| license: apache-2.0 | |
| tags: | |
| - diffsynth | |
| - controlnet | |
| - fp8 | |
| - cyrillic | |
| - benchmark | |
| # Qwen-Image Cyrillic FP8 reproducibility configuration | |
| Configuration and proof bundle for: | |
| - `777Radik/Qwen-Image-2512-FP8-E4M3FN` | |
| - `777Radik/Qwen-Image-Cyrillic-ControlNet` | |
| Files: | |
| - `training-config.json` — ControlNet training inputs, hashes, and hyperparameters | |
| - `runtime-config.json` — component repositories, precision, generation, and retry settings | |
| - `benchmark-config.json` — immutable heldout100 command and quality gates | |
| - `benchmark-results.json` — BF16/FP8 metrics and manual audit result | |
| - `quantization-manifest.json` — per-shard source/output hashes | |
| - `controlnet_benchmark.py` — exact generation harness | |
| - `controlnet_accuracy.py` — OCR/glyph diagnostic and contact-sheet generator | |
| - `verify_hf_release.py` — remote LFS size/SHA-256 and required-file verifier | |
| The accepted production result is 99/100 manual exact Cyrillic on the fixed benchmark. | |
| FP8 regression is zero: all 100 quantized outputs are byte-for-byte identical to the | |
| source-storage run. | |