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.
TypeScript ai-photo-editorcomfyui-alternativegenerative-fillimage-editingimage-generationinpaintinglinuxlocal-aimacosmcpmcp-servernano-banana-alternativeofflinephoto-editorprivacyqwen-imageqwen-image-2-1stable-diffusion-cppvulkanwindows
// "Blend into scene" (LayerDoc): what the AI read about where each object stands, shown before the
// edit so the user can correct it (Gabriele 11-02: "un pop up con possibilità di correggerlo"). The
// reading picks the effects the prompt names (lib/blend.ts): a wrong "wet" draws a reflection on dry
// ground, a wrong "dry" leaves a puddle without one.
import { Loader2, Sun, X } from "lucide-react";
import { useEffect, useRef } from "react";
import { blendEffects, FALLS, SURFACES } from "../lib/blend";
import type { SceneReading } from "../lib/types";
import { IconButton } from "./ui";
export interface BlendRow {
id: string;
name: string;
/** Missing while it is being read. */
reading?: SceneReading;
/** Why the reading is the safe default. */
failed?: string;
}
/** At the fast preset (11-02 bench, RTX 5070 Ti): ~2 s to read each object's place, then one edit of
* the whole scene for all of them, 12-24 s. */
const READ_SECONDS = 2;
const EDIT_SECONDS = 20;
export function BlendChoice({ rows, onChange, onBlend, onCancel }: { rows: BlendRow[]; onChange: (id: string, r: SceneReading) => void; onBlend: () => void; onCancel: () => void }) {
const reading = rows.some((r) => !r.reading);
const go = useRef<HTMLButtonElement>(null);
useEffect(() => {
if (!reading) go.current?.focus();
}, [reading]);
useEffect(() => {
const onKey = (e: KeyboardEvent) => {
if (e.key !== "Escape") return;
e.preventDefault();
e.stopImmediatePropagation();
onCancel();
};
window.addEventListener("keydown", onKey, true);
return () => window.removeEventListener("keydown", onKey, true);
}, [onCancel]);
const s = EDIT_SECONDS + rows.length * READ_SECONDS;
const time = s < 60 ? `about ${Math.max(5, Math.round(s / 5) * 5)} s` : `about ${Math.round(s / 60)} min`;
return (
<div className="modal-scrim" onPointerDown={(e) => e.target === e.currentTarget && onCancel()}>
<div className="split-choice blend-choice" role="dialog" aria-modal="true" aria-labelledby="blend-choice-title">
<span className="split-choice-close">
<IconButton title="Cancel (Esc)" onClick={onCancel}>
<X size={18} />
</IconButton>
</span>
<h2 id="blend-choice-title">
Blend into the scene <span className="exp-tag">Experimental</span>
</h2>
<p className="split-choice-lead">
The AI looked at where {rows.length === 1 ? "the object stands" : "each object stands"}. It redraws the scene with what that place does: the contact shadow, the
shadow, a reflection on wet or shiny ground. Check what it read and change anything that is wrong. It is
experimental and often does not work yet: blending all the objects at once works best.
</p>
<div className="blend-rows">
{rows.map((row) => {
const r = row.reading;
return (
<div key={row.id} className="blend-row">
<div className="blend-row-head">
<span className="blend-row-name">{row.name}</span>
<span className={`blend-row-state ${row.failed ? "warn" : ""}`} title={row.failed}>
{!r ? (
<>
<Loader2 size={13} className="spin" /> Reading the scene…
</>
) : row.failed ? (
"Could not read it: check these"
) : r.complete ? (
"Read by the AI"
) : (
"Partly read: check these"
)}
</span>
</div>
<div className={`blend-fields ${r ? "" : "waiting"}`}>
<label>
<span>Ground</span>
<select className="select" value={r?.surface ?? ""} disabled={!r} onChange={(e) => r && onChange(row.id, { ...r, surface: e.target.value as SceneReading["surface"] })}>
{!r && <option value="" />}
{SURFACES.map((o) => (
<option key={o.id} value={o.id}>
{o.label}
</option>
))}
</select>
</label>
<label>
<span>Shadows</span>
<div className="seg">
{([false, true] as const).map((hard) => (
<button key={String(hard)} className={r && r.hard_shadows === hard ? "on" : ""} disabled={!r} onClick={() => r && onChange(row.id, { ...r, hard_shadows: hard })}>
{hard ? "Hard (sun)" : "Soft"}
</button>
))}
</div>
</label>
<label>
<span>They fall</span>
<select className="select" value={r?.shadows_fall ?? ""} disabled={!r} onChange={(e) => r && onChange(row.id, { ...r, shadows_fall: e.target.value as SceneReading["shadows_fall"] })}>
{!r && <option value="" />}
{FALLS.map((o) => (
<option key={o.id} value={o.id}>
{o.label}
</option>
))}
</select>
</label>
</div>
{r && (
<details className="blend-prompt">
<summary>What the AI is asked</summary>
<p>{blendEffects(row.name, r)}</p>
</details>
)}
</div>
);
})}
</div>
<div className="split-choice-foot">
<span>
{time} · the rest of the picture stays as it is · moved it later? Blend again
</span>
<span className="blend-buttons">
<button className="btn ghost" onClick={onCancel}>
Cancel
</button>
<button ref={go} className="btn primary" onClick={onBlend} disabled={reading}>
<Sun size={15} /> Blend
</button>
</span>
</div>
</div>
</div>
);
}