1,000 liminal-space images from GPT Image 2 released for Stable Diffusion fine-tuning
A community member generated a thousand dreamcore and liminal-space images using OpenAI's GPT Image 2 and released the dataset on HuggingFace for Stable Diffusion fine-tuning.
A user has released a thousand liminal-space and dreamcore images generated with OpenAI's GPT Image 2, packaged as a training dataset on HuggingFace. The collection — empty indoor pools, foggy parking lots, unsettling corridors — targets the aesthetic niche that's become a staple of Stable Diffusion fine-tunes but remains underrepresented in public datasets.
The creator ran GPT Image 2 at 2K medium resolution and curated outputs that hit the liminal feel consistently. Instead of archiving the batch locally, they uploaded the set to HuggingFace for anyone training or experimenting with the style.
What stands out
- 01Aesthetic consistency across a thousand samples. The entire set holds to the liminal / dreamcore look — no random genre drift, which makes it cleaner for fine-tuning than scraping Midjourney grids or Pinterest boards.
- 022K medium resolution from a closed model. GPT Image 2 outputs are higher-fidelity than most open diffusion checkpoints at base, so the dataset could serve as a quality ceiling for open-weight liminal LoRAs.
- 03Ready for SD fine-tuning pipelines. The HuggingFace card includes the full image set; practitioners can drop it into Dreambooth, LoRA, or textual-inversion workflows without additional curation.
- 04A test case for synthetic training data. Closed-model outputs have been used to bootstrap open-weight image models before, but liminal spaces are compositionally tricky — empty rooms with specific lighting and perspective cues. If a LoRA trained on this set generalizes well, it's a signal that GPT Image 2's spatial understanding transfers.