CMU researchers release T-DIG, an open framework for text-conditioned diversity in image generation
simi_97k · x · 2026-10-28
Researchers from CMU and Northeastern introduced T-DIG (Text-Conditioned Diversity for Image Generation), tackling the classic problem where text-to-image models return near-identical images for one prompt.
Key details:
- An LLM policy produces K axis-conditioned prompts or edit instructions for a frozen generator or editor, so it works with open-weight, distilled, and proprietary models without retraining the image model
- The policy is trained with GRPO using a text-conditioned diversity reward plus alignment penalties
- Users enter a prompt, inspect the model's default i.i.d. set, get suggested variation axes (dog breed, lighting, background, camera angle, etc.) or type their own, set a diversity strength, then generate a new set or iteratively edit the current one
- They also release the first dataset for text-conditioned prompt diversity, with human-written training axes and a test set from real user interactions, evaluated with TPIPS
Related event: CMU Open-Sources T-DIG: Text-Conditioned Diversity for Image Generation(7 posts)→
More from Multimodal
- AI band Techhalla reimagines Slipknot's 'Wait and Bleed' as a 1971 soul performance — Mother-Pear-2392 · 2026-10-28
- Inside the Hailuo H3 Max retro-MV workflow: 15-second chunks, one reference image, multi-shot prompts — techhalla · 2026-10-28
- Open-source pipeline turns Seedance AI video into real PBR textures on your 3D mesh — OdinLovis · 2026-10-28
- AI-generated cyberpunk short 'VHS Sirens' shows off consistent characters and neon city detail — brythehuman · 2026-10-28
- First-time creator makes full animated music video with 40 clips via MiniMax H3 — Capable-Insurance-58 · 2026-10-28
- Midjourney sref combo 138606468+1306896884 yields textured flat illustration — sergeantsref · 2026-10-28