---
title: DeepDream
emoji: 🌀
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.16.0
python_version: "3.10"
app_file: app.py
hardware: zero-a10g
pinned: false
short_description: The original 2015 DeepDream, in PyTorch on ZeroGPU
---

# 🌀 DeepDream

A faithful PyTorch reproduction of Alexander Mordvintsev's original 2015
[**DeepDream** notebook](https://github.com/google/deepdream) (`dream.ipynb`) — the one that
started *Inceptionism* — running on **ZeroGPU**.

Same algorithm as the original: normalized-L2 **gradient ascent** on a chosen GoogLeNet layer,
across **octaves**, with random **jitter**. Plus the notebook's two advanced modes:

- **Layer & model choice** — pick what it hallucinates. **ImageNet** (BVLC GoogLeNet) gives the
  classic dogs / eyes / pagodas; **Places365** turns everything into buildings, domes and
  landscapes.
- **Guided dreams** — supply a *guide* image and it dreams in that image's style (the notebook's
  "controlling dreams" dot-product objective).
- **Dream Zoom** — the endless feedback dream: dream, zoom in, repeat → an animation.
- **Live octave-by-octave streaming**, just like watching the original notebook render.

Uses the **original BVLC caffe weights** (converted to PyTorch, caffe layer names preserved),
hosted in this org at
[`mediasynthesismuseum/deepdream-googlenet`](https://huggingface.co/mediasynthesismuseum/deepdream-googlenet).

Preprocessing matches the original: BGR channel order, mean `[104,116,122]`, `0–255` range.

Credit: Google DeepDream (Mordvintsev, Olah, Tyka), BVLC Caffe Model Zoo, MIT Places, and
ProGamerGov's Caffe→PyTorch model conversions.