metadata
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-E4M3FN777Radik/Qwen-Image-Cyrillic-ControlNet
Files:
training-config.json— ControlNet training inputs, hashes, and hyperparametersruntime-config.json— component repositories, precision, generation, and retry settingsbenchmark-config.json— immutable heldout100 command and quality gatesbenchmark-results.json— BF16/FP8 metrics and manual audit resultquantization-manifest.json— per-shard source/output hashescontrolnet_benchmark.py— exact generation harnesscontrolnet_accuracy.py— OCR/glyph diagnostic and contact-sheet generator
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.