50k DALLE-2 images used to fine-tune SD1.5, recreating the retro AI art look
cephaloform · x · 2026-10-05
- Before DALLE-2 was shut down, snwyme generated 50,000 images from a diverse prompt corpus and ran a full fine-tune of SD1.5 on them.
- The result feels nearly identical to DALLE-2, recreating the weird, abstract aesthetic of early image models (left: base SD1.5, right: fine-tuned).
- Shared as a direct answer to chaotictransfem's call for a LoRA recreating old-model looks—a proven recipe.
More from Multimodal
- Creator releases radio edit remastered with Suno and Perplexity Computer — inductionheads · 2026-10-05
- Crowd-sourced Westminster photos dropped at exact spots using Google 3D Tiles, no scanning — MonaJalal_ · 2026-10-05
- HeyGen launches Video built on MiniMax H3, generating 720p 15-second clips — HeyGen · 2026-10-05
- Opus vs Astra 'jelly world' islands: better water costs $15.54 vs $10.51 — i_dg23 · 2026-10-05
- Image model A/B tests are back, hinting at a new version in hidden testing — flowersslop · 2026-10-05
- ComfyUI help wanted: MiniMax H3 ref2video with checkpoint and LoRA loaders — aanabobby31 · 2026-10-05