Qwen Image 2.1 — NextScene Edit

Native DiffSynth-Studio editing LoRA: scene A + original description → scene B. Fresh training on all 11,526 complete available pairs from four subsets, 1024-area, three epochs. 24 established heldout pairs stay separate.

LoRA rank/alpha 32, scaled FP8 DiT, BF16 compute, FP32 adapters, AdamW, peak LR 1e-4 with 200-step warmup and cosine decay. Real batch 1, accumulation 1. Checkpoints and samples are periodically uploaded with verified hashes.

The step-10 smoke verifies execution and learning, not final visual quality. Loading and sampling use the same native DiffSynth model and numerical policy; see evidence/code/qwen21_native.py.

Run stopped

The user cancelled training at step 316 and requested deletion of the instance workspace. The final adapter and complete optimizer/RNG state were uploaded and hash-verified before cleanup. This run did not complete three epochs; no BF16 replacement run was started.