Add files using upload-large-folder tool
Browse files- README.md +57 -0
- artifact-manifest.json +9 -0
- configuration.json +1 -0
- model.safetensors +3 -0
README.md
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: Qwen/Qwen-Image-2512
|
| 4 |
+
tags:
|
| 5 |
+
- diffusers
|
| 6 |
+
- diffsynth
|
| 7 |
+
- controlnet
|
| 8 |
+
- text-to-image
|
| 9 |
+
- cyrillic
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# Qwen-Image Cyrillic Blockwise ControlNet
|
| 13 |
+
|
| 14 |
+
Blockwise Canny ControlNet checkpoint fine-tuned for copying real Cyrillic glyph controls
|
| 15 |
+
into Qwen-Image output. It starts from
|
| 16 |
+
`DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Canny` and keeps the DiffSynth checkpoint
|
| 17 |
+
key layout.
|
| 18 |
+
|
| 19 |
+
## Artifact
|
| 20 |
+
|
| 21 |
+
- `model.safetensors`
|
| 22 |
+
- 2,266,838,080 bytes
|
| 23 |
+
- SHA-256: `8f7ed8860b94283d48b72859cb7ec715a6da7ae522a4a312004e635abfc8f9cb`
|
| 24 |
+
|
| 25 |
+
## Training
|
| 26 |
+
|
| 27 |
+
- base: `Qwen/Qwen-Image-2512`
|
| 28 |
+
- trainable module: full `blockwise_controlnet`
|
| 29 |
+
- 128 prepared edge-control samples at 512×512
|
| 30 |
+
- 128 steps
|
| 31 |
+
- learning rate: `1e-5`
|
| 32 |
+
- trainable precision: BF16
|
| 33 |
+
- frozen transformer, text encoder, and VAE storage/onload: FP8
|
| 34 |
+
- gradient checkpointing enabled
|
| 35 |
+
|
| 36 |
+
Exact hashes and arguments are in the separate configuration repository.
|
| 37 |
+
|
| 38 |
+
## Validation
|
| 39 |
+
|
| 40 |
+
Fixed held-out benchmark: 100 unseen Russian words, 512×512, 20 steps, seed 3000 plus sample
|
| 41 |
+
index, fitted edge control, scale 0.85.
|
| 42 |
+
|
| 43 |
+
- manual exact Cyrillic score: 99/100
|
| 44 |
+
- OCR diagnostic exact: 77/100
|
| 45 |
+
- mean OCR CER: 0.1142
|
| 46 |
+
- mean glyph similarity: 0.9564
|
| 47 |
+
|
| 48 |
+
The single manual failure was `ЕЩЁ → ЁЩЁ`. Edge scale retries 0.9, 1.0, and 1.1 and a
|
| 49 |
+
filled-control retry at 0.7 with the same seed did not correct it. OCR is diagnostic only;
|
| 50 |
+
the reported 99/100 score comes from manual inspection of all five contact sheets.
|
| 51 |
+
|
| 52 |
+
## Limitations
|
| 53 |
+
|
| 54 |
+
Training data uses a narrow synthetic typography distribution: simple backgrounds, limited
|
| 55 |
+
fonts, and isolated words. This checkpoint is intended for glyph-guided generation, not
|
| 56 |
+
unguided spelling, paragraphs, arbitrary layouts, handwriting, or guaranteed OCR-perfect
|
| 57 |
+
output.
|
artifact-manifest.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"file": "model.safetensors",
|
| 3 |
+
"bytes": 2266838080,
|
| 4 |
+
"sha256": "8f7ed8860b94283d48b72859cb7ec715a6da7ae522a4a312004e635abfc8f9cb",
|
| 5 |
+
"format": "DiffSynth Blockwise ControlNet safetensors",
|
| 6 |
+
"base_model": "Qwen/Qwen-Image-2512",
|
| 7 |
+
"initial_controlnet": "DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Canny",
|
| 8 |
+
"training_config_repo": "777Radik/Qwen-Image-Cyrillic-Config"
|
| 9 |
+
}
|
configuration.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"framework":"Pytorch","task":"text-to-image-synthesis"}
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8f7ed8860b94283d48b72859cb7ec715a6da7ae522a4a312004e635abfc8f9cb
|
| 3 |
+
size 2266838080
|