This paper documents a frozen engineering project on personalized Chinese handwriting. The project started from approximately 200 real handwriting images from one user, covering 197 unique Chinese characters, and was initially formulated as few-shot generation of unseen characters. A sequence of canonical-centered personalization routes repeatedly exposed the same conflict: increasing structural pressure made outputs more canonical, while increasing personalization could damage identity-defining strokes. The project was therefore reset around real-human character equivalence classes. A multi-writer CASIA candidate pool showed that a USER-compatible realization often already existed among valid human samples. The task consequently changed from synthesis to character-wise matching, followed by cross-writer composition into a virtual writer. The frozen system uses real-ink features, character-specific human population percentiles, top-20 candidate pruning, and greedy hardest-first whole-row selection. On the covered target set, all 197 USER characters had real-human candidates, and the 100-character evaluation subset was covered 100/100. Knowncharacter held-out comparisons included a row judged visually almost indistinguishable from genuine USER handwriting. A 60- episode stability audit placed every episode in a predefined A-like machine-proxy region, but these were not independent human A-level judgments. The final evidence supports stable practical B-level quality, with many outputs approaching A-level under the USER-defined criterion. The report records why generation became unnecessary for this case without claiming unrestricted or universal handwriting synthesis.
Figures & tables
Fig. 1: Conceptual journey of the frozen project. The decisive change was the candidate space, not an additional generator module.
Row
Unique
Effective
Top-1
Top-3
Top-5
writers
writers
share
share
share
Example A
32
19.04
20%
40%
56%
Example B
29
18.39
20%
40%
54%
TABLE I: Anonymized provenance of two strongly USER-like whole rows.
Fig. 2: Schematic summary of the two recurring early failure modes. A style-positive output could lose structure, while a safe output could remain canonical-looking.
Source
Writers
Samples
Characters
Writer–character cells
Use
USER
1
200
197 unique
–
references and held-out targets
CASIA offline
84
18,662
207
17,153
main real-human candidate pool
CASIA online
84
12,510
137
11,504
provenance / corpus audit
TABLE II: Data scope used in the frozen evaluation record.
Fig. 3: A small crop from the real-human candidate evidence. Multiple valid realizations of the same character differ in slant, pressure, extent, and local execution while retaining character identity.
Fig. 4: Privacy-limited anonymous writer-row crop. Whole-row USER likeness was observed even though the selected characters came from many physical writers; the displayed row labels are anonymous review labels.
Fig. 5: Privacy-limited crop of the known-character held-out real-versus-virtual sheet. Only a finite subset of columns is shown; the complete USER inventory and full row are withheld.
Fig. 6: Final frozen pipeline. All output glyphs are retrieved real-human CASIA samples; the pipeline contains no drawing or synthesis stage.
Stage
Assumption
What was tried
Failure
Lesson
Personalized generation
canonical to personalized
diffusion and adaptation
structure–style conflict
likeness and validity differ
Writer modeling
global writer identity
latent codes, tokens, adapters
machine–human mismatch
a global writer code is insufficient
Local rules
transferable geometry
local correspondences
unsafe execution
reusable local signal exists
Safe execution
constrain deformation
topology-preserving geometry
personalization suppressed
safety can erase identity
Representation search
a better representation solves it
broad feature comparison
canonical-looking outputs remain
the root assumption was wrong
Human equivalence classes
real human realizations
multi-writer samples
direct generation still weak
enter the real-human manifold
TABLE III: Public abstraction of the development path.