Visual Anagrams recreated with Iris 3B Turbo: 1024×1024 illusions on a 12GB RTX 3060
MustBeSomethingThere · reddit · 2026-10-24
The author adapted Visual Anagrams—illusion images that reveal different content under rotation, flips, color inversion, or patch rearrangement—from the original pixel-space DeepFloyd IF to Iris 3B.
Using INT8 weights and the community Turbo r2 LoRA, examples were generated locally at 1024×1024 in four denoising steps on a 12GB RTX 3060. Results are experimental and some prompt pairs work better than others. Code is open-sourced on GitHub.
More from Multimodal
- Dreamina offers Seedance 2.0/2.5 Basic plan for just $0.29 first month (98% off) — thetripathi58 · 2026-10-24
- AI-generated sci-fi villain teaser "SEO-A" shows off cinematic black-gold aesthetics — Foreign-Original124 · 2026-10-24
- Goth house music video made with Freebeat AI and Suno AI v4 — PassengerShoddy2481 · 2026-10-24
- Image generator's NSFW filter keeps blocking SFW prompts yet passes borderline ones — Nomy33 · 2026-10-24
- Kling AI demo shows lifelike female singer performing in a KTV lounge — Sure_Farmer5321 · 2026-10-24
- Creator hooks ChatGPT to Runway via MCP to drive Seedance 2.5 cinematic clips — aziz4ai · 2026-10-24