Competitive intel

What's working, what's open.

Published winner code, open model-type lanes, and a tracker of known techniques — cross-referenced against miners observed using them.

Published code index

RepoTournamentPositionHotkeyTechniquesDiff vs G.O.D
god-image-tourn-20260416-1sttourn_21bd66f314a8ee19_20260416🥇 1st5GU4Xkd3
TIMESTEP_WEIGHT_BIASauto_captionProdigy
+312 −89 vs canonical G.O.D
god-image-tourn-20260416-2ndtourn_21bd66f314a8ee19_20260416🥈 2nd5FpdSckw
multi-resolutionEMAProdigy
+187 −42 vs canonical G.O.D
god-image-tourn-20260409-1sttourn_8c2f91a33bea4411_20260409🥇 1st5GU4Xkd3
TIMESTEP_WEIGHT_BIASauto_caption
+298 −81 vs canonical G.O.D
god-image-tourn-20260402-1sttourn_4bcd77ae902fa823_20260402🥇 1st5DtQ88WH
TIMESTEP_WEIGHT_BIASProdigygradient_accumulation_8
+405 −115 vs canonical G.O.D

Open lane finder

SDXLEXPLOITED
SDXL

Heavy eval-targeting; marginal gains only.

FluxEXPLOITED
Flux

Fluxunchained variants dominate.

Z-imageOPEN
Z-image

Only 3 tasks in Q1; champion sampled twice. Open lane.

Qwen-ImageOPEN
Qwen-Image

Qwen rollout in boss round; no dominant technique yet.

Technique tracker

TIMESTEP_WEIGHT_BIAS
3 miners

Bias training timestep sampling towards eval-critical windows

5GU4Xkd35DtQ88WH5FpdSckw
auto_caption
2 miners

Recaption training pairs with a stronger captioner at ingest

5GU4Xkd35DtQ88WH
Prodigy
3 miners

Prodigy optimizer in lieu of AdamW

5GU4Xkd35FpdSckw5DtQ88WH
multi-resolution
1 miner

Train at 768/1024 alternating buckets

5FpdSckw
EMA
1 miner

Weighted exponential moving average of weights at checkpoint

5FpdSckw
gradient_accumulation_8
1 miner

Large-batch simulation via 8-step accumulation

5DtQ88WH