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
import { Check, CircleAlert, Download, FileSearch, FolderOpen, Loader2 } from "lucide-react";
import { useEffect, useMemo, useState } from "react";
import { api } from "../lib/api";
import { formatBytes } from "../lib/logic";
import { native } from "../lib/native";
import { toastError } from "../lib/store";
import type { GpuBackend, InstallItem, ModelFileInfo, ModelRole, ModelScan, ModelSlot, ModelsStatus } from "../lib/types";
const ROLES: { role: ModelRole; label: string; official: string }[] = [
{ role: "diffusion", label: "Image model", official: "turbo_q4_k" },
{ role: "vae", label: "VAE", official: "vae" },
{ role: "text_encoder", label: "Text encoder", official: "text_encoder" },
{ role: "vision_projector", label: "Vision projector", official: "vision_projector" },
];
/** A file of the user's, the official file downloaded into `downloadDir`, or (vision projector
* only) nothing because the text encoder has the vision weights. */
type Choice = { kind: "path"; path: string } | { kind: "download" } | { kind: "none" };
const fileName = (p: string) => p.split(/[\\/]/).pop() ?? p;
const folderName = (p: string) => p.split(/[\\/]/).filter(Boolean).pop() ?? p;
function describe(f: ModelFileInfo): string {
return [f.weights, formatBytes(f.size), f.official ? "official file" : null].filter(Boolean).join(" · ");
}
/** Setup's "Use models I already have", as a guided flow: choose the folder (e.g. ComfyUI's
* models folder), check what was found (one row per piece, changed only where needed), use
* them. Pieces the user lacks are the official files, downloaded into a folder of their choice.
* Files are used where they are. */
export function OwnModels(props: {
models: ModelsStatus;
/** The chosen preset's official model files (sizes of what can be downloaded). */
officialFiles: InstallItem[];
vramMib: number | null;
/** The engine chosen in Setup (FP8 files are slow on Vulkan). */
backend?: GpuBackend | null;
/** A scan to start from (Setup's "Search another folder"). */
initialScan: ModelScan | null;
/** Resolves once Setup has checked the new files (the button spins until then). */
onSaved: (s: ModelsStatus) => Promise<void>;
}) {
const { models } = props;
const inUse = models.active.source === "user";
const [scan, setScan] = useState<ModelScan | null>(props.initialScan);
const [scanning, setScanning] = useState(false);
const [saving, setSaving] = useState(false);
// Pick each file without a folder scan.
const [manual, setManual] = useState(false);
// Changing files that are in use (otherwise their summary is shown).
const [editing, setEditing] = useState(!inUse || models.missing.length > 0 || !!props.initialScan);
// The row whose options are open.
const [open, setOpen] = useState<ModelRole | null>(null);
// Files chosen with "Choose a file…" (not in the scan), and the current profile's own files.
const [extra, setExtra] = useState<ModelFileInfo[]>([]);
const [downloadDir, setDownloadDir] = useState(models.default_download_dir);
const [choice, setChoice] = useState<Record<ModelRole, Choice>>(() => {
const c: Record<ModelRole, Choice> = { diffusion: { kind: "download" }, vae: { kind: "download" }, text_encoder: { kind: "download" }, vision_projector: { kind: "download" } };
if (inUse) {
for (const s of models.slots) if (s.role) c[s.role] = { kind: "path", path: s.path };
if (!models.slots.some((s) => s.role === "vision_projector")) c.vision_projector = { kind: "none" };
}
return c;
});
// The user profile's files, with what they are (missing ones are shown as such).
useEffect(() => {
if (inUse) setExtra(models.slots.filter((s) => s.info).map((s) => s.info!));
}, [models]);
const known = useMemo(() => {
const all = [...(scan?.files ?? []), ...extra];
return all.filter((f, i) => all.findIndex((g) => g.path === f.path) === i);
}, [scan, extra]);
const info = (path: string) => known.find((f) => f.path === path) ?? null;
const applyPicks = (s: ModelScan) => {
const te = s.picks.text_encoder ? s.files.find((f) => f.path === s.picks.text_encoder) : null;
setChoice(() => {
const n = {} as Record<ModelRole, Choice>;
for (const { role } of ROLES) {
const p = s.picks[role];
n[role] = p ? { kind: "path", path: p } : { kind: "download" };
}
// A text encoder with the vision weights: no projector needed.
if (te?.has_vision) n.vision_projector = { kind: "none" };
return n;
});
};
// A search from the missing-files card that did not find everything: continue from it.
useEffect(() => {
if (!props.initialScan) return;
setScan(props.initialScan);
applyPicks(props.initialScan);
setEditing(true);
}, [props.initialScan]);
const chooseFolder = async () => {
const dir = await native.pickFolder(scan?.dir ?? models.scan_dir ?? models.user_scan_dir ?? undefined);
if (!dir) return;
setScanning(true);
try {
const s = await api.scanModels(dir);
setScan(s);
applyPicks(s);
setEditing(true);
setOpen(null);
} catch (e) {
toastError(e);
} finally {
setScanning(false);
}
};
const browse = async (role: ModelRole) => {
const cur = choice[role];
const p = await native.pickFile(["gguf", "safetensors"], cur.kind === "path" ? cur.path : scan?.dir);
if (!p) return;
try {
const f = await api.inspectModel(p);
setExtra((x) => [...x.filter((g) => g.path !== f.path), f]);
pick(role, { kind: "path", path: p }, f);
} catch (e) {
toastError(e);
}
};
const pick = (role: ModelRole, c: Choice, f: ModelFileInfo | null = c.kind === "path" ? info(c.path) : null) => {
setChoice((x) => ({ ...x, [role]: c, ...(role === "text_encoder" && f?.has_vision ? { vision_projector: { kind: "none" } } : {}) }));
setOpen(null);
};
const teInfo = choice.text_encoder.kind === "path" ? info(choice.text_encoder.path) : null;
const teHasVision = !!teInfo?.has_vision;
const officialSize = (role: ModelRole) => {
const id = ROLES.find((r) => r.role === role)!.official;
// By family: the text encoder of the preset, the image model the official files use.
return props.officialFiles.find((f) => f.id === id || f.id.startsWith(`${id}_`))?.size ?? 0;
};
// What each row says, and whether the set can be used.
const rows = ROLES.map(({ role, label }) => {
const c = role === "vision_projector" && teHasVision ? ({ kind: "none" } as Choice) : choice[role];
let state: "ok" | "download" | "bad";
let text: string;
let detail: string | null = null;
if (c.kind === "none" && teHasVision) [state, text] = ["ok", "Not needed: included in your text encoder"];
else if (c.kind === "none") [state, text] = ["bad", "Needed for editing: this text encoder has no vision weights"];
else if (c.kind === "download") [state, text] = ["download", `${scan ? "Not found in this folder: " : ""}the official file will be downloaded (${formatBytes(officialSize(role))})`];
else {
const f = info(c.path);
const slot = models.slots.find((s) => s.path === c.path);
detail = c.path;
if (slot && !slot.exists) [state, text] = ["bad", "Not found"];
else if (!f) [state, text] = ["ok", fileName(c.path)];
else if (f.role !== role) [state, text] = ["bad", f.role ? `${fileName(c.path)} is a ${f.role.replace("_", " ")} file` : f.problem ?? "Not a Qwen-Image 2.1 file"];
else if (!f.compatible) [state, text] = ["bad", f.problem ?? "Not usable with Qwen-Image 2.1"];
else [state, text] = ["ok", `${fileName(c.path)} · ${describe(f)}`];
}
const candidates = known.filter((f) => f.role === role && f.compatible);
return { role, label, c, state, text, detail, candidates };
});
const canSave = rows.every((r) => r.state !== "bad");
const downloads = rows.filter((r) => r.c.kind === "download");
const found = rows.filter((r) => r.c.kind === "path").length;
const dit = choice.diffusion.kind === "path" ? info(choice.diffusion.path) : null;
const tooLarge = !!dit && props.vramMib != null && dit.size / 2 ** 20 > props.vramMib - 4096;
// RX 9070 XT (#38): FP8 safetensors 3.07 s/it on Vulkan, the Q4_K GGUF 1.75 s/it.
const slowFp8 = !!dit && dit.role === "diffusion" && dit.weights === "FP8" && props.backend === "vulkan";
const skipped = scan ? scan.files.filter((f) => !f.compatible).length : 0;
const save = async () => {
const slot = (c: Choice): ModelSlot | null => (c.kind === "path" ? { path: c.path } : c.kind === "download" ? { download: downloadDir } : null);
setSaving(true);
try {
const s = await api.setUserModels({
diffusion_model: slot(rows[0].c)!,
vae: slot(rows[1].c)!,
text_encoder: slot(rows[2].c)!,
vision_projector: slot(rows[3].c),
scan_dir: scan?.dir ?? models.scan_dir ?? models.user_scan_dir,
});
await props.onSaved(s);
setEditing(false);
} catch (e) {
toastError(e);
} finally {
setSaving(false);
}
};
const icon = (state: "ok" | "download" | "bad") => (state === "ok" ? <Check size={16} /> : state === "download" ? <Download size={16} /> : <CircleAlert size={16} />);
// Step 1: where are the models?
if (editing && !scan && !manual && !inUse) {
return (
<div className="own-start">
<FolderOpen size={26} strokeWidth={1.6} />
<strong>Choose the folder with your models</strong>
<span className="muted">
For example <code>ComfyUI\models</code>. Kokon finds the Qwen-Image 2.1 files in it and in its sub-folders, as GGUF or safetensors. Nothing is copied or moved.
</span>
<button className="btn primary big" onClick={chooseFolder} disabled={scanning}>
{scanning ? <Loader2 className="spin" size={16} /> : <FolderOpen size={16} />} {scanning ? "Looking for models…" : "Choose folder…"}
</button>
<button className="link-btn" onClick={() => setManual(true)}>
or pick each file yourself
</button>
</div>
);
}
return (
<div className="own-models">
<div className="own-head">
<div>
<strong>{!editing ? "Using your own files" : scan ? `In ${folderName(scan.dir)}` : "Your files"}</strong>
<span className="muted">
{!editing
? "Not tested by Kokon. Change them at any time, or switch back to the official files above."
: scan
? `Found ${found} of 4${downloads.length ? `; ${downloads.length === 1 ? "1 piece" : `${downloads.length} pieces`} will be downloaded` : ""}${skipped ? ` · ${skipped} file${skipped === 1 ? "" : "s"} of other models skipped` : ""}`
: "Choose a file for each piece, or download the official one."}
</span>
</div>
{editing ? (
<button className="btn" onClick={chooseFolder} disabled={scanning}>
{scanning ? <Loader2 className="spin" size={14} /> : <FolderOpen size={14} />} {scan || inUse ? "Another folder…" : "Search a folder…"}
</button>
) : (
<button className="btn" onClick={() => setEditing(true)}>
Change files
</button>
)}
</div>
<div className="own-list">
{rows.map((r) => (
<div className={`own-row ${r.state}`} key={r.role}>
<div className="own-row-main">
{icon(r.state)}
<div className="own-row-text">
<span className="own-row-label">{r.label}</span>
<span className="own-row-value">{r.text}</span>
{r.detail && <span className="own-row-path">{r.detail}</span>}
</div>
{editing && (
<button className="btn small" onClick={() => setOpen(open === r.role ? null : r.role)} aria-expanded={open === r.role}>
Change
</button>
)}
</div>
{editing && open === r.role && (
<div className="own-options">
{r.candidates.map((f) => (
<button key={f.path} className={`own-option ${r.c.kind === "path" && r.c.path === f.path ? "on" : ""}`} onClick={() => pick(r.role, { kind: "path", path: f.path }, f)}>
<span>{fileName(f.path)}</span>
<span className="muted">{describe(f)}</span>
</button>
))}
{r.role === "vision_projector" && teHasVision && (
<button className={`own-option ${r.c.kind === "none" ? "on" : ""}`} onClick={() => pick(r.role, { kind: "none" })}>
<span>Not needed</span>
<span className="muted">your text encoder includes the vision weights</span>
</button>
)}
<button className={`own-option ${r.c.kind === "download" ? "on" : ""}`} onClick={() => pick(r.role, { kind: "download" })}>
<span>
<Download size={13} /> Download the official file
</span>
<span className="muted">{formatBytes(officialSize(r.role))}</span>
</button>
<button className="own-option" onClick={() => browse(r.role)}>
<span>
<FileSearch size={13} /> Choose a file…
</span>
<span className="muted">.gguf or .safetensors</span>
</button>
</div>
)}
</div>
))}
</div>
{editing && downloads.length > 0 && (
<div className="own-download">
<span className="muted">
Downloads go to <strong>{downloadDir}</strong>
</span>
<button
className="link-btn"
onClick={async () => {
const d = await native.pickFolder(downloadDir);
if (d) setDownloadDir(d);
}}
>
Change
</button>
</div>
)}
{editing && tooLarge && (
<div className="setup-note">
This image model is {formatBytes(dit!.size)}, probably too large for {Math.round((props.vramMib ?? 0) / 1024)} GB of VRAM. Kokon makes smaller images to fit, but it may still run out of memory
or be slow. A 4-bit or 8-bit version (GGUF Q4_K or Q8_0, or INT8) works better on this card.
</div>
)}
{dit?.role === "diffusion" && dit.compatible && dit.warning && <div className="setup-note">{dit.warning}</div>}
{editing && slowFp8 && (
<div className="setup-note">FP8 files run much slower on AMD and Intel cards with Vulkan than the GGUF (Q4_K) file Kokon downloads. For speed, download the official file or choose a GGUF.</div>
)}
{editing && (
<div className="own-actions">
<button className="btn primary" disabled={!canSave || saving} onClick={save}>
{saving ? <Loader2 className="spin" size={14} /> : <Check size={14} />} {saving ? "Checking the files…" : "Use these files"}
</button>
{inUse && models.missing.length === 0 && (
<button className="btn ghost" onClick={() => setEditing(false)}>
Cancel
</button>
)}
{!tooLarge && <span className="muted">Your own files are not tested by Kokon: quality and speed may differ.</span>}
</div>
)}
</div>
);
}