Skip to content

armgit / kokon

Public Imported from GitHub
0

Edit photos by describing the change, on your own GPU. A free, open-source alternative to Nano Banana and ChatGPT image editing: photos stay on your computer, no daily limit. One installer for Windows, Mac and Linux sets up the best open model (Qwen-Image 2.1). MCP server for AI agents.

Mac: Lanczos is the default upscaler and Superscale-SP is not downloaded

Gabriele Lozupone190e4c44d982
diff --git a/README.md b/README.md
index 5acd497..b2134a9 100644
--- a/README.md
+++ b/README.md
@@ -187,7 +187,7 @@ preset later from Settings → Models → **Set up models…**
| 32 GB+ | Qwen-Image 2.1 Turbo Q4_K | Qwen3-VL 8B Q8_0 (8.7 GB) | ~16 GB | not yet |
Every preset also gets stable-diffusion.cpp (the CUDA 12 build on NVIDIA, the Vulkan build on AMD and Intel, the Metal build on a Mac), the Qwen-Image 2.1 VAE and the
-Qwen3-VL vision projector and the Superscale-SP upscaler (67 MB). The image model is Qwen's
+Qwen3-VL vision projector and, on a PC, the Superscale-SP upscaler (67 MB). The image model is Qwen's
official 8-step [Qwen-Image 2.1 Turbo](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo) through
[DogukanUrker/Qwen-Image-2.1-Turbo-GGUF](https://huggingface.co/DogukanUrker/Qwen-Image-2.1-Turbo-GGUF);
the other sources are
@@ -492,7 +492,9 @@ recent Apple systems show them in HDR, and every other viewer shows a normal JPE
<summary><b>Which upscaler is used, and can I change it?</b></summary>
<br>
-By default **Superscale-SP** by Nmkd (ESRGAN), downloaded by the setup screen. Pick another one
+By default **Superscale-SP** by Nmkd (ESRGAN), downloaded by the setup screen; on a Mac,
+**Lanczos**, because AI upscalers take minutes there (Superscale-SP: 5 min 45 s to enlarge a
+1024×1024 image to 2048×2048 on an M2 Pro, against about a second). Pick another one
in the Advanced menu next to the prompt, or in Settings → Models; **Compare upscalers** there
shows each one on the same crops (a face, a JPEG photo, a night street, text, a landscape, a 4×
upscale):
diff --git a/crates/qls-core/src/catalog.rs b/crates/qls-core/src/catalog.rs
index 055f3c8..53530e7 100644
--- a/crates/qls-core/src/catalog.rs
+++ b/crates/qls-core/src/catalog.rs
@@ -418,9 +418,14 @@ pub fn recommended_preset(vram_mib: Option<u64>) -> &'static Preset {
}
/// Everything a preset needs on `backend`, in download order (small, essential files first).
+/// Superscale-SP is the default upscaler on PCs only (on Metal it takes minutes per image).
pub fn downloads_for(p: &Preset, backend: GpuBackend) -> Vec<Download> {
let mut v = engine_downloads(backend);
- v.extend([UPSCALER, SUPERSCALE, VAE, VISION_PROJECTOR, p.text_encoder, p.diffusion]);
+ v.push(UPSCALER);
+ if backend != GpuBackend::Metal {
+ v.push(SUPERSCALE);
+ }
+ v.extend([VAE, VISION_PROJECTOR, p.text_encoder, p.diffusion]);
v
}
@@ -479,7 +484,9 @@ mod tests {
for (name, _) in crate::config::UPSCALER_CHOICES {
assert_eq!(upscaler_download(name).is_some(), *name != crate::config::LANCZOS, "{name}");
}
- assert_eq!(upscaler_download(crate::config::DEFAULT_UPSCALER).unwrap().id, "upscaler_superscale");
+ assert_eq!(upscaler_download(crate::config::UPSCALER_CHOICES[0].0).unwrap().id, "upscaler_superscale");
+ let metal = downloads_for(&PRESETS[0], GpuBackend::Metal);
+ assert!(metal.iter().all(|d| d.id != "upscaler_superscale"), "Macs default to Lanczos: no Superscale download");
for p in PRESETS {
let cuda = downloads_for(p, GpuBackend::Cuda);
assert_eq!(cuda.len(), 8);
diff --git a/crates/qls-core/src/config.rs b/crates/qls-core/src/config.rs
index 4e4a7b2..2c3bbc7 100644
--- a/crates/qls-core/src/config.rs
+++ b/crates/qls-core/src/config.rs
@@ -30,8 +30,10 @@ pub const UPSCALER_CHOICES: &[(&str, &str)] = &[
("4x-UltraSharp", "UltraSharp"),
(LANCZOS, "Lanczos (plain resize)"),
];
-/// Default model (user choice, 2026-10-12, after MoSR drew a grid on clean renders).
-pub const DEFAULT_UPSCALER: &str = UPSCALER_CHOICES[0].0;
+/// Default model (user choice, 2026-10-12, after MoSR drew a grid on clean renders): Superscale-SP
+/// on PCs, Lanczos on Macs, where ESRGAN takes minutes (1024² to 2048² on an M2 Pro: Superscale-SP
+/// 5 min 45 s, Lanczos 1 s).
+pub const DEFAULT_UPSCALER: &str = if cfg!(target_os = "macos") { LANCZOS } else { UPSCALER_CHOICES[0].0 };
/// Extensions sd-server accepts for model-backed upscalers (common.cpp, commit 3f8527a).
pub const UPSCALER_EXTENSIONS: &[&str] = &["pth", "safetensors", "gguf", "pt"];
/// Profile id of the official Qwen-Image 2.1 weights (the default since schema 5).
@@ -619,7 +621,8 @@ mod tests {
assert!(c.validate().is_ok());
assert_eq!(p.files().len(), 4);
assert_eq!(c.upscale_tile_size, 0, "automatic");
- assert_eq!((c.upscaler_model.as_str(), c.upscale_blend, c.upscale_prescale), ("4x_NMKD-Superscale-SP_178000_G", 1.0, false));
+ let default_model = if cfg!(target_os = "macos") { "lanczos" } else { "4x_NMKD-Superscale-SP_178000_G" };
+ assert_eq!((c.upscaler_model.as_str(), c.upscale_blend, c.upscale_prescale), (default_model, 1.0, false));
assert_eq!(c.upscalers_dir(), default_sdcpp_dir().join("models").join("upscalers"));
}
diff --git a/crates/qls-core/src/planner.rs b/crates/qls-core/src/planner.rs
index 7e9f907..bd2cd6a 100644
--- a/crates/qls-core/src/planner.rs
+++ b/crates/qls-core/src/planner.rs
@@ -593,7 +593,8 @@ pub fn upscale_tile(override_px: u32, vram_mib: Option<u64>, model: &str) -> u32
/// One x4 pass over `input` pixels, model load included. Measured on a 1472² edit
/// (2026-10-12), RTX 5070 Ti, CUDA with direct convolutions: MoSR 3.1 s at 768 px tiles and
/// ~5 s at 128–256 px (Vulkan within 10%), Superscale-SP (ESRGAN) 13.2 s at 128 px. M2 Pro,
-/// Metal: MoSR 54.6 s at 768 px and 78 s at 128 px; ESRGAN not measured (estimated ~3x MoSR).
+/// Metal: MoSR 54.6 s at 768 px and 78 s at 128 px; Superscale-SP 5 min 45 s for 1024² (M2 Pro,
+/// 2026-10-12: ~345 s/MP, which is why Macs default to Lanczos).
pub fn esrgan_seconds(input: u64, tile: u32, metal: bool, mosr: bool) -> f32 {
let large = tile == 0 || tile >= 512;
let per_mp = match (mosr, metal, large) {
@@ -602,7 +603,7 @@ pub fn esrgan_seconds(input: u64, tile: u32, metal: bool, mosr: bool) -> f32 {
(true, true, true) => 25.0,
(true, true, false) => 32.0,
(false, false, _) => 6.1,
- (false, true, _) => 90.0,
+ (false, true, _) => 345.0,
};
0.5 + per_mp * input as f32 / 1_048_576.0
}
diff --git a/docs/engineering-notes.md b/docs/engineering-notes.md
index 0806107..8f09fc4 100644
--- a/docs/engineering-notes.md
+++ b/docs/engineering-notes.md
@@ -121,6 +121,8 @@ resized down to the exact size; Lanczos is only used for shrinking or factors be
Default: Superscale-SP (ESRGAN, WTFPL, setup download), 128 px tiles with direct convolutions
(13.2 s / +455 MB for 1472² on an RTX 5070 Ti, vs 24.7 s / +747 MB without). Also offered:
MoSR (experimental, bundled), UltraSharp (optional download, non-commercial), Lanczos.
+Macs default to Lanczos and skip the Superscale download: on an M2 Pro (Metal) Superscale-SP
+took 5 min 45 s for 1024² → 2048² (~345 s/MP) against ~1 s for Lanczos; MoSR ~25 s/MP.
`upscaleBlend` mixes the result with Lanczos; `upscalePrescale` shrinks the input so x4 lands
~1.41x the target. Why MoSR is not the default: on clean full-size renders enlarged < ~2.8x it
invents a fine crosshatch that the final downscale turns into a visible grid (the port matches
diff --git a/src/components/Upscalers.tsx b/src/components/Upscalers.tsx
index 6a00056..69fd7b3 100644
--- a/src/components/Upscalers.tsx
+++ b/src/components/Upscalers.tsx
@@ -2,6 +2,7 @@ import { Download, Images, Loader2, X } from "lucide-react";
import { useEffect, useState } from "react";
import { api } from "../lib/api";
import { formatBytes, installProgress } from "../lib/logic";
+import { isMac } from "../lib/platform";
import { refreshStatus, setState, toastError, useStore } from "../lib/store";
import type { InstallStatus, UpscalerItem } from "../lib/types";
import { Seg } from "./ui";
@@ -9,27 +10,34 @@ import { Seg } from "./ui";
const img = import.meta.glob("../assets/upscalers/*.png", { eager: true, import: "default" }) as Record<string, string>;
const crop = (scenario: string, method: string) => img[`../assets/upscalers/${scenario}-${method}.png`];
-/** What each upscaler is good and bad at, from an A/B on 8 images at x1.4, x2 and x4 (2026-10-12). */
+/** What each upscaler is good and bad at, from an A/B on 8 images at x1.4, x2 and x4, and the
+ * time on an M2 Pro (2026-10-12): Macs default to Lanczos, ESRGAN models take minutes there. */
export const UPSCALER_NOTES: Record<string, { key: string; label: string; text: string }> = {
"4x_NMKD-Superscale-SP_178000_G": {
key: "superscale",
label: "Superscale-SP",
- text: "The default. Sharp and natural on most photos and renders. On JPEG photos and dark areas it can draw a faint grid: lower the strength if you see one.",
+ text: isMac
+ ? "Sharp and natural, but very slow on a Mac: about 6 minutes to enlarge a 1 MP image to 2K on an M2 Pro. On JPEG photos and dark areas it can draw a faint grid."
+ : "The default. Sharp and natural on most photos and renders. On JPEG photos and dark areas it can draw a faint grid: lower the strength if you see one.",
},
"4xNomos2_hq_mosr": {
key: "mosr",
label: "MoSR (experimental)",
- text: "Included with Kokon and the fastest. Good for big enlargements (4×), but on clean images enlarged less than about 3× it draws a fine grid pattern.",
+ text: isMac
+ ? "Included with Kokon and the fastest AI upscaler on a Mac (about half a minute per megapixel on an M2 Pro), but on clean images enlarged less than about 3× it draws a fine grid pattern."
+ : "Included with Kokon and the fastest. Good for big enlargements (4×), but on clean images enlarged less than about 3× it draws a fine grid pattern.",
},
"4x-UltraSharp": {
key: "ultrasharp",
label: "UltraSharp",
- text: "Very clean, never a grid. But it smooths skin and textures into an artificial look and can invent lines and shapes, especially at night. Non-commercial license.",
+ text: `Very clean, never a grid. But it smooths skin and textures into an artificial look and can invent lines and shapes, especially at night. Non-commercial license.${isMac ? " Very slow on a Mac (minutes per image)." : ""}`,
},
lanczos: {
key: "lanczos",
label: "Lanczos",
- text: "A plain resize: never invents anything, but enlargements look soft.",
+ text: isMac
+ ? "The default on a Mac: a plain resize in about a second. It never invents anything, but enlargements look soft."
+ : "A plain resize: never invents anything, but enlargements look soft.",
},
};
diff --git a/src/components/WhatsNew.tsx b/src/components/WhatsNew.tsx
index 7a8d613..15e69a3 100644
--- a/src/components/WhatsNew.tsx
+++ b/src/components/WhatsNew.tsx
@@ -2,6 +2,7 @@ import { CircleAlert, Images, LayoutGrid, Maximize, SlidersHorizontal, Sparkles
import { useEffect, useState } from "react";
import { api } from "../lib/api";
import { formatBytes, installProgress, installStep, installTimeLeft } from "../lib/logic";
+import { isMac } from "../lib/platform";
import { refreshStatus, setState, toastError, useStore } from "../lib/store";
import type { GpuBackend, InstallItem, InstallStatus } from "../lib/types";
import { KokonMark } from "./Welcome";
@@ -10,11 +11,17 @@ import { KokonMark } from "./Welcome";
export const WHATS_NEW_VERSION = "0.9.1";
const NEWS = [
- {
- icon: Sparkles,
- title: "Superscale-SP by default",
- text: "Results are enlarged with a sharper, more natural upscaler. The previous one, UltraSharp, was great on some images but smoothed skin and textures into an artificial look on others. You can still download it.",
- },
+ isMac
+ ? {
+ icon: Sparkles,
+ title: "Plain resize by default on a Mac",
+ text: "AI upscalers take minutes on a Mac (Superscale-SP: about 6 minutes for a 1 MP image to 2K on an M2 Pro), so results are enlarged with Lanczos in about a second. You can still pick one.",
+ }
+ : {
+ icon: Sparkles,
+ title: "Superscale-SP by default",
+ text: "Results are enlarged with a sharper, more natural upscaler. The previous one, UltraSharp, was great on some images but smoothed skin and textures into an artificial look on others. You can still download it.",
+ },
{
icon: SlidersHorizontal,
title: "Choose your upscaler",
@@ -84,6 +91,8 @@ export function WhatsNew() {
const size = (missing ?? []).reduce((a, i) => a + Math.max(0, i.size - i.received), 0);
const prog = installProgress(items);
const showEngine = needsEngine || !!inst;
+ // Macs default to Lanczos: their update has no Superscale-SP download.
+ const getsSuperscale = (missing ?? []).some((i) => i.id === "upscaler_superscale");
const update = async () => {
if (!target) return;
@@ -121,8 +130,8 @@ export function WhatsNew() {
{showEngine && !done && (
<p className="welcome-lead">
- This version needs a newer stable-diffusion.cpp, the engine that runs the models on your graphics chip, and the Superscale-SP model. Your images, models and settings stay
- as they are.
+ This version needs a newer stable-diffusion.cpp, the engine that runs the models on your graphics chip{getsSuperscale ? ", and the Superscale-SP model" : ""}. Your images,
+ models and settings stay as they are.
</p>
)}
{(running || (inst && !done && !failed.length)) && (
@@ -161,7 +170,9 @@ export function WhatsNew() {
Later
</button>
</div>
- <span className="muted welcome-foot">Until you update, enlargements use a plain resize. You can update later from Settings → Models → Set up models…</span>
+ <span className="muted welcome-foot">
+ {isMac ? "" : "Until you update, enlargements use a plain resize. "}You can update later from Settings → Models → Set up models…
+ </span>
</>
)}
</div>