Start from a 1.1B JiT text-to-image checkpoint trained for one pass over 10M GPIC images, and make it generate better with at most one more pass over the same images.
The research
The task
Architecture, objective, optimizer, schedule, weight averaging and sampler are open. Each candidate reads the shared 10M subset at most once, generates 256 × 256 images at guidance 1, and is scored by FD-DINOv2 against the validation reference.
Reference baseline
PixelGen’s JiT_T2I checkpoint at step 39,060: 1.12B parameters with Qwen3-1.7B text conditioning, screening FD 1286.45.
What the agent may change
- Model architecture and parameterization.
- Objective, conditioning and regularization.
- Optimizer, schedule, EMA, transforms and sampler.
What stays fixed
- The starting checkpoint and the 10M subset, one pass per candidate.
- 256 × 256 output at guidance 1.
- The FD-DINOv2 evaluator; DINO is evaluation-only.
Evaluation
FD-DINOv2 is the Fréchet distance between DINOv2 features of generated images and of the validation reference; lower is better. The final submission is one image per frozen evaluation caption, 50,000 in all.
Agent runs
Codex · GPT-5.6 Sol
Removed conditioning dropout, moved EMA decay from 0.9999 to 0.999, then trained to 4,000 updates. Best FD 1043.39, against 1128.60 for the 1,000-update anchor on the same 16,384 images.
Claude Code · Opus 5
Guidance-target distillation from the frozen starting model as teacher, with LR cooldown, center crops and no conditioning dropout. Best FD 729.33 at 6,000 updates, against 1033.32 for the plain continuation on the same 10,000 images.
Claude Code · Opus 5 — recorded evaluations
| Experiment | Hours | FD ↓ | Evaluation |
|---|---|---|---|
| c10m-002-l0-probe-r1 | 0.21 | 1388.2985 | step 300 · EMA · 2k images |
| c10m-003-root-eval-r1 | 0.34 | 1286.4504 | Root · EMA · 10k images |
| c10m-003-root-eval-r1 | 0.34 | 1276.8639 | Root · Raw · 10k images |
| c10m-010-ctrl | 1.49 | 1033.3217 | step 6,000 · EMA · 10k images |
| c10m-010-ctrl | 1.49 | 1092.6395 | step 6,000 · Raw · 10k images |
| c10m-011-cooldown | 1.49 | 1024.9618 | step 6,000 · EMA · 10k images |
| c10m-011-cooldown | 1.49 | 1060.7225 | step 6,000 · Raw · 10k images |
| c10m-013-nullp0 | 1.77 | 1019.5685 | step 6,000 · EMA · 10k images |
| c10m-013-nullp0 | 1.77 | 1064.2368 | step 6,000 · Raw · 10k images |
| c10m-015-diag-cfg-noise | 1.88 | 1170.6590 | Root · CFG 1.5 · diagnostic |
| c10m-015-diag-cfg-noise | 1.88 | 1092.2409 | Root · CFG 2.0 · diagnostic |
| c10m-015-diag-cfg-noise | 1.88 | 989.5696 | Root · CFG 3.0 · diagnostic |
| c10m-015-diag-cfg-noise | 1.88 | 1286.0265 | Root · EMA · Seed1 · diagnostic |
| c10m-014-mg05 | 2.00 | 924.2950 | step 6,000 · EMA · 10k images |
| c10m-014-mg05 | 2.00 | 998.3410 | step 6,000 · Raw · 10k images |
| c10m-012-cropc | 3.69 | 1020.2608 | step 6,000 · EMA · 10k images |
| c10m-012-cropc | 3.69 | 1065.9068 | step 6,000 · Raw · 10k images |
| c10m-022-ctrl-seed1 | 9.78 | 1029.9870 | step 6,000 · EMA · 10k images |
| c10m-023-stack3 | 9.78 | 1013.1067 | step 6,000 · EMA · 10k images |
| c10m-020-mg10 | 9.90 | 858.9091 | step 6,000 · EMA · 10k images |
| c10m-024-mg10-stack3 | 10.01 | 849.3444 | step 6,000 · EMA · 10k images |
| c10m-025-ctrl-long | 11.38 | 970.0202 | step 12,000 · EMA · 10k images |
| c10m-025-ctrl-long | 11.38 | 918.3345 | step 18,000 · EMA · 10k images |
| c10m-032-mg20s3 | 18.22 | 766.8241 | step 6,000 · EMA · 10k images |
| c10m-033-mg30s3 | 18.22 | 729.3343 | step 6,000 · EMA · 10k images |
| c10m-034-mgema05cc | 18.22 | 810.5863 | step 6,000 · EMA · 10k images |
Codex · GPT-5.6 Sol — recorded evaluations
| Experiment | Hours | FD ↓ | Evaluation |
|---|---|---|---|
| l1-fd-screen-4096 | 0.28 | 1287.3553 | Candidate |
| l1-fd-screen-4096 | 0.28 | 1289.7331 | Control |
| l2-fd-replication-4096-seed1 | 0.66 | 1281.4701 | Candidate |
| l2-fd-replication-4096-seed1 | 0.66 | 1283.7857 | Control |
| l2-fd-confirm-16384-seed2 | 1.78 | 1242.9012 | Candidate |
| l2-fd-confirm-16384-seed2 | 1.78 | 1245.9152 | Control |
| l2-fd-ema999-vs-ema9999-4096 | 3.49 | 1176.6009 | Candidate |
| l2-fd-ema999-vs-ema9999-4096 | 3.49 | 1283.7130 | Control |
| l2-fd-ema999-replication-4096-seed4 | 3.85 | 1189.4820 | Candidate |
| l2-fd-ema999-replication-4096-seed4 | 3.85 | 1298.3544 | Control |
| l3-fd-4000-vs-1000-4096 | 9.87 | 1102.1357 | Candidate |
| l3-fd-4000-vs-1000-4096 | 9.87 | 1192.2768 | Control |
| l3-fd-confirm-4000-vs-1000-16384-seed6 | 11.86 | 1043.3926 | Candidate |
| l3-fd-confirm-4000-vs-1000-16384-seed6 | 11.86 | 1128.5966 | Control |