Improving Inference for Neural Image Compression

Yibo Yang, Robert Bamler, Stephan Mandt

Department of Computer Science
University of California, Irvine
{yibo.yang, rbamler, mandt}@uci.edu
arXiv:2006.04240v4 [eess.IV] 8 Jan 2021

Abstract
We consider the problem of lossy image compression with deep latent variable
models. State-of-the-art methods [Ballé et al., 2018, Minnen et al., 2018, Lee et al.,
2019] build on hierarchical variational autoencoders (VAEs) and learn inference
networks to predict a compressible latent representation of each data point. Drawing
on the variational inference perspective on compression [Alemi et al., 2018], we
identify three approximation gaps which limit performance in the conventional
approach: an amortization gap, a discretization gap, and a marginalization gap.
We propose remedies for each of these three limitations based on ideas related to
iterative inference, stochastic annealing for discrete optimization, and bits-back
coding, resulting in the first application of bits-back coding to lossy compression.
In our experiments, which include extensive baseline comparisons and ablation
studies, we achieve new state-of-the-art performance on lossy image compression
using an established VAE architecture, by changing only the inference method.

1 Introduction
Deep learning methods are reshaping the field of data compression, and recently started to outperform
state-of-the-art classical codecs on image compression [Minnen et al., 2018]. Besides useful on its
own, image compression is a stepping stone towards better video codecs [Lombardo et al., 2019,
Habibian et al., 2019, Yang et al., 2020a], which can reduce a sizable amount of global internet traffic.
State-of-the-art neural methods for lossy image compression [Ballé et al., 2018, Minnen et al.,
2018, Lee et al., 2019] learn a mapping between images and latent variables with a variational
autoencoder (VAE). An inference network maps a given image to a compressible latent representation,
which a generative network can then map back to a reconstructed image. In fact, compression can
be more broadly seen as a form of inference: to compress—or “encode”—data, one has to perform
inference over a well-specified decompression—or “decoding”—algorithm.
In classical compression codecs, the decoder has to follow a well-specified procedure to ensure
interoperability between different implementations of the same codec. By contrast, the encoding
process is typically not uniquely defined: different encoder implementations of the same codec often
compress the same input data to different bitstrings. For example, pngcrush [Randers-Pehrson, 1997]
typically produces smaller PNG files than ImageMagick. Yet, both programs are standard-compliant
PNG encoder implementations as both produce a compressed file that decodes to the same image. In
other words, both encoder implementations perform correct inference over the standardized decoder
specification, but use different inference algorithms with different performance characteristics.
The insight that better inference leads to better compression performance even within the same codec
motivates us to reconsider how inference is typically done in neural data compression with VAEs. In
this paper, we show that the conventional amortized inference [Kingma and Welling, 2013, Rezende
et al., 2014] in VAEs leaves substantial room for improvement when used for data compression.
We propose an improved inference method for data compression tasks with three main innovations:

34th Conference on Neural Information Processing Systems (NeurIPS 2020), Vancouver, Canada.
Figure 1: Graphical model and control flow charts. a) generative model with hyperlatents z, latents y,
and image x [Minnen et al., 2018]; b) conventional method for compression (dashed blue, see Eq. 1)
and decompression (solid black); c) common training objective due to [Ballé et al., 2017] (see
Eq. 3); d) proposed hybrid amortized (dashed blue) / iterative (dotted red) inference (Section 3.1);
e)-f) inference in the proposed lossy bitsback method (Section 3.3); the encoder first executes e) and
then keeps ŷ fixed while executing f); the decoder reconstructs ŷ and then executes f) to get bits back.

1. Improved amortization: The amortized inference strategy in VAEs speeds up training, but is
restrictive at compression time. We draw a connection between a recently proposed iterative
procedure for compression [Campos et al., 2019] to the broader literature of VI that closes
the amortization gap, which provides the basis of the following two novel inference methods.
2. Improved discretization: Compression requires discretizing the latent representation from
VAEs, because only discrete values can be entropy coded. As inference over discrete
variables is difficult, existing methods typically relax the discretization constraint in some
way during inference and then discretize afterwards. We instead propose a novel method
based on a stochastic annealing scheme that performs inference directly over discrete points.
3. Improved entropy coding: In lossless compression with latent-variable models, bits-back
coding [Wallace, 1990, Hinton and Van Camp, 1993] allows approximately coding the
latents with the marginal prior. It is so far believed that bits-back coding is incompatible
with lossy compression [Habibian et al., 2019] because it requires inference on the decoder
side, which does not have access to the exact (undistorted) input data. We propose a remedy
to this limitation, resulting in the first application of bits-back coding to lossy compression.

We evaluate the above three innovations on an otherwise unchanged architecture of an established

model and compare against a wide range of baselines and ablations. Our proposals significantly
improve compression performance and results in a new state of the art in lossy image compression.
The rest of the paper is structured as follows: Section 2 summarizes lossy compression with VAEs,
for which Section 3 proposes the above three improvements. Section 4 reports experimental results.
We conclude in Section 5. We review related work in each relevant subsection.

2 Background: Lossy Neural Image Compression as Variational Inference

In this section, we summarize an existing framework for lossy image compression with deep latent
variable models, which will be the basis of three proposed improvements in Section 3.

Related Work. Ballé et al. [2017] and Theis et al. [2017] were among the first to recognize a
connection between the rate-distortion objective of lossy compression and the loss function of a certain
kind of Variational Autoencoders (VAEs), and to apply it to end-to-end image compression. [Ballé
et al., 2018] proposes a hierarchical model, upon which current state-of-the-art methods [Minnen
et al., 2018, Lee et al., 2019] further improve by adding an auto-regressive component. For simplicity,
we adopt the VAE of [Minnen et al., 2018] without the auto-regressive component, reviewed in this
section below, which has also been a basis for recent compression research [Johnston et al., 2019].

Generative Model. Figure 1a shows the generative process of an image x. The Gaussian likelihood
p(x|y) = N (x; g(y; θ), σx2 I) has a fixed variance σx2 and its mean is computed from latent variables y
by a deconvolutional neural network (DNN) g with weights θ. A second DNN gh with weights θh
outputs the parameters (location and scale) of the prior p(y|z) conditioned on “hyperlatents” z.

2
Compression and Decompression. Figure 1b illustrates compression (“encoding”, dashed blue)
and decompression (“decoding”, solid black) as proposed in [Minnen et al., 2018]. The encoder passes
a target image x through trained inference networks f and fh with weights φ and φh , respectively.
The resulting continuous latent representations, (µy , µz ), are rounded (denoted b·e) to discrete (ŷ, ẑ),
ŷ = bµy e; ẑ = bµz e where µy = f (x; φ); µz = fh (f (x; φ); φh ). (1)
The encoder then entropy-codes ẑ and ŷ, using discretized versions P (ẑ) and P (ŷ|ẑ) of the hyperprior
p(z) and (conditional) prior p(y|z), respectively, as entropy models. This step is simplified by a
clever restriction on the shape of p, such that it agrees with the discretized P on all integers (see
[Ballé et al., 2017, 2018, Minnen et al., 2018]). The decoder then recovers ẑ and ŷ and obtains a
lossy image reconstruction x0 := arg maxx p(x|ŷ) = g(ŷ; θ).

Model Training. Let x be sampled from a training set of images. To train the above VAE for lossy
compression, Theis et al. [2017], Minnen et al. [2018] consider minimizing a rate-distortion objective:
Lλ (bµy e, bµz e) = R(bµy e, bµz e) + λD(bµy e, x)



2
= − log2 p bµz e − log2 p bµy e bµz e + λ x − g bµy e; θ 2 (2)
| {z }| {z } | {z }
information content information content of distortion D
of ẑ = bµz e ŷ = bµy e given ẑ = bµz e
where (µy , µz ) is computed from the inference networks as in Eq. 1, and the parameter λ > 0
controls the trade-off between bitrate R under entropy coding, and distortion (reconstruction error) D.
As the rounding operations prevent gradient-based optimization, Ballé et al. [2017] propose to replace
rounding during training by adding uniform noise from the interval [− 12 , 12 ] to each coordinate of µy
and µz (see Figure 1c). This is equivalent to sampling from uniform distributions q(y|x) and q(z|x)
with a fixed width of one centered around µy and µz , respectively. One thus obtains the following
relaxed rate-distortion objective, for a given data point x,
h i
2
L̃λ (θ, θh , φ, φh ) = Eq(y|x) q(z|x) − log2 p(z) − log2 p(y|z) + λ||x − g(y; θ)||2 . (3)

Connection to Variational Inference. As pointed out in [Ballé et al., 2017], the relaxed objective
in Eq. 3 is the negative evidence lower bound (NELBO) of variational inference (VI) if we identify
λ = 1/(2σx2 log 2). This draws a connection between lossy compression and VI [Blei et al., 2017,
Zhang et al., 2019]. We emphasize a distinction between variational inference and variational
expectation maximization (EM) [Beal and Ghahramani, 2003]: whereas variational EM trains a model
(and employs VI as a subroutine), VI is used to compress data using a trained model. A central result
of this paper is that improving inference in a fixed generative model at compression time already
suffices to significantly improve compression performance.

3 Novel Inference Techniques for Data Compression

This section presents our main contributions. We identify three approximation gaps in VAE-based
compression methods (see Section 2): an amortization gap, a discretization gap, and a marginalization
gap. Recently, the amortization gap was considered by [Campos et al., 2019]; we expand on this
idea in Section 3.1, bringing it under a wider framework of algorithms in the VI literature. We
then propose two specific methods that improve inference at compression time: a novel inference
method over discrete representations (Section 3.2) that closes the discretization gap, and a novel lossy
bits-back coding method (Section 3.3) that closes the marginalization gap.

3.1 Amortization Gap and Hybrid Amortized-Iterative Inference

Amortization Gap. Amortized variational inference [Kingma and Welling, 2013, Rezende et al.,
2014] turns optimization over local (per-data) variational parameters into optimization over the
global weights of an inference network. In the VAE of Section 2, the inference networks f and fh
map an image x to local parameters µy and µz of the variational distributions q(y|x) and q(z|x),
respectively (Eq. 1). Amortized VI speeds up training by avoiding an expensive inner inference loop
of variational EM, but it leads to an amortization gap [Cremer et al., 2018, Krishnan et al., 2018],
which is the difference between the value of the NELBO (Eq. 3) when µy and µz are obtained from
the inference networks, compared to its true minimum when µy and µz are directly minimized. Since

3
the NELBO approximates the rate-distortion objective (Section 2), the amortization gap translates
into sub-optimal performance when amortization is used at compression time. As discussed below,
this gap can be closed by refining the output of the inference networks by iterative inference.

Related Work. The idea of combining amortized inference with iterative optimization has been
studied and refined by various authors [Hjelm et al., 2016, Kim et al., 2018, Krishnan et al., 2018,
Marino et al., 2018]. While not formulated in the language of variational autoencoders and variational
inference, Campos et al. [2019] apply a simple version of this idea to compression. Drawing on
the connection to hybrid amortized-iterative inference, we show that the method can be drastically
improved by addressing the discretization gap (Section 3.2) and the marginalization gap (Section 3.3).

Hybrid Amortized-Iterative Inference. We reinterpret the proposal in [Campos et al., 2019] as a

basic version of a hybrid amortized-iterative inference idea that changes inference only at compression
time but not during model training. Figure 1d illustrates the approach. When compressing a target
image x, one initializes µy and µz from the trained inference networks f and fh , see Eq. 1. One then
treats µy and µz as local variational parameters, and minimizes the NELBO (Eq. 3) over (µy , µz )
with the reparameterization trick and stochastic gradient descent. This approach thus separates
inference at test time from inference during model training. We show in the next two sections
that this simple idea forms a powerful basis for new inference approaches that drastically improve
compression performance.

3.2 Discretization Gap and Stochastic Gumbel Annealing (SGA)

Discretization Gap. Compressing data to a bitstring is an inherently discrete optimization problem.

As discrete optimization in high dimensions is difficult, neural compression methods instead optimize
some relaxed objective function (such as the NELBO L̃λ in Eq. 3) over continuous representations
(such as µy and µz ), which are then discretized afterwards for entropy coding (see Eq. 1). This leads
to a discretization gap: the difference between the true rate-distortion objective Lλ at the discretized
representation (ŷ, ẑ), and the relaxed objective function L̃λ at the continuous approximation (µy , µz ).

Related Work. Current neural compression methods use a differentiable approximation to dis-
cretization during model training, such as Straight-Through Estimator (STE) [Bengio et al., 2013,
Oord et al., 2017, Yin et al., 2019], adding uniform noise [Ballé et al., 2017] (see Eq. 3), stochastic
binarization [Toderici et al., 2016], and soft-to-hard quantization [Agustsson et al., 2017]. Discretiza-
tion at compression time was addressed in [Yang et al., 2020b] without assuming that the training
procedure takes discretization into account, whereas our work makes this additional assumption.

Stochastic Gumbel Annealing (SGA). We refine the hybrid amortized-iterative inference ap-
proach of Section 3.1 to close the discretization gap: for a given image x, we aim to find its discrete
(in our case, integer) representation (ŷ, ẑ) that optimizes the rate-distortion objective Lλ in Eq. 2, this
time re-interpreted as a function of (ŷ, ẑ) directly. Our proposed annealing scheme approaches this
discretization optimization problem with the help of continuous proxy variables µy and µz , which
we initialize from the inference networks as in Eq. 1. We limit the following discussion to the latents
µy and ŷ, treating the hyperlatents µz and ẑ analogously. For each dimension i of µy , we map the
continuous coordinate µy,i ∈ R to an integer coordinate ŷi ∈ Z by rounding either up or down. Let
ry,i be a one-hot vector that indicates the rounding direction, with ry,i = (1, 0) for rounding down
(denoted b·c) and ry,i = (0, 1) for rounding up (denoted d·e). Thus, the result of rounding is the
inner product, ŷi = ry,i · (bµy,i c, dµy,i e). Now we let ry,i be a Bernoulli variable with a “tempered”
distribution
( 

exp − ψ µy,i − bµy,i c /τ if ry,i = (1, 0)
qτ (ry,i |µy,i ) ∝ 
(4)
exp − ψ dµy,i e − µy,i /τ if ry,i = (0, 1)
with temperature parameter τ > 0, and ψ : [0, 1) → R is an increasing function satisfying
limα→1 ψ(α) = ∞ (this is chosen to ensure continuity of the resulting objective (5); we used
ψ = tanh−1 ). Thus, as µy,i approaches Q an integer, the probability of rounding to that integer
approaches one. Defining qτ (ry |µy ) = i qτ (ry,i |µy,i ), we thus minimize the stochastic objective
h 

i
L̃λ,τ (µy , µz ) = Eqτ (ry |µy ) qτ (rz |µz ) Lλ ry · bµy c, dµy e , rz · bµz c, dµz e (5)

4
 +2.698×105 0.45
5 95

true R-D loss

L (Eq. 3)
Inference Network Prediction
85
Trajectory of SGA Parameters
0.40
SGA Samples 75
latent coordinate 2 4
65 0 500 1000 1500 2000
number of optimization iterations
55

discretization gap
3
45 0.2

L − Le
35
0.0
2
25

15 0 500 1000 1500 2000

number of optimization iterations
1 5 [A1] MAP
1 2 3 4 5
[A2] STE
latent coordinate 1 [A3] Uniform Noise
[A4] Deterministic Annealing
Figure 2: Visualizing SGA optimization. White dashed lines [M1] SGA (proposed)
indicate the discretization grid. SGA samples are colored
by temperature (lighter color corresponds to lower tempera- Figure 3: Comparing Stochastic
ture). 10 Gumbel-softmax samples were used to approximate Gumbel Annealing (Section 3.2) to
expectations in the objective (5), and temperature τ = 0.1. alternatives (Table 1, [A1]-[A4]).

where we reintroduce the hyperlatents z, and ry · (bµy c, dµy e) denotes the discrete vector ŷ obtained
by rounding each coordinate µy,i of µy according to its rounding direction ry,i . At any temperature
τ , the objective in Eq. 5 smoothly interpolates the R-D objective in Eq. 2 between all integer points.
We minimize Eq. 5 over (µy , µz ) with stochastic gradient descent, propagating gradients through
Bernoulli samples via the Gumbel-softmax trick [Jang et al., 2016, Maddison et al., 2016] using, for
simplicity, the same temperature τ as in qτ . We note that in principle, REINFORCE [Williams, 1992]
can also work instead of Gumbel-softmax. We anneal τ towards zero over the course of optimization,
such that the Gumbel approximation becomes exact and the stochastic rounding operation converges
to the deterministic one in Eq. 1. We thus name this method “Stochastic Gumbel Annealing” (SGA).
Figure 2 illustrates the loss function (Eq. 5) landscape of SGA near its converged solution, for a given
Kodak image. Here, we visualize along the two coordinates of µy whose optimization trajectories
have the most variance among all coordinates, plotting the optimization trajectory of µy (magenta

line) and the stochastically rounded samples (red dots) from the inner product ry · bµy c, dµy e . The
final SGA sample converges to the deterministically rounded solution (white dot, (4, 4)), which is
significantly better than the initialization predicted by inference network (blue diamond, near (2, 2)).

Comparison to Alternatives. Alternatively, we could have started with Eq. 2, and optimized it as a
function of (µy , µz ), using existing discretization methods to relax rounding; the optimized (µy , µz )
would subsequently be rounded to (ŷ, ẑ). However, our method outperforms all these alternatives.
Specifically, we tried (with shorthands [A1]-[A4] for ‘ablation’, referring to Table 1 of Section 4):
[A1] MAP: completely ignoring rounding during optimization; [A2] Straight-Through Estimator
(STE) [Bengio et al., 2013]: rounding only on the forward pass of automatic differentiation; [A3] Uni-
form Noise [Campos et al., 2019]: optimizing Eq. 3 over (µy , µz ) at compression time, following the
Hybrid Amortized-Iterative Inference approach of Section 3.1; and [A4] Deterministic Annealing:
turning the stochastic rounding operations of SGA into deterministic weighted averages, by pushing
theexpectation operators w.r.t q in Eq. 5 to inside the argumentsof Lλ , resulting in the loss function



Lλ Eqτ (ry |µy ) [ry · bµy c, dµy e ], Eqτ (rz |µz ) [rz · bµz c, dµz e ] resembling [Agustsson et al., 2017].
We tested the above discretization methods on images from Kodak [Kodak], using a pre-trained hyper-
prior model (Section 2). Figure 3 (left) shows learning curves for the true rate-distortion objective Lλ ,
evaluated via rounding the intermediate (µy , µz ) throughout optimization. Figure 3 (right) shows the
discretization gap Lλ − L̃λ , where L̃λ is the relaxed objective of each method. Our proposed SGA
method closes the discretization gap and achieves the lowest true rate-distortion loss among all meth-
ods compared, using the same initialization. Deterministic Annealing also closed the discretization
gap, but converged to a worse solution. MAP naively converged to a non-integer solution, and STE

5
consistently diverged even with a tiny learning rate, as also observed by Yin et al. [2019] (following
their results, we changed the gradient of the backward pass from identity to that of ReLU or clipped
ReLU; neither helped). Uniform Noise produced a consistently negative discretization gap, as the
noise-perturbed objective in Eq. 3 empirically overestimates the true rate-distortion objective Lλ on a
large neighborhood around the initial (µy , µz ) computed by the inference network; by contrast, SGA
shrinks this gap by letting its variational distribution q increasingly concentrate on integer solutions.

3.3 Marginalization Gap and Lossy Bits-Back Coding

Marginalization Gap. To compress an image x with the hierarchical model of Section 2, ideally we
would only need to encode and transmit its latent representation ŷ but not the hyper-latents ẑ, as only ŷ
is needed for reconstruction.
R This would require entropy coding with (a discretization of) the marginal
prior p(y) = p(z) p(y|z) dz, which unfortunately is computationally intractable. Therefore, the
standard compression approach reviewed in Section 2 instead encodes some hyperlatent ẑ using the
discretized hyperprior P (ẑ) first, and then encodes ŷ using the entropy model P (ŷ|ẑ). This leads to
a marginalization gap, which is the difference between the information content − log2 P (ẑ, ŷ) =
− log2 P (ẑ) − log2 P (ŷ|ẑ) of the transmitted tuple (ẑ, ŷ) and the information content of ŷ alone,
− log2 P (ẑ, ŷ) − (− log2 P (ŷ)) = − log2 P (ẑ|ŷ), (6)
i.e., the marginalization gap is the information content of the hyperlatent ẑ under the ideal posterior
P (ẑ|ŷ), in the (discrete) latent generative process ẑ → ŷ.

Related Work. A similar marginalization gap has been addressed in lossless compression by bits-
back coding [Wallace, 1990, Hinton and Van Camp, 1993] and made more practical by the BB-ANS
algorithm [Townsend et al., 2019a]. Recent work has improved its efficiency in hierarchical latent
variable models [Townsend et al., 2019b, Kingma et al., 2019], and extended it to flow models [Ho
et al., 2019]. To our knowledge, bits-back coding has not yet been used in lossy compression. This is
likely since bits-back coding requires Bayesian posterior inference on the decoder side, which seems
incompatible with lossy compression, as the decoder does not have access to the undistorted data x.

Lossy Bits-Back Coding. We extend the bits-back idea to lossy compression by noting that the
discretized latent representation ŷ is encoded losslessly (Figure 1b) and thus amenable to bits-back
coding. A complication arises as bits-back coding requires that the encoder’s inference over z can be
exactly reproduced by the decoder (see below). This requirement is violated by the inference network
in Eq. 1, where µz = fh (f (x; φ); φh ) depends on x, which is not available to the decoder in lossy
compression. It turns out that the naive fix of setting instead µz = fh (ŷ; φh ) would hurt performance
by more than what bits-back saves (see ablations in Section 4). We propose instead a two-stage
inference algorithm that cuts the dependency on x after an initial joint inference over ŷ and z.
Bits-back coding bridges the marginalization gap in Eq. 6 by encoding a limited number of bits of
some additional side information (e.g., an image caption or a previously encoded image of a slide
show) into the choice of ẑ using an entropy model Q(ẑ|µz , σz2 ) with parameters µz and σz2 . We obtain
Q(ẑ|µz , σz2 ) by discretizing a Gaussian variational distribution q(z|µz , σz2 ) = N (z; µz , diag(σz2 )),
thus extending the last layer of the hyper-inference network fh (Eq. 1) to output now a tuple (µz , σz2 )
of means and diagonal variances. This replaces the form of variational distribution q(z|x) of Eq 3
as the restriction to a box-shaped distribution with fixed width is not necessary in bits-back coding.
Additionally, we drop the restriction on the shape of the hyperprior p(z) reviewed in Section 2, simply
adopting the flexible density model as proposed in [Ballé et al., 2018] without convolving it with
a uniform distribution, as µz is no longer discretized to integers. We train the resulting VAE by
minimizing the NELBO. Since q(z|µz , σz2 ) now has a variable width, the NELBO has an extra term
compared to Eq. 3 that subtracts the entropy of q(z|µz , σz2 ), reflecting the expected number of bits
we ‘get back’. Thus, the NELBO is again a relaxed rate-distortion objective, where now the net rate
is the compressed file size minus the amount of embedded side information.
Algorithm 1 describes lossy compression and decompression with the trained model. Subroutine
encode initializes the variational parameters µy , µz , and σz2 conditioned on the target image x using
the trained inference networks (line 1). It then jointly performs SGA over ŷ by following Section 3.2,
and Black-Box Variational Inference (BBVI) over µz and σz2 by minimizing the the NELBO (line 2).
At the end of the routine, the encoder decodes the provided side information ξ (an arbitrary bitstring)
into ẑ using Q(ẑ|µz , σz2 ) as entropy model (line 4) and then encodes ẑ and ŷ as usual (line 5).

6
 Algorithm 1: Proposed lossy bits-back coding method (Section 3.3 and Figure 1e-f).
Global Constants: Trained hierarchical VAE with model p(z, y, x) = p(z) p(y|z) p(x|y) and
inference networks f ( · ; φ) and fh ( · ; φh ), see Figure 1e (f is only used in subroutine encode).
Subroutine encode(image x, side information ξ) 7→ returns compressed bitstring s
1 Initialize µy ← f (x; φ) and (µz , σz2 ) ← fh (µy ; φh ). . Figure 1e (blue)
2 Optimize over ŷ using SGA (Section 3.2), and over µz and σz2 using BBVI. . Figure 1e (red)
3 Reset (µz , σz2 ) ← reproducible_BBVI(ŷ). . See below.
4 Decode side information ξ into ẑ using Q(ẑ|µz , σz2 ) as entropy model.
5 Encode ẑ and ŷ into s using P (ẑ) and P (ŷ|ẑ) as entropy models, respectively.
Subroutine decode(compressed bitstring s) 7→ returns (lossy reconstruction x0 , side info ξ)
6 Decode s into ẑ and ŷ; then get reconstructed image x0 = arg maxx p(x|ŷ).
7 Set (µz , σz2 ) ← reproducible_BBVI(ŷ). . See below.
8 Encode ẑ into ξ using Q(ẑ|µz , σz2 ) as entropy model.
Subroutine reproducible_BBVI(discrete latents ŷ) 7→ returns variational parameters (µz , σz2 )
9 Initialize (µz , σz2 ) ← fh (ŷ; φh ); seed random number generator reproducibly. . Figure 1f (blue)
10 Refine µz and σz2 by running BBVI for fixed ŷ. . Figure 1f (red)

The important step happens on line 3: up until this step, the fitted variational parameters µz and σz2
depend on the target image x due to their initialization on line 1. This would prevent the decoder,
which does not have access to x, from reconstructing the side information ξ by encoding ẑ with the
entropy model Q(ẑ|µz , σz2 ). Line 3 therefore re-fits the variational parameters µz and σz2 in a way
that is exactly reproducible based on ŷ alone. This is done in the subroutine reproducible_BBVI,
which performs BBVI in the prior model p(z, y) = p(z)p(y|z), treating y = ŷ as observed and
only z as latent. Although we reset µz and σz2 on line 3 immediately after optimizing over them on
line 2, optimizing jointly over both ŷ and (µz , σz2 ) on line 2 allows the method to find a better ŷ.
The decoder decodes ẑ and ŷ as usual (line 6). Since the subroutine reproducible_BBVI depends
only on ŷ and uses a fixed random seed (line 9), calling it from the decoder on line 7 yields the exact
same entropy model Q(ẑ|µz , σz2 ) as used by the encoder, allowing the decoder to recover ξ (line 8).

4 Experiments

We demonstrate the empirical effectiveness of our approach by applying two variants of it —with
and without bits-back coding—to an established (but not state-of-the-art) base model [Minnen et al.,
2018]. We improve its performance drastically, achieving an average of over 15% BD rate savings on
Kodak and 20% on Tecnick [Asuni and Giachetti, 2014], outperforming the previous state-of-the-art
of both classical and neural lossy image compression methods. We conclude with ablation studies.

Proposed Methods. Table 1 describes all compared methods, marked as [M1]-[M9] for short. We
tested two variants of our method ([M1] and [M2]): SGA builds on the exact same model and training
procedure as in [Minnen et al., 2018] (the Base Hyperprior model [M3]) and changes only how the
trained model is used for compression by introducing hybrid amortized-iterative inference [Campos
et al., 2019] (Section 3.1) and Stochastic Gumbel Annealing (Section 3.2). SGA+BB adds to it
bits-back coding (Section 3.3), which requires changing the inference model over hyperlatents z to
admit a more flexible variational distribution q(z|x), and no longer restricts the shape of p(z), as
discussed in Section 3.3. The two proposed variants address different use cases: SGA is a “standalone”
variant of our method that compresses individual images, while SGA+BB encodes images with
additional side information. In all results, we used Adam [Kingma and Ba, 2015] for optimization,
and annealed the temperature of SGA by an exponential decay schedule, and found good convergence
without per-model hyperparameter tuning. We provide details in the Supplementary Material.

Baselines. We compare to the Base Hyperprior model from [Minnen et al., 2018] without our
proposed improvements ([M3] in Table 1), two state-of-the-art neural methods ([M4] and [M5]), two

7
Table 1: Compared methods ([M1]-[M9]) and ablations ([A1]-[A6]). We propose two variants:
[M1] compresses images, and [M2] compresses images + side information via bits-back coding.
Name Explanation and Reference (for baselines)
[M1] SGA Proposed standalone variant: same trained model as [M3], SGA (Section 3.2) at compression time.
ours

[M2] SGA + BB Proposed variant with bits-back coding: both proposed improvements of Sections 3.2, and 3.3.
[M3] Base Hyperprior Base method of our two proposals, reviewed in Section 2 of the present paper [Minnen et al., 2018].
[M4] Context + Hyperprior Like [M3] but with an extra context model defining the prior p(y|z) [Minnen et al., 2018].
baselines

[M5] Context-Adaptive Context-Adaptive Entropy Model proposed in [Lee et al., 2019].

[M6] Hyperprior Scale-Only Like [M3] but the hyperlatents z model only the scale (not the mean) of p(y|z) [Ballé et al., 2018].
[M7] CAE Pioneering “Compressive Autoencoder” model proposed in [Theis et al., 2017].
[M8] BPG 4:4:4 State-of-the-art classical lossy image compression codec ‘Better Portable Graphics’ [Bellard, 2014].
[M9] JPEG 2000 Classical lossy image compression method [Adams, 2001].
[A1] MAP Like [M1] but with continuous optimization over µy and µz followed by rounding instead of SGA.
[A2] STE Like [M1] but with straight-through estimation (i.e., round only on backpropagation) instead of SGA.
ablations

[A3] Uniform Noise Like [M1] but with uniform noise injection instead of SGA (see Section 3.1) [Campos et al., 2019].
[A4] Deterministic Annealing Like [M1] but with deterministic version of Stochastic Gumbel Annealing
[A5] BB without SGA Like [M2] but without optimization over ŷ at compression time.
[A6] BB without iterative inference Like [M2] but without optimization over ŷ, µz , and σz2 at compression time.

other neural methods ([M6] and [M7]), the state-of-the art classical codec BPG [M8], and JPEG
2000 [M9]. We reproduced the Base Hyperprior results, and took the other results from [Ballé et al.].

Results. Figure 4 compares the compression performance of our proposed method to existing
baselines on the Kodak dataset, using the standard Peak Signal-to-Noise Ratio (PSNR) quality metric
(higher is better), averaged over all images for each considered quality setting λ ∈ [0.001, 0.08]. The
bitrates in our own results ([M1-3], [A1-6]) are based on rate estimates from entropy models, but
do not differ significantly from actual file sizes produced by our codec implementation (which has
negligible (< 0.5%) overhead). The left panel of Figure 4 plots PSNR vs. bitrate, and the right panel
shows the resulting BD rate savings [Bjontegaard, 2001] computed relative to BPG as a function
of PSNR for readability (higher is better). The BD plot cuts out CAE [M7] and JPEG 2000 [M9] at
the bottom, as they performed much worse. Overall, both variants of the proposed method (blue
and orange lines in Figure 4) improve substantially over the Base Hyperprior model (brown), and
outperform the previous state-of-the-art. We report similar results on the Tecnick dataset [Asuni
and Giachetti, 2014] in the Supplementary Material. Figure 6 shows a qualitative comparison of a
compressed image (in the order of BPG, SGA (proposed), and the Base Hyperprior), in which we
see that our method notably enhances the baseline image reconstruction at comparable bitrate, while
avoiding unpleasant visual artifacts of BPG.
Ablations. Figure 5 compares BD rate improvements of the two variants of our proposal (blue and
orange) to six alternative choices (ablations [A1]-[A6] in Table 1), measured relative to the Base
Hyperprior model [Minnen et al., 2018] (zero line; [M3] in Table 1), on which our proposals build.
[A1]-[A4] replace the discretization method of SGA [M1] with four alternatives discussed at the end
of Section 3.2. [A5] and [A6] are ablations of our proposed bits-back variant SGA+BB [M2], which
remove iterative optimization over ŷ, or over ŷ and q(z), respectively, at compression time. Going
from red [A3] to orange [A2] to blue [M1] traces our proposed improvements over [Campos et al.,
2019] by adding SGA (Section 3.2) and then bits-back (Section 3.3). It shows that iterative inference

40 20
BD rate saving (%) over Base Model

20
BD rate saving (%) over BPG

15
38

10
36 10
5
PSNR

34
0
0
[M2] SGA+BB [M3] Minnen (2018) [M2] SGA+BB
32 −5 (proposed, bits-back) Base Hyperprior (proposed, bits-back) [A5] BB without SGA
[M1] SGA [M6] Ballé (2018) [M1] SGA [A2] Straight-Through
−10 (proposed, standalone) Hyperprior Scale-Only Estimator
−10 (proposed, standalone)
30 [M7] Theis (2017) CAE [M3] Base Hyperprior
[M4] Minnen (2018) [A4] Deterministic
Context+Hyperprior Minnen (2018)
−15 [M9] JPEG2K Annealing
[M5] Lee (2019) [A3] Uniform Noise [A6] BB without
28 Iterative Inference
Context Adaptive Campos (2019)
−20 −20
[M8] BPG 4:4:4 [A1] MAP
26
0.2 0.4 0.6 0.8 1.0 1.2 28 30 32 34 36 38 28 30 32 34 36 38
Bits per pixel (BPP) PSNR PSNR

Figure 4: Compression performance comparisons on Kodak against Figure 5: BD rate savings of var-
existing baselines. Left: R-D curves. Right: BD rate savings (%) ious ablation methods relative to
relative to BPG. Legend shared; higher values are better in both. Base Hyperprior on Kodak.

8
(a) Original image (b) BPG 4:4:4; (c) Proposed (SGA); (d) [Minnen et al., 2018];
(from Kodak dataset) 0.14 BPP (PSNR = 30.4) 0.13 BPP (PSNR = 31.2) 0.12 BPP (PSNR = 29.8)
Figure 6: Qualitative comparison of lossy compression performance. Our method (c; [M1] in Table 1)
significantly boosts the visual quality of the Base Hyperprior method (d; [M3] in Table 1) at similar
bit rates. It sharpens details of the hair and face (see insets) obscured by the baseline method (d),
while avoiding the ringing artifacts around the jaw and ear pendant produced by a classical codec (b).

(Section 3.1) and SGA contribute approximately equally to the performance gain, whereas the gain
from bits-back coding is smaller. Interestingly, lossy bits-back coding without iterative inference and
SGA actually hurts performance [A6]. As Section 3.3 mentioned, this is likely because bits-back
coding constrains the posterior over z to be conditioned only on ŷ and not on the original image x.

5 Discussion

Starting from the variational inference view on data compression, we proposed three enhancements to
the standard inference procedure in VAEs: hybrid amortized-iterative inference, Stochastic Gumbel
Annealing, and lossy bits-back coding, which translated to dramatic performance gains on lossy image
compression. Improved inference provides a new promising direction to improved compression,
orthogonal to modeling choices (e.g., with auto-regressive priors [Minnen et al., 2018], which can
harm decoding efficiency). Although lossy-bits-back coding in the present VAE only gave relatively
minor benefits, it may reach its full potential in more hierarchical architectures as in [Kingma
et al., 2019]. Similarly, carrying out iterative inference also at training time may lead to even more
performance improvement, with techniques like iterative amortized inference [Marino et al., 2018].

Broader Impacts

Improved image and video compression is becoming more and more important as our lives become
more digital. For instance, video compression is of enormous societal relevance, as over 80% of web
traffic is due to video streaming, and the share of video data is expected to increase even further in the
future [Cisco, 2017]. Moreover, there is an explosion of different new data formats that need to be
efficiently compressed; examples include high resolution medical images, LIDAR data, 3D graphics
data, DNA sequences, etc. It would be very costly to design custom codecs for such data, making
learnable compression indispensable. Ultimately, better compression algorithms, both for data and
models, will also facilitate the possibility to carry-out machine learning on local devices, as opposed
to large centralized servers. This may address many of our privacy concerns.
Currently, neural compression comes with a few drawbacks compared to hand-engineered codecs,
such as higher computation/storage/energy cost, and lack of robustness, which we hope can be
overcome by active research in this area. Our work on improving the encoding procedure already
has the potential to lower the overall resource consumption of a neural codec, for applications where
the increased cost of encoding can be amortized across many file transfer/decoding operations and
translate to net savings, e.g., on a content hosting website that has millions of visitors per day.

9
Acknowledgments and Disclosure of Funding
We thank Yang Yang for valuable feedback on the manuscript. Yibo Yang acknowledges funding
from the Hasso Plattner Foundation. This material is based upon work supported by the Defense
Advanced Research Projects Agency (DARPA) under Contract No. HR001120C0021. Any opinions,
findings and conclusions or recommendations expressed in this material are those of the author(s) and
do not necessarily reflect the views of the Defense Advanced Research Projects Agency (DARPA).
Furthermore, this work was supported by the National Science Foundation under Grants 1928718,
2003237 and 2007719, and by Qualcomm. Stephan Mandt consulted for Disney and Google.

References
Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational
image compression with a scale hyperprior. International Conference on Learning Representations,
2018.
David Minnen, Johannes Ballé, and George D Toderici. Joint autoregressive and hierarchical priors
for learned image compression. In Advances in Neural Information Processing Systems, pages
10771–10780, 2018.
Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beack. Context-adaptive entropy model for
end-to-end optimized image compression. In the 7th Int. Conf. on Learning Representations, May
2019.
Alexander Alemi, Ben Poole, Ian Fischer, Joshua Dillon, Rif A Saurous, and Kevin Murphy. Fixing a
broken ELBO. In International Conference on Machine Learning, pages 159–168. PMLR, 2018.
Salvator Lombardo, Jun Han, Christopher Schroers, and Stephan Mandt. Deep generative video
compression. In Advances in Neural Information Processing Systems 32, pages 9287–9298, 2019.
Amirhossein Habibian, Ties van Rozendaal, Jakub M Tomczak, and Taco S Cohen. Video compression
with rate-distortion autoencoders. In Proceedings of the IEEE International Conference on
Computer Vision, pages 7033–7042, 2019.
Ruihan Yang, Yibo Yang, Joseph Marino, and Stephan Mandt. Hierarchical autoregressive modeling
for neural video compression. arXiv preprint arXiv:2010.10258, 2020a.
Glenn Randers-Pehrson. Mng: a multiple-image format in the png family. World Wide Web Journal,
2(1):209–211, 1997.
Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint
arXiv:1312.6114, 2013.
Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and
approximate inference in deep generative models. In International Conference on Machine
Learning, pages 1278–1286, 2014.
Joaquim Campos, Simon Meierhans, Abdelaziz Djelouah, and Christopher Schroers. Content adaptive
optimization for neural image compression. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition Workshops, 2019.
Chris S Wallace. Classification by minimum-message-length inference. In International Conference
on Computing and Information, pages 72–81. Springer, 1990.
Geoffrey E Hinton and Drew Van Camp. Keeping the neural networks simple by minimizing the
description length of the weights. In Proceedings of the sixth annual conference on Computational
learning theory, pages 5–13, 1993.
Johannes Ballé, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression.
International Conference on Learning Representations, 2017.
Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with
compressive autoencoders. International Conference on Learning Representations, 2017.

10
Nick Johnston, Elad Eban, Ariel Gordon, and Johannes Ballé. Computationally efficient neural image
compression. arXiv preprint arXiv:1912.08771, 2019.
David M Blei, Alp Kucukelbir, and Jon D McAuliffe. Variational inference: A review for statisticians.
Journal of the American statistical Association, 112(518):859–877, 2017.
Cheng Zhang, Judith Butepage, Hedvig Kjellstrom, and Stephan Mandt. Advances in variational
inference. IEEE transactions on pattern analysis and machine intelligence, 2019.
Matthew J Beal and Zoubin Ghahramani. The variational Bayesian EM algorithm for incomplete
data: with application to scoring graphical model structures. In Bayesian Statistics 7: Proceedings
of the Seventh Valencia International Meeting, volume 7, pages 453–464. Oxford University Press,
2003.
Chris Cremer, Xuechen Li, and David Duvenaud. Inference suboptimality in variational autoencoders.
In International Conference on Machine Learning, pages 1078–1086, 2018.
Rahul Krishnan, Dawen Liang, and Matthew Hoffman. On the challenges of learning with inference
networks on sparse, high-dimensional data. In International Conference on Artificial Intelligence
and Statistics, pages 143–151, 2018.
Devon Hjelm, Russ R Salakhutdinov, Kyunghyun Cho, Nebojsa Jojic, Vince Calhoun, and Junyoung
Chung. Iterative refinement of the approximate posterior for directed belief networks. In Advances
in Neural Information Processing Systems, pages 4691–4699, 2016.
Yoon Kim, Sam Wiseman, Andrew Miller, David Sontag, and Alexander Rush. Semi-amortized
variational autoencoders. In International Conference on Machine Learning, pages 2678–2687,
2018.
Joe Marino, Yisong Yue, and Stephan Mandt. Iterative amortized inference. In Proceedings of the
35th International Conference on Machine Learning, pages 3403–3412, 2018.
Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through
stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.
Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning.
arXiv preprint arXiv:1711.00937, 2017.
Penghang Yin, Jiancheng Lyu, Shuai Zhang, Stanley Osher, Yingyong Qi, and Jack Xin. Under-
standing straight-through estimator in training activation quantized neural nets. arXiv preprint
arXiv:1903.05662, 2019.
George Toderici, Sean M O’Malley, Sung Jin Hwang, Damien Vincent, David Minnen, Shumeet
Baluja, Michele Covell, and Rahul Sukthankar. Variable rate image compression with recurrent
neural networks. International Conference on Learning Representations, 2016.
Eirikur Agustsson, Fabian Mentzer, Michael Tschannen, Lukas Cavigelli, Radu Timofte, Luca
Benini, and Luc V Gool. Soft-to-hard vector quantization for end-to-end learning compressible
representations. In Advances in Neural Information Processing Systems, 2017.
Yibo Yang, Robert Bamler, and Stephan Mandt. Variational Bayesian Quantization. In International
Conference on Machine Learning, 2020b.
Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv
preprint arXiv:1611.01144, 2016.
Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous
relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016.
Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement
learning. Machine learning, 8(3-4):229–256, 1992.
Eastman Kodak. Kodak lossless true color image suite (PhotoCD PCD0992). URL http://r0k.
us/graphics/kodak.

11
James Townsend, Tom Bird, and David Barber. Practical lossless compression with latent variables
using bits back coding. arXiv preprint arXiv:1901.04866, 2019a.
James Townsend, Thomas Bird, Julius Kunze, and David Barber. Hilloc: Lossless image compression
with hierarchical latent variable models. arXiv preprint arXiv:1912.09953, 2019b.
Friso H Kingma, Pieter Abbeel, and Jonathan Ho. Bit-swap: Recursive bits-back coding for lossless
compression with hierarchical latent variables. arXiv preprint arXiv:1905.06845, 2019.
Jonathan Ho, Evan Lohn, and Pieter Abbeel. Compression with flows via local bits-back coding. In
Advances in Neural Information Processing Systems, pages 3874–3883, 2019.
N. Asuni and A. Giachetti. TESTIMAGES: A large-scale archive for testing visual devices and
basic image processing algorithms (SAMPLING 1200 RGB set). In STAG: Smart Tools and Apps
for Graphics, 2014. URL https://sourceforge.net/projects/testimages/files/OLD/
OLD_SAMPLING/testimages.zip.
Diederik P Kingma and Jimmy Lei Ba. Adam: A method for stochastic gradient descent. In
International Conference on Learning Representations, 2015.
Fabrice Bellard. BPG specification, 2014. URL https://bellard.org/bpg/bpg_spec.txt.
(accessed June 3, 2020).
Michael D Adams. The jpeg-2000 still image compression standard. 2001.
Johannes Ballé, Sung Jin Hwang, Nick Johnston, and David Minnen. Tensorflow-compression: Data
compression in tensorflow. URL https://github.com/tensorflow/compression.
Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. VCEG-M33, 2001.
Visual Network Index Cisco. Forecast and methodology, 2016-2021. White Paper, 2017.

12
 Supplementary Material to
Improving Inference for Neural Compression

S1 Stochastic Annealing
Here we provide conceptual illustrations of our stochastic annealing idea on a simple example.
Consider the following scalar optimization problem over integers,
minimize f (z) = z 2
z∈Z

Following our stochastic annealing method in Section. 3.2, we let x ∈ R be a continuous proxy
variable, and r ∈ {(1, 0), (0, 1)} be the one-hot vector of stochastic rounding direction. We let r
follow a Bernoulli distribution with temperature τ > 0,
( 

exp − ψ x − bxc /τ /C(x, τ ) if r = (1, 0)
qτ (r|x) = 

exp − ψ dxe − x /τ /C(x, τ ) if r = (0, 1)
where
C(x, τ ) = exp{−ψ(x − bxc)/τ } + exp{−ψ(dxe − x)/τ }
is the normalizing constant.
The stochastic optimization problem for SGA is then
minimize `τ (x) := Eqτ (r|x) [f (r · (bxc, dxe)]
x∈R
exp{−ψ(x − bxc)/τ } exp{−ψ(dxe − x)/τ }
= f (bxc) + f (dxe)
C(x, τ ) C(x, τ )

Below we plot the rounding probability qτ (r = (0, 1)|x) and the SGA objective `τ as a function of
x, at various temperatures.

1.0 4

0.8 3
0.6
2 τ = 0.5
0.4 τ = 0.2
τ = 0.5 1 τ = 0.01
0.2 τ = 0.2 z2
τ = 0.01 0
0.0 −2 −1 0 1 2
0.00 0.25 0.50 0.75 1.00
Figure S2: Graphs of stochastic annealing objec-
Figure S1: Illustration of the probability qτ (r = tive `τ at various temperatures. `τ interpolates
(0, 1)|x) of rounding up, on the interval (0, 1), at the discrete optimization problem between inte-
various temperatures. gers, and approaches round(x)2 as τ → 0.

S2 Model Architecture and Training

As mentioned in the main text, the Base Hyperprior model uses the same architecture as in Table 1 of
Minnen et al. [2018] except without the “Context Prediction” and “Entropy Parameters” components
(this model was referred to as “Mean & Scale Hyperprior” in this work). The model is mostly
already implemented in bmshj2018.py1 from Ballé et al., Ballé et al. [2018], and we modified their
implementation to double the number of output channels of the HyperSynthesisTransform to
predict both the mean and (log) scale of the conditional prior p(y|z) over the latents y.
As mentioned in Section 3.3 and 4, lossy bits-back modifies the above Base Hyperprior as follows:
1
https://github.com/tensorflow/compression/blob/master/examples/bmshj2018.py

1
 1. The number of output channels of the hyper inference network is doubled to compute both
the mean and diagonal (log) variance of Gaussian q(z|x);
2. The hyperprior p(z) is no longer restricted to the form of a flexible density model convolved
with the uniform distribution on [−0.5, 0.5]; instead it simply uses the flexible density, as
described in the Appendix of Ballé et al. [2018].

We trained our models (both the Base Hyperprior and the bits-back variant) on CLIC-2018
2
images using minibatches of eight 256 × 256 randomly-cropped image patches, for λ ∈
{0.001, 0.0025, 0.005, 0.01, 0.02, 0.04, 0.08}. Following Ballé et al. [2018], we found that in-
creasing the number of latent channels helps avoid the “bottleneck” effect and improve rate-
distortion performance at higher rates, and increased the number of latent channels from 192
to 256 for models trained with λ = 0.04 and 0.08. All models were trained for 2 million
steps, except the ones with λ = 0.001 which were trained for 1 million steps, and λ = 0.08
which were trained for 3 million steps. Our code and pre-trained models can be found at https:
//github.com/mandt-lab/improving-inference-for-neural-image-compression.

S3 Hyper-Parameters of Various Methods Considered

Given a pre-trained model (see above) and image(s) x to be compressed, our experiments in Section 4
explored hybrid amortized-iterative inference to improve compression performance. Below we
provide hyper-parameters of each method considered:

Stochastic Gumbel Annealing. In all experiments using SGA (including the SGA+BB experiments,
which optimized over ŷ using SGA), we used the Adam optimizer with an initial learning rate of
0.005, and an exponentially decaying temperature schedule τ (t) = min(exp{−ct}, 0.5), where t is
the iteration number, and c > 0 controls the speed of the decay. We found that lower c generally
increases the number of iterations needed for convergence, but can also yield slightly better solutions;
in all of our experiments we set c = 0.001, and obtained good convergence with 2000 iterations.

Temperature Annealing Schedule for SGA. We initially used a naive temperature schedule,
τ (t) = τ0 exp{−ct} for SGA, where τ0 is the initial temperature (typically set to 0.5 to simulate soft
quantization), and found that aggressive annealing with a large decay factor c can lead to suboptimal
solutions, as seen with c = 0.002 or c = 0.001 in the left subplot of Figure S3. We found that we
can overcome the suboptimality from fast annealing by an initial stage of optimization at a fixed
temperature before annealing: in the initial stage, we fix the temperature to some relatively high value
τ0 (to simulate soft discretization), and run the optimization for t0 steps such that the R-D objective
roughly converges. We demonstrate this in the right subplot of Figure S3, where we modify the naive
schedule to τ (t) = min{τ0 , τ0 exp{−c(t − t0 )}}, so that SGA roughly converges after t0 = 700
steps at temperature τ0 = 0.5 before annealing. As can be seen, the results of SGA are robust to
different choices of decay factor c, with c = 0.002 and c = 0.001 giving comparably good results to
c = 0.0005 but with faster convergence.

0.42
τ (t) = 0.5e−ct τ (t) = min{0.5, 0.5e−c(t−700)}
true R-D loss
L (Eq. 3)

c = 0.002
0.40
c = 0.001
c = 0.0005
0.38

0 1000 2000 3000 0 1000 2000 3000

number of optimization iterations (t) number of optimization iterations (t)
Figure S3: Comparing different annealing schedules for SGA (Section 3.2).

Lossy bits-back. In the SGA+BB experiments, line 2 of the encode subroutine of Algorithm 1
performs joint optimization w.r.t. µy , µz , and σz2 with Black-box Variational Inference (BBVI), using
2
https://www.compression.cc/2018/challenge/

2
the reparameterization trick to differentiate through Gumbel-softmax samples for µy , and Gaussian
samples for (µz , σz2 ). We used Adam with an initial learning rate of 0.005, and ran 2000 stochastic
gradient descent iterations; the optimization w.r.t. µy (SGA) used the same temperature annealing
schedule as above. The reproducible_BBVI subroutine of Algorithm 1 used Adam with an initial
learning rate of 0.003 and 2000 stochastic gradient descent iterations; we used the same random seed
for the encoder and decoder for simplicity, although a hash of ŷ can also be used instead.

Alternative Discretization Methods In Figure 3 and ablation studies of Section 4, all alternative
discretization methods were optimized with Adam for 2000 iterations to compare with SGA. The
initial learning rate was 0.005 for MAP, Uniform Noise, and Deterministic Annealing, and 0.0001
for STE. All methods were tuned on a best-effort basis to ensure convergence, except that STE
consistently encountered convergence issues even with a tiny learning rate (see [Yin et al., 2019]).
The rate-distortion results for MAP and STE were calculated with early stopping (i.e., using the
intermediate (ŷ, ẑ) with the lowest true rate-distortion objective during optimization), just to give
them a fair chance. Lastly, the comparisons in Figure 3 used the Base Hyperprior model trained with
λ = 0.0025.

S4 Additional Results
On Kodak, we achieve the following BD rate savings: 17% for SGA+BB and 15% for SGA relative to
Base Hyperprior; 17% for SGA+BB and 15% for SGA relative to BPG. On Tecnick, we achieve the
following BD rate savings: 20% for SGA+BB and 19% for SGA relative to Base Hyperprior; 22%
for SGA+BB and 21% for SGA relative to BPG.
Our experiments were conducted on a Titan RTX GPU with 24GB RAM; we observed about a 100×
slowdown from our proposed iterative inference methods (compared to standard encoder network
prediction), similar to what has been reported in [Campos et al., 2019]. Note that our proposed
standalone variant [M1] with SGA changes only compression and does not affect decompression
speed (which is more relevant, e.g., for images on a website with many visitors).
Below we provide an additional qualitative comparison on the Kodak dataset, and report detailed
rate-distortion performance on the Tecnick [Asuni and Giachetti, 2014] dataset.

3
(a) Original image (b) BPG 4:4:4; (c) Proposed (SGA); (d) [Minnen et al., 2018];
(kodim19 from Kodak) 0.143 BPP (PSNR = 29.3) 0.142 BPP (PSNR = 29.8) 0.130 BPP (PSNR = 28.5)
Figure S4: Qualitative comparison of lossy compression performance on an image from the Kodak
dataset. Figures in the bottom row focus on the same cropped region of images in the top row. Our
method (c; [M1] in Table 1) significantly boosts the visual quality of the Base Hyperprior method (d;
[M3] in Table 1) at similar bit rates. Unlike the learning-based methods (subfigures c,d), the classical
codec BPG (subfigure b) introduces blocking and geometric artifacts near the rooftop, and ringing
artifacts around the contour of the lighthouse.

4
 40

38

36

[M2] SGA+BB
PSNR

34 (proposed, bits-back)
[M1] SGA
(proposed, standalone)
32 [M3] Minnen (2018)
Base Hyperprior
[M6] Ballé (2018)
30 Hyperprior Scale-Only
[M8] BPG 4:4:4
[M9] JPEG2K
28
0.2 0.4 0.6 0.8 1.0
bits per pixel (BPP)
Figure S5: Rate-distortion performance comparisons on the Tecnick [Asuni and Giachetti, 2014]
dataset against existing baselines. Image quality measured in Peak Signal-to-Noise Ratio (PSNR) in
RGB; higher values are better.

20
BD rate saving (%) over [M8] BPG 4:4:4

10

[M2] SGA+BB [M6] Ballé (2018)

−10 (proposed, bits-back) Hyperprior Scale-Only
[M1] SGA [M8] BPG 4:4:4
(proposed, standalone) [M9] JPEG2K
[M3] Minnen (2018)
−20 Base Hyperprior

30 32 34 36 38
PSNR
Figure S6: BD rate savings (%) relative to BPG as a function of PSNR, computed from R-D curves
(Figure S5) on Tecnick. Higher values are better.