12B dense DiT backbone comes with nine LoRAs, right up my alley

They open-sourced a text-to-image foundation model with a 12B dense DiT backbone, paired with Qwen3-VL for text encoding — that combo is right up my alley. It’s built around creative exploration, not the type that just translates your prompt literally and calls it done, it actually pushes you toward more divergent tries.

The official release also dropped nine matching LoRAs in one go, basically laying out starter moves for common styles so you don’t have to train from scratch yourself. They claim it’s near the top of the leaderboards, I’ll reserve judgment on that, you have to pull it down and run the benchmark yourself for it to count.

What actually matters to me is that the weights are open for local deployment — a dense 12B is not exactly kind to VRAM, but being free to fine-tune it matters way more than ranking a few spots higher.

1 Like

Giving you the matching LoRAs straight up saves a lot of hassle

12B is quite the VRAM load

Leaderboard rank is just for reference, running it yourself is what counts

Being fine-tunable matters more than rank, agreed

Waiting on the benchmarks

12B 是真吃显存,量化版没出来之前先别惦记本地跑