VFX Atlas
H 308 / 3212020s

Latent-Space Morphing

Latent space / diffusion morphing. GAN history. Open diffusion experiments, especially from 2022/2023 onwards.

Typeface · Syne

DiffMorpher, 2023/2024: Original demonstration with generated intermediate images. Temporal effect: still image or image sequence.
DiffMorpher, 2023/2024: Original demonstration with generated intermediate images. Temporal effect: still image or image sequence.Kaiwen Zhang, Yifan Zhou, Xudong Xu, Xingang Pan & Bo Dai. Source illustration of the authors; Project page CC BY-SA 4.0View source ↗

Prompt

Apply a Latent-Space Morphing effect as a single surreal intermediate frame, blending existing surfaces into related organic or architectural forms. Use continuous contours, hybrid textures, and smoothly changing materials rather than transparent crossfades, keeping defining features recognizable throughout the transformation. Keep the original subject and composition.

Paste it into LetzAI with your own image to apply the effect ↗

Identify

Faces become landscapes, architecture becomes organisms. New intermediate forms continuously change their meaning.

Principle

A model generates images along interpolated latent representations1 or parameters. Learned image structures replace simple pixel blending.

Example

Wang/Golland and DiffMorpher show different research approaches to image interpolation in 2023.

Try it out

Compare dissolve, classic morphing, and generative interpolation2 between two motifs. Look for newly invented intermediate forms.

Background

GAN prehistory in the 2010s; open diffusion experiments from 2022/2023.

It's not just finished pixels that are blended. A model generates intermediate states along interpolated latent representations, conditionings, or model parameters. The visible form can change semantically in the process.

Wang and Golland, Interpolating between Images with Diffusion Models (2023); DiffMorpher (2023) interpolates, among other things, LoRA parameters and latent noise states.

Regarding the illustration: Historical research example. Plausible new image states are created between two inputs, instead of simply blending pixels. The original research image shows portrait, architectural, and dog morphs.

Distinction

Frame interpolation fills in temporal intermediate frames of a movement. Generative image interpolation can switch between different image content. Not every AI zoom is a true latent morph. Latent interpolation predates the 2020s. In that decade, with the widespread use of image generators, it became a highly visible creative practice.

Aftermath

The transition itself became a search for images. Its intermediate forms can be more surprising than the two target images.

Further examples · 7 images

AI illustration: A chair transforms into a bonsai tree through four discrete steps
02 AI illustration: A chair transforms into a bonsai tree through four discrete stepsA chair transforms into a bonsai tree through four discrete steps. The intermediate frames show hybrid structures where wood and upholstery evolve into biological branches and foliage.Illustration generated for the Atlas (AI)
AI illustration: A cathedral morphs into a human heart across twelve frames
03 AI illustration: A cathedral morphs into a human heart across twelve framesA cathedral morphs into a human heart across twelve frames. The architectural elements gradually reconfigure into anatomical veins and chambers, demonstrating semantic reinterpretation of learned structures.Illustration generated for the Atlas (AI)
AI illustration: The panels compare cross-dissolve, geometric warping, and generative inte
04 AI illustration: The panels compare cross-dissolve, geometric warping, and generative inteThe panels compare cross-dissolve, geometric warping, and generative interpolation. The middle and right columns show a butterfly becoming a stained-glass window via novel hybrid forms.Illustration generated for the Atlas (AI)
AI illustration: A triptych displays a woman's face, an intermediate moss-covered figure,
05 AI illustration: A triptych displays a woman's face, an intermediate moss-covered figure, A triptych displays a woman's face, an intermediate moss-covered figure, and a forest. This illustrates latent-space morphing by generating a semantically meaningful hybrid form that exists between two distinct conceptual endpoints.Illustration generated for the Atlas (AI)
AI illustration: A concrete building transitions into a metallic beetle across four frames
06 AI illustration: A concrete building transitions into a metallic beetle across four framesA concrete building transitions into a metallic beetle across four frames. It illustrates latent-space morphing by generating coherent intermediate states where architectural geometry and biological anatomy structurally merge.Illustration generated for the Atlas (AI)
AI illustration: A digital artist works on a sequence morphing a bust into a flower
07 AI illustration: A digital artist works on a sequence morphing a bust into a flowerA digital artist works on a sequence morphing a bust into a flower. The screen displays a mathematical latent trajectory where one semantic form evolves into another via learned features.Illustration generated for the Atlas (AI)

Sources