A Developer Cloned DALL-E 2's Dying Aesthetic Into an Open 0.9B Model
A developer scraped 50,000 DALL-E 2 outputs before OpenAI pulled the plug, then fine-tuned Stable Diffusion 1.5 to preserve its signature look.
PRO- snwy released SD1.5-DALLE-2, a Stable Diffusion 1.5 fine-tune that mimics the DALL-E 2 aesthetic.
- Trained on roughly 50,000 images generated from DALL-E 2 before OpenAI shut it down.
- OpenAI retired DALL-E 2 and DALL-E 3 API endpoints, pushing users to gpt-image-1.5 and gpt-image-1-mini.
- Ships as a 0.9B parameter safetensors checkpoint compatible with the standard StableDiffusionPipeline.
- Author argues newer image models are too realism-focused and lost DALL-E 2's distinctive charm.
- Establishes a playbook for preserving proprietary model aesthetics through output-based distillation into open checkpoints.
DALL-E 2 had a look. Slightly painterly, slightly wrong, with faces that felt hand-drawn even when they were photorealistic. That aesthetic died quietly when OpenAI shut down its API endpoints, but a developer going by snwy managed to bottle it up first.
The project, called SD1.5-DALLE-2, is a full fine-tune of Stable Diffusion 1.5 on roughly 50,000 images generated from DALL-E 2 using a diverse prompt corpus, captured in the weeks before the model went dark. The author claims the outputs feel nearly identical to the original. It is now live on Hugging Face as a 0.9B parameter checkpoint that plugs directly into the StableDiffusionPipeline.
Why clone a two-generation-old model
Why replicate something even OpenAI stopped believing in? OpenAI removed both DALL-E 2 and DALL-E 3 from the API, pointing developers to gpt-image-1.5 and gpt-image-1-mini, which are sharper on text rendering, photorealism, and complex prompts.
Sharper is not more interesting. snwy argues that DALL-E 2 outputs had a distinct quality no other image model has, and one that is hard to reproduce by hand. Newer models, in the author's framing, are too realism-maxxed and have lost that charm. If you have ever generated a swirly, dreamlike image from 2022 and then tried to recreate it in FLUX or Imagen, you know what they mean. The seams do not match.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
