Asuka-FM: 206M dual-stream DiT pretrained on one RTX 4090 in 160h for under $55
bdsqlsz · x · 2026-11-20
A new arXiv paper, Asuka-FM, presents a system-architectural co-design framework for text-to-image pretraining on consumer hardware: a 206M-parameter dual-stream DiT pretrained from scratch on a single RTX 4090 (24GB) in 160 GPU-hours for under $55 at 512px, reaching 0.609 GenEval and 78.51 DPG-Bench — Pareto-optimal DPG at just 0.066×10²¹ FLOPs of denoiser training compute.
Key designs targeting single-GPU bottlenecks:
- AdaLN-free dual-stream DiT backbone to cut parameter count
- SPRINT sparse-dense residual fusion reducing attention FLOPs by over 50%
- Continuous aspect-normalized RoPE with shifted square crop training, guaranteeing static tensor shapes and avoiding JIT compilation graph breaks from aspect-ratio bucketing
- Mitigating the frozen-VAE throughput wall that dominates step latency
The charts show Asuka matching MiniT2I-family DPG scores with roughly an order of magnitude less compute, positioning it as an accessible foundation-modeling baseline for academic labs.
More from Multimodal
- Studio trolled everyone: real artist made poster look like AI slop for OpenAI movie — aakashgupta · 2026-11-20
- Reverb-TTS trends on Hugging Face: on-device zero-shot voice cloning TTS — software-mansion · 2026-11-20
- Dev builds low-poly/pixel-art RPG Runespell on AI platform Astrocade — andrey_kurenkov · 2026-11-20
- China now leads AI video: ByteDance Seedance and MiniMax set the frontier pace — alexvoica · 2026-11-20
- Prompt template recreates 1980s vintage anime cel-animation aesthetic — azed_ai · 2026-11-20
- AI-animated series debuts with episode on young Ramanujan asking about 0/0 — visl_creations · 2026-11-20