination | Letting a machine make identifiers or one-off tickets that must all differ · Exhibitions or works where designing the rule is what is on show |
| Type | Add type after processing, keeping it short, bold and simple enough not to fight the detail. | Letters handled as coordinates and repetitions with the face itself fixed |
| Composition | Preserve the source's large composition and amplify related features at several scales to create recursion into depth. | Seed and random range set, outputs laid side by side to inspect the spread |
| Material | Choose layers in a trained vision model and iteratively increase their activation through gradient ascent. | Lines assumed at a plotter's constant weight, tone made by overlap density |
| Caution | It is not a swirl filter or pasted dog faces. Preserve which layer and feature were amplified and how they bind to the source. | A random range set too wide widens the gap between good and bad results and leaves a pile of output instead of a work. |