---
kind: concept
slug: ddim-inversion
title: DDIM inversion
summary: >-
  Recovers the noise latent a deterministic DDIM sampler would turn into a given image, so the noise can be resampled
  under new conditions. Plug-and-Play inverts with an empty prompt and about 1000 steps; Composer uses the image's own
  conditions; StyleAligned inverts one style reference under a caption. Trade-offs: cost, and inversion can fail.
tags: [diffusion, image-editing, sampling, controllable-generation]
papers: ["2211.12572", "2302.09778", "2312.02133"]
related: [classifier-free-guidance, diffusion-models, image-to-image-translation, text-guided-image-editing, text-to-image-personalization]
status: draft
updated_at: 2026-10-18
---

## Overview

A diffusion model normally maps noise to an image. DDIM inversion runs the
deterministic DDIM sampler in the other direction: it starts from a real image,
encodes it, and walks toward noise to get a latent x_T that a sampler could map
back to (approximately) that image [arxiv:2211.12572 §4] [arxiv:2211.12572 app. C].
The useful property is that x_T is a concrete starting point. If you sample from
the same x_T with a different prompt, the output keeps some of the layout of the
original while the text changes its content [arxiv:2211.12572 §4].

Three reuse-licensed cards cover this concept so far. They differ in the
conditions used during inversion: Plug-and-Play inverts with an empty prompt,
Composer inverts with the image's own set of conditions [arxiv:2302.09778
§3.2], and StyleAligned inverts a single style reference under a caption
[arxiv:2312.02133 §4.3]. All three rely on sampling from the inverted latent of
the image that is being edited or used as a style source.

The invariant that Plug-and-Play relies on is that the translated image must
start from the inverted noise of the guidance image. With a different seed, the
injected features no longer match the encoder skip features and the method
fails [arxiv:2211.12572 app. B] [arxiv:2211.12572 fig. 14].

## Variants

### Empty-prompt inversion with 1000 steps (Plug-and-Play)

The guidance image is encoded with the autoencoder and inverted with the empty
prompt and a CFG scale of 1 [arxiv:2211.12572 §5.1]. The paper uses 1000
inversion steps and 1000 reconstruction steps for real images [arxiv:2211.12572
app. C]. The guidance trajectory is then re-run deterministically, and its
internal features are saved at the 50 timesteps used for translation
[arxiv:2211.12572 app. C]. Card: [Plug-and-Play](/papers/2211.12572).

Inverting with the empty prompt is the design choice the authors give for
why any prompt and CFG scale can be used at sampling time. They attribute the
weaker results of Prompt-to-Prompt on real images to inverting with a source
text [arxiv:2211.12572 §5.1]. Inverting with the target or source prompt is
therefore not what this variant does.

### Inversion of generated guidance images

For guidance images the model generated itself, the paper's guidance features
come from the generating prompt P_G instead of the empty prompt, with a CFG
scale of 7.5 [arxiv:2211.12572 §4] [arxiv:2211.12572 app. C]. The paper does
not state whether the latent is still obtained by DDIM inversion in this case
or taken directly from the generation run.

### Inversion under the image's own conditions (Composer)

Composer inverts an image to a latent with DDIM inversion conditioned on a set
of the image's representations, c_i. It then runs DDIM sampling from that
latent under a modified set, c_j [arxiv:2302.09778 §3.2]. The same procedure
serves reconfiguration, where one aspect of the image changes and the rest is
kept [arxiv:2302.09778 §3.2] [fig. 3]. Card: [Composer](/papers/2302.09778).

The paper states neither the number of inversion steps nor the guidance scale
used during inversion [arxiv:2302.09778 §3.2] [arxiv:2302.09778 §3.5]. The
reconfiguration results are shown only as figures, without metrics
[arxiv:2302.09778 fig. 3].

### Inversion of one style reference under a caption (StyleAligned)

StyleAligned uses a real image as the style source. It inverts that image with
DDIM inversion, conditioned on a text caption produced with BLIP, to recover a
noise trajectory from x_T to x_0. New images are then generated while the
reference follows that trajectory [arxiv:2312.02133 §4.3]. Card:
[StyleAligned](/papers/2312.02133).

The paper states no inversion step count and no inversion guidance scale
[arxiv:2312.02133 §4.3]. The official notebook uses 50 steps with guidance 2,
which the card lists in its paper-vs-code section, so these values come from the
code and not from the paper [StyleAligned card](/papers/2312.02133). The paper
says that inversion sometimes fails or gives an erroneous trajectory, and it
describes no fallback [arxiv:2312.02133 §4.3].

## When to use which

| Situation | Use | Why |
|---|---|---|
| You have a real photo, a frozen diffusion model, and need a zero-shot edit that keeps layout | DDIM inversion with an empty prompt, then resample under the new prompt [arxiv:2211.12572 §4] [arxiv:2211.12572 §5.1] | This is the setup the paper evaluates; it needs no training and lets any prompt be used at sampling time [arxiv:2211.12572 §5.1] |
| You have an image in a multi-condition model such as Composer and want to change one aspect, such as layout or palette, while keeping the rest | Invert under the image's current condition set, then sample under the modified set [arxiv:2302.09778 §3.2] [fig. 3] | The paper uses this for reconfiguration, but only shows the results as figures [arxiv:2302.09778 fig. 3] |
| You have one reference image in a style you want, need several new images in that style, and cannot fine-tune per style | Invert the reference under a caption, then generate the set with shared self-attention [arxiv:2312.02133 §4.3] [arxiv:2312.02133 §3.2] | Zero-shot. The paper reports no metric for the inverted-reference case, and its quantitative evaluation uses style phrases in the text prompt [arxiv:2312.02133 §4] |
| The latency or compute budget cannot absorb roughly 1000 extra U-Net calls per image | Plug-and-Play's schedule costs about 1000 inversion and 1000 reconstruction steps for real images [arxiv:2211.12572 app. C]; Plug-and-Play does not discuss cheaper inversion schemes | Composer and StyleAligned state no inversion step count [arxiv:2302.09778 §3.2] [arxiv:2312.02133 §4.3], so their cost cannot be estimated from the paper |
| The guidance image is a textureless shape, silhouette or segmentation mask | Expect poor results; the paper flags this case as a limitation | Inversion of textureless images can encode dominant low-frequency appearance, which then leaks into the output [arxiv:2211.12572 §6]. Masks with arbitrary colors have no semantic link to the text [arxiv:2211.12572 §6] |
| You need the output to match the source appearance, not only its layout | Not established by this paper; its results report appearance change alongside structure preservation | Negative prompting raises appearance change at a small structure cost, and the paper's metrics include a deviation-from-guidance score [arxiv:2211.12572 §5.2] [arxiv:2211.12572 §5.1] |

The Plug-and-Play paper compares its full method against SDEdit,
Prompt-to-Prompt, DiffuseIT and VQGAN-CLIP, but reports its results as plotted
curves without numeric tables for the main comparison [arxiv:2211.12572 §5.1].
Composer's comparison is on COCO text-to-image generation, not on editing
[arxiv:2302.09778 §3.5]. StyleAligned's quantitative table compares style
consistency for text prompts against personalisation and editing baselines
[arxiv:2312.02133 tab. 2]. The table above therefore reflects what the papers
say, not a benchmark of inversion alone.

## Implementation sketch

The Plug-and-Play steps, expressed as an inversion loop. Each inversion step
runs the U-Net on the current latent with the empty prompt and moves the latent
one step toward noise, using the deterministic DDIM update in reverse. The exact
update rule and its scaling are standard DDIM and are not restated in the
paper's main text.

```pseudocode
# DDIM inversion of a guidance image with an empty prompt
# [arxiv:2211.12572 §4] [arxiv:2211.12572 app. C]
# Input:  image I_G, number of inversion steps N (paper: 1000)
# Output: latent x_T used as the shared starting noise

def ddim_invert(I_G, N):
    # Encode the image into the model's latent space  [arxiv:2211.12572 §4]
    x = encode(I_G)
    # Walk from the clean latent toward noise, one step at a time  [arxiv:2211.12572 app. C]
    for t in ddim_timesteps_ascending(N):
        # Predict noise with the empty prompt, CFG scale 1  [arxiv:2211.12572 §5.1]
        eps = unet(x, t, prompt="")
        x = ddim_step_inverse(x, eps, t)
    return x  # the same x_T is the start of the translated sample  [arxiv:2211.12572 app. B]
```

Notes on where variants differ:

- **Step count.** The paper's 1000 steps are what it states; its shipped code
  uses 999 for the same schedule. This mismatch is noted on the Plug-and-Play
  card, not here. Composer states no inversion step count. StyleAligned states
  none either; its notebook uses 50 [StyleAligned card](/papers/2312.02133).
- **Conditions for inversion.** Empty prompt in the Plug-and-Play real-image
  path. Composer inverts under the image's own condition set c_i instead
  [arxiv:2302.09778 §3.2]. StyleAligned conditions on a BLIP caption
  [arxiv:2312.02133 §4.3]. A non-empty prompt changes the latent that comes out
  of inversion.
- **Sampling afterwards.** Plug-and-Play uses the same x_T with deterministic
  DDIM sampling (eta = 0) [arxiv:2211.12572 app. C]. For Composer's sampler,
  see Contradictions.

## Contradictions

The papers invert with different conditioning, and Composer's sampler is
described inconsistently. No paper tests another's choice, so this page does
not settle either point.

- **Conditioning for inversion.** Plug-and-Play inverts with an empty prompt
  and attributes weaker real-image results to inverting with a source text
  [arxiv:2211.12572 §5.1]. Composer inverts under the image's own conditions,
  one of which can be a caption [arxiv:2302.09778 §2.2] [arxiv:2302.09778 §3.2].
  StyleAligned inverts under a BLIP caption [arxiv:2312.02133 §4.3]. These are
  different setups for different tasks, not conflicting numbers.
- **Sampler after inversion.** Composer's text describes DDIM sampling from the
  inverted latent [arxiv:2302.09778 §3.2]. Its hyperparameter table caption says
  DPM-Solver++ is the sampler for all diffusion models [arxiv:2302.09778 tab. 1].
  The paper does not say which one the reconfiguration step uses.

## Papers

- [Plug-and-Play](/papers/2211.12572) (Tumanyan et al., text-driven image-to-image translation): uses DDIM inversion with an empty prompt as the first step of zero-shot, structure-preserving image editing on frozen Stable Diffusion v1.4.
- [Composer](/papers/2302.09778) (Huang et al., controllable image synthesis from composable conditions): inverts an image under its own condition set and resamples under a modified set, for reconfiguration.
- [StyleAligned](/papers/2312.02133) (Hertz et al., style-aligned image generation via shared attention): inverts one style reference under a BLIP caption, then shares attention with it to generate a style-consistent set on SDXL.