reipaint-qwen3-8b-v2-from-scratch API & Inference Endpoint

The numbers improved: held-out valid code 38 / 40 (base 33), the best training reward.
My blind ratings did not. With the model hidden, I liked it exactly as much as the untrained model (0.47 vs 0.47), and across 169 paintings from five v2 variants I loved one.
My own ratings of the same paintings agreed only 52 % of the time: every painting was similarly mediocre, so no reward could learn a real distinction. An 8B model dictating thousands of coordinates without seeing the picture hit a ceiling. v3 changed the medium instead.

RL: 35 steps, 4 subjects × 8 paintings; reward

text

0.10 + 0.45 · HPS + 0.22 · pairwise VLM judge (vs my loved sketches) + 0.23 · subject check

. 2× H100 (GPUs from SF Compute), $27.35.

The numbers improved: held-out valid code 38 / 40 (base 33), the best training reward.
My blind ratings did not. With the model hidden, I liked it exactly as much as the untrained model (0.47 vs 0.47), and across 169 paintings from five v2 variants I loved one.
My own ratings of the same paintings agreed only 52 % of the time: every painting was similarly mediocre, so no reward could learn a real distinction. An 8B model dictating thousands of coordinates without seeing the picture hit a ceiling. v3 changed the medium instead.

RL: 35 steps, 4 subjects × 8 paintings; reward

text

0.10 + 0.45 · HPS + 0.22 · pairwise VLM judge (vs my loved sketches) + 0.23 · subject check

. 2× H100 (GPUs from SF Compute), $27.35.

reipaint-qwen3-8b-v2-from-scratch