SVD-compressed AdaLN shrinks a MageFlow 4B finetune to a 2.8B SDXL-class T2I model
Turbulent-Bass-649 · reddit · 2026-09-18
An indie dev updates MageTrail: a full finetune of Microsoft's MageFlow 4B text-to-image model on Danbooru/E621-style data, using a diversity-maximized condensed 41k image dataset to inject booru tag prompting and illustration skills without the full booru dataset (potentially $20k-50k+).
The update approximates the model's AdaLN layers with rank-256 SVD, cutting the architecture to 2.8B — same class as SDXL — with slightly faster generation, much lower VRAM for inference and training, and minimal quality loss. Weights are on Hugging Face and Civitai; an open-source trainer (bluvoll/mage-flow-trainer, SDNQ support) enables consumer LoRA finetuning. V0.3 is now training on the 2.8B arch on a single H100.
More from Multimodal
- Free course covers the 4 hard problems of production generative media agents — Al_Grigor · 2026-11-04
- PixelBench: testing multimodal models pixel-by-pixel instead of with QA — bdsqlsz · 2026-11-04
- Kitten-TTS 2 offers free voice cloning on laptops, no GPU needed — solyarisoftware · 2026-11-04
- "We'll Take Your RAM": a fully local-hardware cyberpunk meme from 2030 San Francisco — Wonderful-Yak-3348 · 2026-11-04
- Openfilm: open-source AI video agent turns coding agents into video makers — solyarisoftware · 2026-11-04
- Microsoft reopens VibeVoice: 10-second voice cloning, 90-minute local TTS — solyarisoftware · 2026-11-04