Turbo-DDCM: Fast and Flexible Zero-Shot Diffusion-Based Image Compression

1. Problem Statement

While zero-shot diffusion-based image compression methods have achieved state-of-the-art (SOTA) rate-distortion-perception trade-offs, they suffer from prohibitive computational costs and slow inference times. Existing methods (e.g., DDCM, PSC, DiffC) often require hundreds or thousands of diffusion steps and computationally expensive iterative search processes (like Matching Pursuit) to select noise vectors. This results in round-trip compression/decompression times ranging from tens of seconds to several minutes, rendering them impractical for real-world applications. Furthermore, existing zero-shot methods often lack flexibility in handling user-defined priorities (Region-of-Interest) or controlling distortion levels (PSNR) rather than just bitrates.

2. Methodology: Turbo-DDCM

The authors propose Turbo-DDCM, a highly efficient zero-shot compression framework built upon the Denoising Diffusion Codebook Models (DDCM). The core innovation lies in replacing the slow, iterative noise selection process with a closed-form, thresholding-based optimization that allows for the simultaneous combination of many noise vectors.

A. Efficient Multi-Atom Selection (The Core Innovation)

In standard DDCM, the encoder selects a single noise vector from a reproducible Gaussian codebook at each step to minimize the reconstruction error. To improve quality, DDCM uses a Matching Pursuit (MP) strategy to select a convex combination of multiple atoms, but this is iterative and slow.

Turbo-DDCM reformulates this as a constrained sparse least-squares problem:
$$s^*_t = \arg \min_{s_t} \| C_t s_t - (x_0 - \hat{x}_{0|t}) \|_2^2 \quad \text{s.t.} \quad \|s_t\|_0 = M, \quad (s_t)_i \in V \cup \{0\}$$
Where:

• $C_t$ is the codebook of $K$ i.i.d. Gaussian atoms.
• $x_0 - \hat{x}_{0|t}$ is the residual between the target image and the denoiser's prediction.
• $M$ is the number of atoms to combine.
• $V$ is a quantization set (e.g., $\{-1, +1\}$).

Key Insight: In high-dimensional spaces, Gaussian random vectors are nearly orthogonal. Leveraging this property, the authors derive a closed-form solution using a thresholding algorithm:

1. Compute inner products between all codebook atoms and the residual.
2. Select the top $M$ atoms with the largest absolute inner products.
3. Assign coefficients based on the sign of the inner product (effectively doubling the search space by allowing negative coefficients, unlike DDCM's convex combinations).

This approach eliminates the need for iterative greedy search, reducing the complexity from $\Theta(M^2 C K d)$ (DDCM MP) to $\Theta(Kd)$, achieving orders of magnitude speedup.

B. Efficient Bitstream Protocol

The original DDCM protocol encodes the indices of selected atoms sequentially, leading to factorial redundancy when order doesn't matter (since the combination is commutative). Turbo-DDCM introduces a lexicographical encoding protocol:

• It encodes the combination of $M$ indices from $K$ atoms as a single integer (the rank of the combination), rather than an ordered sequence.
• This reduces the bitrate overhead significantly (by ~40% in typical configurations) and allows for a more efficient representation of the selected noise.

C. Flexible Variants

1. Priority-Aware Compression: The method incorporates a latent priority mask $w$. The optimization problem is modified to weight the residual by $w$, allowing the model to allocate more "bits" (better reconstruction) to user-specified regions (e.g., faces in a photo) without changing the total bitrate.
2. Distortion-Controlled Compression: Instead of targeting a fixed bitrate, the system predicts the necessary bitrate to achieve a target PSNR. It leverages a correlation between JPEG file size (a fast proxy) and the resulting PSNR of the diffusion method to dynamically adjust the bitrate per image.

3. Key Contributions

• Turbo-DDCM Algorithm: A novel zero-shot compression method that reduces the number of required diffusion steps by 92% (from ~1000 to ~30) while maintaining SOTA performance.
• Closed-Form Noise Combination: Replaces iterative Matching Pursuit with a fast, threshold-based selection of multiple noise vectors, enabling the use of large $M$ (up to 300 atoms) with negligible runtime cost.
• New Bit Protocol: A lexicographical encoding scheme that eliminates factorial redundancy, significantly improving encoding efficiency.
• Flexibility: Introduction of Priority-Aware (ROI) and Distortion-Controlled (PSNR-targeted) variants, features previously absent in zero-shot diffusion compression.
• Speed: Achieves a round-trip time of ~1.8 seconds per image (512x512) on a standard GPU, which is nearly an order of magnitude faster than the fastest existing zero-shot methods (DiffC) and 30-50x faster than DDCM.

4. Experimental Results

The method was evaluated on Kodak24 and DIV2K datasets against SOTA zero-shot methods (PSC, DDCM, DiffC) and training-based methods (StableCodec, HiFiC, BPG).

• Performance (Rate-Distortion-Perception): Turbo-DDCM achieves competitive or superior PSNR, LPIPS, and FID scores compared to other zero-shot methods. It outperforms DDCM and PSC significantly. While it slightly lags behind specialized training-based methods (like StableCodec) at very high bitrates due to the inherent limits of the SD 2.1 latent space, it excels at low-to-mid bitrates.
• Speed:
  • Turbo-DDCM: ~1.3s – 2.5s (depending on bitrate).
  • DiffC (SOTA Zero-Shot): ~2.2s – 19s (with custom CUDA kernels).
  • DDCM: ~16s – 87s.
  • PSC: >300s.
  • Turbo-DDCM is ~34x faster than DDCM and ~3x faster than DiffC at high bitrates, despite DiffC using custom hardware acceleration.
• Scalability: The runtime remains nearly constant across different bitrates, unlike DiffC where runtime increases significantly with bitrate.
• Variants: The Priority-Aware variant successfully reconstructs specific regions with high fidelity even at extremely low bitrates where other methods fail. The Distortion-Controlled variant reduces PSNR variance across images by ~40% compared to fixed-bitrate approaches.

5. Significance

Turbo-DDCM represents a major breakthrough in making zero-shot diffusion compression practical. By solving the computational bottleneck of noise selection, it bridges the gap between the high perceptual quality of diffusion models and the speed requirements of real-world applications.

• Universal Backbone: It retains the "zero-shot" advantage, allowing the same pre-trained model (e.g., Stable Diffusion) to be used for compression, restoration, and editing without retraining.
• Hardware Agnostic: It achieves its speedup through algorithmic efficiency (closed-form solutions) rather than custom hardware (CUDA kernels), making it accessible on standard GPUs.
• Future Direction: It opens the door for adaptive, user-centric compression (ROI, target PSNR) in generative AI pipelines, moving beyond static bitrate constraints.

In summary, Turbo-DDCM transforms diffusion-based image compression from a slow, experimental technique into a fast, flexible, and competitive solution suitable for practical deployment.