PoolDINO cuts RAE image generation tokens 4-16x, runs on an M1 Pro CPU
francoisfleuret · x · 2026-10-09
- PoolDINO pools tokens in RAE-style image generation, using 4-16x fewer tokens with similar generation quality.
- Both training and sampling get cheaper; a demo compares RAE vs. 8x compression.
- Demo reportedly runs on an M1 Pro CPU with no accelerator; details in the linked thread.
More from Multimodal
- Qwen's two voice tracks: rent Qwen-Audio-3.1 or self-host TTS/ASR under Apache-2.0 — lmoroney · 2026-10-10
- Drama launches AI model to make acting editable and kill reshoots — iamfakhrealam · 2026-10-10
- HeyGen announces #1 TTS model, offers API at 50% off through 10/31 — HeyGen · 2026-10-10
- HeyGen's new voice model tops blind TTS leaderboard, API price cut 50% to $15/M chars — HeyGen · 2026-10-10
- Self-built image model Latent Blend launches, using pure embedding interpolation for unique images — graycrawford · 2026-10-10
- HeyGen Voice tops Artificial Analysis TTS Arena with Elo 1201, beats Qwen and ElevenLabs — ArtificialAnlys · 2026-10-10