Torch implementation of inceptionism/deep-dream algorithm
Switch branches/tags
Nothing to show
Clone or download
Pull request Compare This branch is 4 commits ahead, 61 commits behind jcjohnson:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.


This is a fork of jcjohnson's neural art style project, modified to generate images using the techniques described in the Google Research blog Inceptionism. Hence the name "torch-inception"! I considered calling it nightmare-sauce, but decided that's for another project.

Most of what I did was delete code that wasn't relevant in this use case. The main addition is to multiply the syle layer's activations by the "acid_dose" constant (which can be negative!!) in order to train the image to accentuate that layer's features. Also I changed the default tag of gpu to be -1 so newbs that don't read README's won't think it requires a gpu ;)

I omit lbfgs as an optimization option because for some reason it always stops early with this technique.


  1. Install dependencies:


  1. Pick a model to use (some are provided in the models folder)
  2. Pick an input image
  3. Pick a level of abstaction ("macro" relu5_1, relu4_1, ..., relu1_1 "micro") (conv5_4 etc is also available)
  4. Run a command that looks like this
th neural_dream.lua -content_image ./examples/inputs/frida_kahlo.jpg -style_layers relu5_1 -save_iter 1 -num_iterations 30 -gpu -1 -acid_dose 1000

It can make weird images.

Depending on the dose, it can make brad look very different.

Which layer you choose can make a big difference. The above were taken from relu5_1. If we change to conv5_4, it gives brad a third eye.

There's significant difference between nearby conv layers. These are taken from conv5_1 and conv5_4, respectively.

Looking across the network gives a good feel for what the network is doing. These are taken from relu1_1, relu2_1, ..., relu5_1.

You can even do a negative dose, which results in subtracting the abstraction layer from the image.


  • -image_size: Maximum side length (in pixels) of of the generated image. Default is 512.
  • -gpu: Zero-indexed ID of the GPU to use; for CPU mode set -gpu to -1.

Optimization options:

  • -content_weight: How much to weight the content reconstruction term. Default is 5e0.
  • -num_iterations: Default is 1000.
  • -learning_rate: Learning rate to use with the ADAM optimizer. Default is 1e1.
  • -normalize_gradients: If this flag is present, style and content gradients from each layer will be L1 normalized. Idea from andersbll/neural_artistic_style.

Output options:

  • -output_image: Name of the output image. Default is out.png.
  • -print_iter: Print progress every print_iter iterations. Set to 0 to disable printing.
  • -save_iter: Save the image every save_iter iterations. Set to 0 to disable saving intermediate results.

Layer options:

  • -style_layers: Comman-separated list of layer names to use for style reconstruction. Default is relu1_1,relu2_1,relu3_1,relu4_1,relu5_1.

Other options:

  • -style_scale: Scale at which to extract features from the style image. Default is 1.0.
  • -proto_file: Path to the deploy.txt file for the VGG Caffe model.
  • -model_file: Path to the .caffemodel file for the VGG Caffe model. Default is the original VGG-19 model; you can also try the normalized VGG-19 model used in the paper.
  • -pooling: The type of pooling layers to use; one of max or avg. Default is max. The VGG-19 models uses max pooling layers, but the paper mentions that replacing these layers with average pooling layers can improve the results. I haven't been able to get good results using average pooling, but the option is here.
  • -backend: nn or cudnn. Default is nn. cudnn requires cudnn.torch and may reduce memory usage.

Frequently Asked Questions

Problem: Generated image has saturation artifacts:

Solution: Update the image packge to the latest version: luarocks install image

Problem: Running without a GPU gives an error message complaining about cutorch not found

Solution: Pass the flag -gpu -1 when running in CPU-only mode

Problem: The program runs out of memory and dies

Solution: Try reducing the image size: -image_size 256 (or lower). Note that different image sizes will likely require non-default values for -style_weight and -content_weight for optimal results. If you are running on a GPU, you can also try running with -backend cudnn to reduce memory usage.

Problem: Get the following error message:

models/VGG_ILSVRC_19_layers_deploy.prototxt.cpu.lua:7: attempt to call method 'ceil' (a nil value)

Solution: Update nn package to the latest version: luarocks install nn

Problem: Get an error message complaining about paths.extname

Solution: Update torch.paths package to the latest version: luarocks install paths

Memory Usage

By default, neural-style uses the nn backend for convolutions and L-BFGS for optimization. These give good results, but can both use a lot of memory. You can reduce memory usage with the following:

  • Use cuDNN: Add the flag -backend cudnn to use the cuDNN backend. This will only work in GPU mode.
  • Use ADAM: Add the flag -optimizer adam to use ADAM instead of L-BFGS. This should significantly reduce memory usage, but may require tuning of other parameters for good results; in particular you should play with the learning rate, content weight, style weight, and also consider using gradient normalization. This should work in both CPU and GPU modes.
  • Reduce image size: If the above tricks are not enough, you can reduce the size of the generated image; pass the flag -image_size 256 to generate an image at half the default size.

With the default settings, neural-style uses about 3.5GB of GPU memory on my system; switching to ADAM and cuDNN reduces the GPU memory footprint to about 1GB.


On a GTX Titan X, running 1000 iterations of gradient descent with -image_size=512 takes about 2 minutes. In CPU mode on an Intel Core i7-4790k, running the same takes around 40 minutes. Most of the examples shown here were run for 2000 iterations, but with a bit of parameter tuning most images will give good results within 1000 iterations.