| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen-Image-2512 |
| tags: |
| - diffusers |
| - diffsynth |
| - controlnet |
| - text-to-image |
| - cyrillic |
| --- |
| |
| # Qwen-Image Cyrillic Blockwise ControlNet |
|
|
| Blockwise Canny ControlNet checkpoint fine-tuned for copying real Cyrillic glyph controls |
| into Qwen-Image output. It starts from |
| `DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Canny` and keeps the DiffSynth checkpoint |
| key layout. |
|
|
| ## Artifact |
|
|
| - `model.safetensors` |
| - 2,266,838,080 bytes |
| - SHA-256: `8f7ed8860b94283d48b72859cb7ec715a6da7ae522a4a312004e635abfc8f9cb` |
|
|
| ## Training |
|
|
| - base: `Qwen/Qwen-Image-2512` |
| - trainable module: full `blockwise_controlnet` |
| - 128 prepared edge-control samples at 512×512 |
| - 128 steps |
| - learning rate: `1e-5` |
| - trainable precision: BF16 |
| - frozen transformer, text encoder, and VAE storage/onload: FP8 |
| - gradient checkpointing enabled |
|
|
| Exact hashes and arguments are in the separate configuration repository. |
|
|
| ## Validation |
|
|
| Fixed held-out benchmark: 100 unseen Russian words, 512×512, 20 steps, seed 3000 plus sample |
| index, fitted edge control, scale 0.85. |
|
|
| - manual exact Cyrillic score: 99/100 |
| - OCR diagnostic exact: 77/100 |
| - mean OCR CER: 0.1142 |
| - mean glyph similarity: 0.9564 |
|
|
| The single manual failure was `ЕЩЁ → ЁЩЁ`. Edge scale retries 0.9, 1.0, and 1.1 and a |
| filled-control retry at 0.7 with the same seed did not correct it. OCR is diagnostic only; |
| the reported 99/100 score comes from manual inspection of all five contact sheets. |
|
|
| ## Limitations |
|
|
| Training data uses a narrow synthetic typography distribution: simple backgrounds, limited |
| fonts, and isolated words. This checkpoint is intended for glyph-guided generation, not |
| unguided spelling, paragraphs, arbitrary layouts, handwriting, or guaranteed OCR-perfect |
| output. |
|
|
