简体中文
EnglishWhat Projects Are Featured on Justin Pinkney's Blog?
Updated
Justin Pinkney's blog highlights five recent projects: the Trailer Faces HQ Dataset (187k high-resolution face images with varied expressions taken from movie trailers, August 2024), Experiments with Swapping Autoencoder (using a model that separates an image into style and content components, May 2024), Image Mixer Diffusion (blending images by fine-tuning an Image Variation Diffusion model, April 2024), Face Mixer Diffusion (training face-conditioned diffusion models, January 2024), and Würstchen v2 Pokémon (fine-tuning Würstchen v2 for text-to-Pokémon generation, January 2024). All of them sit at the intersection of machine learning and creative image work, which is how Pinkney describes the blog's focus.
The Five Featured Projects
| Project | Date | What it involves |
|---|---|---|
| Trailer Faces HQ Dataset | August 2024 | A dataset of 187k high-resolution face images with varied expressions, sourced from movie trailers |
| Experiments with Swapping Autoencoder | May 2024 | Exploring a model that breaks images into style and content components, and the creative uses that follow |
| Image Mixer Diffusion | April 2024 | Blending or mixing images together by fine-tuning an Image Variation Diffusion model |
| Face Mixer Diffusion | January 2024 | Experiments training face-conditioned diffusion models |
| Würstchen v2 Pokémon | January 2024 | A quick experiment fine-tuning Würstchen v2 for text-to-Pokémon generation |
What Each Project Tells You
Trailer Faces HQ Dataset
This is a resource release rather than a model experiment: a large collection of high-resolution face images with varied expressions, drawn from movie trailers. If you work on face-related generation or analysis and need diverse expressions at high resolution, this is the entry point on the blog.
Experiments with Swapping Autoencoder
The premise is a model that decomposes an image into style and content. The post covers the "fun things" that decomposition enables — useful if you are curious about image manipulation through component separation rather than end-to-end prompting.
Image Mixer Diffusion and Face Mixer Diffusion
These two posts are companion experiments in blending. Image Mixer Diffusion fine-tunes an Image Variation Diffusion model to mix images; Face Mixer Diffusion trains face-conditioned diffusion models. Pinkney frames the latter with a candid note about "the best model checkpoint I ever lost," so expect practical lessons alongside results.
Würstchen v2 Pokémon
A short fine-tuning experiment: take Würstchen v2 and adapt it for text-to-Pokémon generation. It is the most self-contained of the five if you want a small, concrete fine-tuning example.
How to Decide Where to Start
- Looking for data? Start with Trailer Faces HQ Dataset.
- Interested in image decomposition and style/content control? Read Experiments with Swapping Autoencoder.
- Want image blending or face-conditioned generation? Read Image Mixer Diffusion and Face Mixer Diffusion together.
- Want a compact fine-tuning walkthrough? Start with Würstchen v2 Pokémon.
The blog notes that it is "occasionally updated," and a "See more posts" link leads to the full archive, so the five above are the latest rather than the complete list.