A controlled study · 4 × RTX PRO 6000
Prepare a studyMake beauty, deliberately.
Keep the subject. Change one craft decision. Compare the actual images, then choose what deserves to continue.
01 / FLUXRender one candidateSame seed, fixed subject and directions. One generation at a time.
02 / GEMMALook at both imagesActual PNGs in a blind A/B and B/A comparison.
03 / ORNITHPropose a craft changeText advice grounded in Gemma’s visual observations.
04 / YOUChoose the championYour choice is final. Every original and candidate is retained.
Your study
This prepares a plan. Generation and jury connections still need verification.
Prepare the runtime connections
- FLUX TP4 · adapter requiredThe existing FLUX worker has no TP4 flag. Confirm the checkpoint and verify a compatible multi-GPU adapter.
- Gemma · download and vision checkRecord the 12B model and assistant artifacts, then verify actual-image comparisons.
- Ornith · verify current launchCheck TP4, MTP3, concurrency 4, 256K contexts, 10 GiB KV/card, tools on and thinking off.
- Memory · measured peaksRead a fresh four-GPU snapshot and profile co-resident peaks. Planning arithmetic alone does not establish a fit.
- Controlled comparisonsUse the same seed. Reverse blind presentation order; scores stay advisory.
- Preserve and publishAppend receipts and frame hashes. Store every candidate; publish only your selected frames.