Datasets:
File size: 2,820 Bytes
24e04bb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 | # Feasibility findings — Hindi text LoRA on SD1.5
Measured 2026-09-28, CPU-only, on this machine. Scripts in `src/diag/`.
## 1. Shaping (`verify_shaping.py`) — SOLVED
Pillow's wheel ships **without Raqm/HarfBuzz**, so `ImageDraw.text()` renders
Devanagari unshaped (broken conjuncts, ि on the wrong side). Silent corruption:
it looks plausible unless you check.
Fix: `src/render/shaper.py` = uharfbuzz (shaping) + freetype-py (raster).
Verified: क्ष/त्र/ज्ञ collapse 3 cp -> 1 glyph; ि reorders before its consonant;
स्त्री -> 2 glyphs. Continuous shirorekha confirmed visually.
## 2. VAE capacity (`vae_size_sweep.py`, `vae_visual.py`) — HARD FLOOR AT ~32px EM
SD1.5's 4-ch VAE round-trip, measured inside the ink box:
| font em size | result |
|---|---|
| 12-20px | destroyed; nuktas gone, conjuncts illegible |
| 24px | marginal; nukta merges into stroke |
| **32px** | **clean — minimum safe size** |
| 48-64px | near-perfect |
=> Every training sample must render text at **font em size >= 32px**
(FreeType set_pixel_sizes). NOT ink-box height: the measured ink/em ratio
varies 0.67 (bhaarat, no tall matras) to 1.08 (hindii) depending on which
matras a word carries, so ink height is content-dependent and a bad assert
target. Em size controls stroke width and is what the sweep actually varied.
=> At 512 canvas this caps usable text at roughly 2-4 words.
=> SD1.5 can do headline/signage text. It cannot do small body text. Ever.
## 3. Text conditioning (`tokenizer_test.py`) — THE REAL BOTTLENECK
CLIP ViT-L/14 tokenizer vs Devanagari:
- **0 of 49,408 vocab entries contain any Devanagari codepoint.**
- Falls back to raw UTF-8 byte BPE: ~1-2 tokens per character.
- Character identity IS preserved (decode round-trips), so the signal is not
information-theoretically lost — but CLIP never trained on these tokens, so
the *embeddings* are meaningless.
- 77-token cap: MEASURED ~1.1 tokens/char even for conjunct-heavy strings
(kshatriya 8ch->9tok; pradhanmantri karyalaya 21ch->23tok). BPE merges the
3-byte UTF-8 sequences efficiently -- there is no conjunct blowup. With a
12-token caption template that leaves ~65 tokens = ~55-60 chars.
=> Must train the **text encoder** alongside the UNet. A UNet-only LoRA cannot
work: the conditioning vectors it would read are noise.
=> The token cap is NOT the binding constraint. At em 32px (~17.6px/char
advance) only ~27 chars fit across a 512px line, so PIXELS bind first.
Enforce length in tokens anyway (exact, template included), but scope is
set by the VAE floor.
## Verdict
Viable, scope bounded by the VAE floor: short, large Devanagari text
(~1-4 words, <=~27 chars per 512px line). Not viable: small or dense text.
|