Teaching diffusion models the art of the slow fade

What it does DiffMorpher takes two images and corresponding text prompts, then produces a sequence of intermediate frames that transition from one subject to the other. It is built on top of the diffusers library and uses Stable Diffusion v2.1-base as its backbone. The result is a frame sequence—by default 50 frames—that sits somewhere between image editing and video generation.

The interesting bit Instead of blending pixels in image space, the project manipulates the model’s internal representations through LoRA interpolation, AdaIN, and a tunable self-attention replacement governed by a hyperparameter lamb. The authors also released MorphBench, noting that evaluating general object morphing had no prior benchmark.

Key highlights

• Generates frame sequences from two still images using Stable Diffusion v2.1-base.
• Interpolates LoRA weights and replaces self-attention features via a tunable lamb parameter.
• Ships with a Gradio UI and hosted Hugging Face/OpenXLab demos.
• Introduces MorphBench, a benchmark dataset for assessing image morphing of general objects.
• Supports AdaIN and reschedule sampling, exposed as runtime flags.

Caveats

• The README is essentially a quickstart manual; the theory behind the flags is absent without reading the paper.
• It is largely glue code atop the diffusers library, as the authors acknowledge.
• MorphBench is offered as a raw download with no description of its structure or evaluation metrics in the repository.

Verdict Worth a look if you are experimenting with diffusion-based video generation or need controllable transitions between two subjects. Skip it if you want a lightweight, CPU-friendly morphing tool without model dependencies.