here's a bit of fun: i've been distillation training the Qwen3 0.6B text embedder on T5-xxl outputs. This is the first coherent output i've gotten running its output through Chroma.

The image doesn't at all resemble the prompt, which was:

"Hatsune Miku, depicted in anime style, holding up a sign that reads 'Qwen3'. In the background there is an anthroporphic muscular wolf, rendered like a high-resolution 3D model, giving a thumbs up to the camera as a priapic bulge rises inside his tight shorts."

but it is cool as fuck to see something resembling a thing