Feature Request

[SleepyOldOrbs]

Hi,

Thanks for sharing your ref2font program. I am having a go at using it.

In the Docs you are very clear to "..just use Aa" as the starting point.

1. What if you have other capitals of a typeface A H T Q etc... can these be used to guide the program. i.e.
2. What if you do NOT have any lowercase letters? Is it possible to use your program to generate the typeface still?

Sincerely

[SnJake]

Hi, thanks for the question!

The short answer is: No, currently you must use "Aa" (or shapes that look like them).

The technical reason:
The LoRA was trained specifically to recognize the visual structure and shapes of "A" and "a" as the trigger for style transfer. It doesn't "read" the letters; it looks for those specific shapes to calculate the style weights. If you feed it "H T" or "Q Z", the model won't recognize them as a valid reference and will likely generate noise or a standard fallback font.

If you have an All-Caps font (no lowercase):
You should try creating an input image with two "A"s (effectively "AA"). Since the visual shape is close to what the model expects, it has a much better chance of working than using random letters like H or Q.

Adding support for random letters (arbitrary reference) would require a completely different training strategy (multi-reference), which is a great idea for the future but not supported in V2.

[SleepyOldOrbs]

OK, thank you for the clarification. I shall try with two capitals and see what happens. Cheers.

…

On Sun, 8 Feb 2026, 12:34 SnJake, ***@***.***> wrote: *SnJake* left a comment (SnJake/Ref2Font#4) <#4 (comment)> Hi, thanks for the question! The short answer is: *No, currently you must use "Aa" (or shapes that look like them).* *The technical reason:* The LoRA was trained specifically to recognize the visual structure and shapes of *"A"* and *"a"* as the trigger for style transfer. It doesn't "read" the letters; it looks for those specific shapes to calculate the style weights. If you feed it "H T" or "Q Z", the model won't recognize them as a valid reference and will likely generate noise or a standard fallback font. *If you have an All-Caps font (no lowercase):* You should try creating an input image with two **"A"**s (effectively "AA"). Since the visual shape is close to what the model expects, it has a much better chance of working than using random letters like H or Q. Adding support for random letters (arbitrary reference) would require a completely different training strategy (multi-reference), which is a great idea for the future but not supported in V2. — Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ARP34J4QRKNR5XLDI7MAR6T4K4UONAVCNFSM6AAAAACUMADYG6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTQNRXGA4TSOJTHE> . You are receiving this because you authored the thread.Message ID: ***@***.***>