wallpaper: upscale a too-small staged image in place

A staged wallpaper smaller than its screen (yellow or red border) gets an
"Upscale 4x" button. It rewrites the file with Real-ESRGAN through
stable-diffusion.cpp's `sd-cli -M upscale`, keeping the name, so wallp's
state and the engine's lists, which name it by path, still find it.
sd-cli writes to a dot-named sibling the scan skips, and only a
successful run is moved over the original.

Success needs no [E] line as well as exit 0: with a model that fails to
load, sd-cli logs "[E] new_upscaler_ctx failed", saves the input
unscaled and exits 0 (measured).

The button takes two clicks so the model choice shows before anything is
rewritten. The first runs Imajev-4B, a small Jev-style vision decision
model, through llama-mtmd-cli on a 512 px thumbnail, with a grammar that
allows only A (anime) or B (photo or render); greedy decoding then
returns the higher-logit option, about 1.7 s. The label becomes
"Upscale 4x (anime)" or "(photo)", and the second click upscales with
the anime or the general model. Anything but A, including a failed
decider, picks the general model, safe on any image.

Imajev-4B was chosen by measurement on 22 wallpapers: all 12 anime and 4
photos right, 4 of 6 painted 3D renders called anime (those come out
smoother than painted). The 2B got 17 of 22, leaning to photo; a
three-way prompt adding "painted or render" dropped both to 15 or
fewer, with most anime called painted.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 30 insertions, 0 deletions

diff --git a/desktop/modules/appearance/README.md b/desktop/modules/appearance/README.md
index f933ac0..59868b4 100644
--- a/desktop/modules/appearance/README.md
+++ b/desktop/modules/appearance/README.md
@@ -111,6 +111,36 @@ and closes the settings window, as a successful Apply does, so the desk
 shows. It goes through `gtk-launch` and the app's desktop
 entry, so `~/bin` need not be on quickshell's PATH.
 
+When the image staged for the target screen is too small for it (the yellow
+or red border), "Upscale 4×" appears beside it. It runs Real-ESRGAN through
+stable-diffusion.cpp's `sd-cli -M upscale` and replaces the file in place,
+keeping its name, so wallp's state and the engine's lists still find it;
+Apply then sets the larger copy. Only files under `~/Pictures/wallpapers` are
+offered.
+
+Which Real-ESRGAN model is decided per image by Imajev-4B, a small Jev-style
+vision decision model, run once through `llama-mtmd-cli` on a 512 px
+thumbnail with a grammar that allows only "A" (anime) or "B" (photo or
+render). Greedy decoding then returns the higher-logit option, which is the
+decision; about 1.7 s. Anime goes to the anime model, which keeps line art
+clean; anything else, and any failure of the decider, to the general one.
+Measured on 22 wallpapers, it got all 12 anime and all 4 photos right and
+called 4 of 6 painted 3D renders anime, which then come out smoother than
+painted. The 2B got 17 of 22, and a three-way prompt adding "painted or
+render" dropped both to 15 or fewer, with most anime called painted.
+
+The button takes two clicks so the choice shows before anything is
+rewritten: "Upscale 4×: pick model" runs the decider and becomes
+"Upscale 4× (anime)" or "(photo)", and that click upscales. The decision is
+kept for the image it was made for, so staging another and coming back
+still shows it.
+
+The models are read from `/data/LLM-models/SD/upscalers/` and
+`/data/LLM-models/SD/imajev/` (`Wallpapers.upscalers`, `Wallpapers.decider`).
+Deciding takes about 2 seconds, upscaling a 1280x720 image about 7. sd-cli exits 0 even when the
+model fails to load, saving the input unscaled, so an `[E]` line in its log
+counts as a failure too, and a failure leaves the original untouched.
+
 `wallp` does the actual work, including running `udt-accent`, so the accent
 follows the new wallpaper exactly as it does from a terminal. This page
 replaces its qarma file dialog, not the script.