Denoising Diffusion Implicit Model (DDIM)
Definition
DDIM refers to a method for generating images by stepwise denoising with a diffusion model.
Denoising Diffusion Implicit Models (DDIM) use a non-Markovian sampling process to generate images from a trained diffusion model. The original paper shows that compared to certain DDPM sampling methods, the number of required generation steps can be reduced without changing the training objective.
Simply explained
Think of an image that initially consists only of noise and becomes clearer in several steps. DDIM chooses a different sequence of these steps. The comparison explains the idea but not the mathematical relationships or the quality of each individual result.
Why it matters in practice
The choice of sampling method influences how much computational effort is required for image generation. The result also depends on the model, settings, and quality requirements.
Example from business
Possible example: A creative team tests the same image generation task with different sampling configurations and evaluates image quality and runtime separately.
Limits and pitfalls
- A lower number of steps does not guarantee the same image quality or a specific speed advantage on every hardware. DDIM is not standalone proof of the suitability of a generated motif.
Origin & history
Jiaming Song, Chenlin Meng, and Stefano Ermon introduced DDIM in the first arXiv version submitted on October 6, 2020. The paper compares the proposed sampling approach with DDPMs.
Use in mid-sized companies
- Possible uses include experimental image generation, evaluation of quality/runtime trade-offs, and selection of technical inference parameters before deployment.
Frequently asked questions
Is it mandatory to train a new model for DDIM?
No. The original paper describes the same training objective as DDPM.
Are faster results guaranteed?
No. The mentioned measurements come from the experiments studied.
Sources
- Denoising Diffusion Implicit Models · Song et al. (2020): Denoising Diffusion Implicit Models
Sources checked on 20.10.2026. Editorial use cases are hypothetical; no evidence of real customer results.
Editorially reviewed on: