v4.1.8: 이미지/동영상 생성 스튜디오 앱 + SDXL/FLUX 품질 개선

- 이미지·동영상 생성 전용 앱(studio-app.html) + API 라우트 추가, 갤러리 기능 포함
- SD1.5 → SDXL 교체 (프롬프트 반영력/해상도 개선), 한글 프롬프트 자동 영어 번역 추가
- image_generate에 quality="high" 옵션 추가 (FLUX.1-schnell 4비트 양자화, 게이트 없는 커뮤니티 미러 사용)
- video_generate에 guidance_scale 노출, 기본 steps 상향
- 생성기 서브프로세스가 결과 없이 죽는 경우의 크래시 방지 + 에러 로깅 강화
- 멀티 GPU 환경에서 시스템 통계 GPU 미터가 깨지던 nvidia-smi 파싱 버그 수정

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

package.json

@@ -1,6 +1,6 @@
 {
   "name": "smallclaw",
-  "version": "4.1.7",
+  "version": "4.1.8",
   "description": "Local AI agent framework powered by Ollama - OpenClaw alternative",
   "main": "dist/index.js",
   "bin": {

src/gateway/routes/imagegen.ts

@@ -0,0 +1,119 @@
+import { Express, Request, Response } from 'express';
+import fs from 'fs';
+import path from 'path';
+import { imageGenerateTool, videoGenerateTool } from '../../tools/imagegen.js';
+
+const IMAGE_EXTS = new Set(['.png', '.jpg', '.jpeg', '.webp']);
+const VIDEO_EXTS = new Set(['.mp4']);
+
+function galleryDir(workspace: string): string {
+  return path.join(workspace, 'generated-media');
+}
+
+export function registerImagegenRoutes(app: Express): void {
+  app.post('/api/imagegen/generate-image', async (req: Request, res: Response) => {
+    const user = (req as any).user;
+    if (!user) return res.status(401).json({ error: 'Unauthorized' });
+
+    console.log(`[imagegen] generate-image start user=${user.username} prompt="${String(req.body?.prompt || '').slice(0, 120)}"`);
+    try {
+      const outDir = galleryDir(user.workspace);
+      fs.mkdirSync(outDir, { recursive: true });
+      const filename = `img_${Date.now()}.png`;
+
+      const result = await imageGenerateTool.execute({
+        ...req.body,
+        output: path.join('generated-media', filename),
+        _workspacePath: user.workspace,
+      });
+
+      if (!result.success) {
+        console.error(`[imagegen] generate-image failed for user=${user.username}: ${result.error}`, result.stderr ? `\nstderr: ${result.stderr}` : '');
+        return res.status(500).json({ error: result.error, stderr: result.stderr });
+      }
+      res.json({
+        success: true,
+        url: `/api/files/${result.data.rel_path}`,
+        width: result.data.width,
+        height: result.data.height,
+      });
+    } catch (err: any) {
+      // Belt-and-suspenders: never let an unexpected throw fall through to Express's
+      // default HTML error page — the studio app's fetch always expects JSON back.
+      console.error(`[imagegen] generate-image threw for user=${user.username}:`, err);
+      res.status(500).json({ error: err?.message || 'image_generate failed unexpectedly' });
+    }
+  });
+
+  app.post('/api/imagegen/generate-video', async (req: Request, res: Response) => {
+    const user = (req as any).user;
+    if (!user) return res.status(401).json({ error: 'Unauthorized' });
+
+    console.log(`[imagegen] generate-video start user=${user.username} prompt="${String(req.body?.prompt || '').slice(0, 120)}"`);
+    try {
+      const outDir = galleryDir(user.workspace);
+      fs.mkdirSync(outDir, { recursive: true });
+      const filename = `vid_${Date.now()}.mp4`;
+
+      const result = await videoGenerateTool.execute({
+        ...req.body,
+        output: path.join('generated-media', filename),
+        _workspacePath: user.workspace,
+      });
+
+      if (!result.success) {
+        console.error(`[imagegen] generate-video failed for user=${user.username}: ${result.error}`, result.stderr ? `\nstderr: ${result.stderr}` : '');
+        return res.status(500).json({ error: result.error, stderr: result.stderr });
+      }
+      res.json({
+        success: true,
+        url: `/api/files/${result.data.rel_path}`,
+        width: result.data.width,
+        height: result.data.height,
+        num_frames: result.data.num_frames,
+        fps: result.data.fps,
+      });
+    } catch (err: any) {
+      console.error(`[imagegen] generate-video threw for user=${user.username}:`, err);
+      res.status(500).json({ error: err?.message || 'video_generate failed unexpectedly' });
+    }
+  });
+
+  app.get('/api/imagegen/gallery', (req: Request, res: Response) => {
+    const user = (req as any).user;
+    if (!user) return res.status(401).json({ error: 'Unauthorized' });
+
+    const outDir = galleryDir(user.workspace);
+    if (!fs.existsSync(outDir)) return res.json({ items: [] });
+
+    const items = fs.readdirSync(outDir)
+      .map((name) => {
+        const ext = path.extname(name).toLowerCase();
+        const type = IMAGE_EXTS.has(ext) ? 'image' : VIDEO_EXTS.has(ext) ? 'video' : null;
+        if (!type) return null;
+        const stat = fs.statSync(path.join(outDir, name));
+        return {
+          name,
+          type,
+          url: `/api/files/generated-media/${encodeURIComponent(name)}`,
+          size: stat.size,
+          mtime: stat.mtimeMs,
+        };
+      })
+      .filter((x): x is NonNullable<typeof x> => x !== null)
+      .sort((a, b) => b.mtime - a.mtime);
+
+    res.json({ items });
+  });
+
+  app.delete('/api/imagegen/gallery/:name', (req: Request, res: Response) => {
+    const user = (req as any).user;
+    if (!user) return res.status(401).json({ error: 'Unauthorized' });
+
+    const name = path.basename(String(req.params.name || ''));
+    const target = path.join(galleryDir(user.workspace), name);
+    if (!fs.existsSync(target)) return res.status(404).json({ error: 'Not found' });
+    fs.unlinkSync(target);
+    res.json({ success: true });
+  });
+}

src/gateway/server-v2.ts

@@ -59,6 +59,7 @@ import { writeSkillPackFromContent } from '../skills/processor.js';
 import { summarizeSkillForApi } from '../tools/skills.js';
 import { getToolRegistry } from '../tools/registry.js';
 import { registerSettingsRoutes } from './routes/settings.js';
+import { registerImagegenRoutes } from './routes/imagegen.js';
 import {
   browserOpen,
   browserSnapshot,
@@ -5785,7 +5786,7 @@ async function handleChat(
       content: isTranslateSession ? `You are a medical translator. Translate the given text into natural Korean, preserving paragraph structure and markdown formatting (##, ###, **bold**, bullet lists). Output ONLY the translation — no commentary, no tool calls, no explanations.` : isProjSession ? `You are a project file designer. Output ONLY the project-files JSON block as instructed. No tool calls. No extra text.` : `${executionModeSystemBlock ? `${executionModeSystemBlock}\n\n` : ''}You are SmallClaw 🦞, a local AI assistant.\nCurrent date: ${dateStr}, ${timeStr}.\nNever search for or link SmallClaw repos unless the user is asking about SmallClaw itself.\nThis app runs on the user's own machine — browser/desktop automation requests are pre-authorized.\nKeep responses SHORT (1-2 sentences). Don't think out loud. Act and report. Greet naturally without tools.
 ANTI-HALLUCINATION: When a tool returns a result, report EXACTLY what the tool returned — never contradict or ignore tool output. If a tool says "(no rows)", say so. Never invent data, file contents, table names, or command output. If you don't know something, call a tool to find out or say you don't know.
 IMAGE EDITING RULE: NEVER call image_edit (or any editing tool) when a user uploads a photo without explicitly requesting edits. Uploading a photo is NOT a request to edit it. Only call image_edit when the user's message explicitly asks for an edit (e.g. "수채화로 바꿔줘", "회전해줘"). Violating this rule is a critical error.
-IMAGE/VIDEO GENERATION: When a user asks to create/draw/generate a NEW image from a description (no existing photo involved), call image_generate (local SD1.5, a few seconds). When they ask for a short video/clip/animation from a description, call video_generate (local LTX-Video, 30–90 seconds — tell the user it'll take a bit before calling it). Both run entirely on local GPU hardware, no external API or cost. Do not confuse these with image_edit, which only modifies an existing uploaded/generated image.${browserRuleBlock}
+IMAGE/VIDEO GENERATION: When a user asks to create/draw/generate a NEW image from a description (no existing photo involved), call image_generate (local SDXL, 10-20 seconds by default). If the user explicitly asks for higher quality/detail/photorealism (e.g. "고품질로", "디테일 살려서"), pass quality="high" instead — this uses FLUX.1-schnell, noticeably better detail but 40-60 seconds total, so tell the user it'll take a bit before calling it. When they ask for a short video/clip/animation from a description, call video_generate (local LTX-Video, 30–90 seconds — tell the user it'll take a bit before calling it). All run entirely on local GPU hardware, no external API or cost. Do not confuse these with image_edit, which only modifies an existing uploaded/generated image.${browserRuleBlock}
 CHEMISTRY NOTATION: NEVER draw molecular structures as ASCII art (H/C/#/=/\\/| characters arranged to look like a diagram) — it always renders as garbled, misaligned text. Instead: for a formula or reaction, use LaTeX inside $...$ (e.g. $\\ce{C4H10}$, $\\ce{2H2 + O2 -> 2H2O}$ — mhchem extension is loaded). For an actual 2D structure with real bond lines (rings, branches), output a \`\`\`smiles\`\`\` code block containing the SMILES string (e.g. \`\`\`smiles\\nc1ccccc1\\n\`\`\` for benzene, \`\`\`smiles\\nCC(=O)Oc1ccccc1C(=O)O\\n\`\`\` for aspirin) — the client automatically renders it as a proper 2D diagram with bond lines.
 CODE OUTPUT: When writing code in a fenced code block, start with a filename comment on line 1: \`# filename: snake_game.py\` (Python), \`// filename: app.js\` (JS/C), \`<!-- filename: index.html -->\` (HTML). Never repeat code already written in this conversation. For modifications to existing files, use coder_overwrite_lines or coder_insert_lines (not coder_write_file — it only works for NEW files). All code changes are presented as diffs for the user to review before being applied. Write code directly — do not ask for permission.
 PACKAGE INSTALL: NEVER run pip install, npm install, apt-get, or any package installation command. If a package is missing, just write the code and mention the package name in a comment — let the user decide whether to install it. Do NOT attempt to install packages yourself.
@@ -12671,6 +12672,7 @@ app.post('/api/schedules/parse', (req: any, res: any) => {
 
 // ── Settings routes (extracted to routes/settings.ts) ──────────────────────────
 registerSettingsRoutes(app);
+registerImagegenRoutes(app);
 
 // GET /api/credentials/status — list which vault keys are currently stored (names only, no values)
 app.get('/api/credentials/status', (_req, res) => {
@@ -12904,17 +12906,32 @@ app.get('/api/system-stats', async (req, res) => {
         'nvidia-smi --query-gpu=name,utilization.gpu,memory.used,memory.total --format=csv,noheader,nounits',
         { timeout: 3000, encoding: 'utf8', stdio: ['ignore', 'pipe', 'pipe'] },
       );
-      const parts = smiOut.trim().split(',').map((s: string) => s.trim());
-      if (parts.length >= 4) {
-        const vramUsedMb = Number(parts[2]);
-        const vramTotalMb = Number(parts[3]);
+      // One line per physical GPU — a machine with 2+ cards must split by line
+      // before splitting by comma, or fields from adjacent lines bleed together
+      // (e.g. memory.total gets "12288\nNVIDIA GeForce RTX 3060" glued on, NaN's
+      // the VRAM percent). Aggregate across all detected GPUs: sum VRAM (so the
+      // meter reflects total usage), max utilization (so activity on any card shows).
+      const gpuLines = smiOut.trim().split('\n')
+        .map((line: string) => line.split(',').map((s: string) => s.trim()))
+        .filter((parts: string[]) => parts.length >= 4);
+      if (gpuLines.length > 0) {
+        let vramUsedMbSum = 0;
+        let vramTotalMbSum = 0;
+        let utilMax = 0;
+        const names: string[] = [];
+        for (const parts of gpuLines) {
+          vramUsedMbSum += Number(parts[2]) || 0;
+          vramTotalMbSum += Number(parts[3]) || 0;
+          utilMax = Math.max(utilMax, Number(parts[1]) || 0);
+          names.push(parts[0]);
+        }
         gpuStats = {
           available: true,
-          name: parts[0],
-          gpu_util_percent: Number(parts[1]),
-          vram_used_percent: vramTotalMb > 0 ? (vramUsedMb / vramTotalMb) * 100 : 0,
-          vram_used_gb: vramUsedMb / 1024,
-          vram_total_gb: vramTotalMb / 1024,
+          name: gpuLines.length > 1 ? `${names[0]} ×${gpuLines.length}` : names[0],
+          gpu_util_percent: utilMax,
+          vram_used_percent: vramTotalMbSum > 0 ? (vramUsedMbSum / vramTotalMbSum) * 100 : 0,
+          vram_used_gb: vramUsedMbSum / 1024,
+          vram_total_gb: vramTotalMbSum / 1024,
         };
       }
     } catch { /* nvidia-smi already confirmed working at startup; ignore transient errors */ }

src/tools/imagegen.ts

@@ -4,8 +4,39 @@ import fs from 'fs';
 import { ToolResult } from '../types.js';
 import { getWorkspacePath } from '../config/paths.js';
 import { buildImageMarkdown } from './image.js';
+import { getOllamaConfig } from './web.js';
 
-// Local diffusion models (SD1.5, LTX-Video) run in a dedicated venv with their
+// SDXL/LTX-Video's text encoders (CLIP/T5) are trained overwhelmingly on English
+// captions, so non-English prompts (e.g. Korean) produce poor prompt adherence.
+// Route non-ASCII prompts through the configured Ollama chat model for a quick
+// English translation before handing them to the diffusion pipeline.
+async function translatePromptToEnglish(text: string): Promise<string | null> {
+  if (!text || !/[^\x00-\x7F]/.test(text)) return null;
+  try {
+    const { endpoint, model } = getOllamaConfig();
+    const res = await fetch(`${endpoint}/api/chat`, {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({
+        model,
+        messages: [{
+          role: 'user',
+          content: `Translate the following image/video generation prompt into natural, descriptive English. Output ONLY the translated English text — no quotes, no explanation:\n\n${text}`,
+        }],
+        stream: false,
+      }),
+      signal: AbortSignal.timeout(30_000),
+    });
+    if (!res.ok) return null;
+    const data: any = await res.json();
+    const translated = String(data.message?.content || '').trim();
+    return translated || null;
+  } catch {
+    return null;
+  }
+}
+
+// Local diffusion models (SDXL, LTX-Video) run in a dedicated venv with their
 // own torch/diffusers stack, pinned to the second GPU (04:00.0 — kept free of
 // the voice engine that permanently resides on GPU0). See
 // /srv/homeclaw/.smallclaw/imagegen-venv.
@@ -17,17 +48,31 @@ function runVenvPython(script: string, timeoutMs: number): Promise<any> {
   return new Promise((resolve) => {
     const child = spawn(VENV_PYTHON, ['-c', script], {
       timeout: timeoutMs,
-      env: { ...process.env, HF_HOME, CUDA_VISIBLE_DEVICES: GEN_GPU },
+      // BNB_CUDA_VERSION: bitsandbytes (used by the FLUX high-quality path) ships no
+      // prebuilt binary for our CUDA 13.2 torch build yet — pin it to the newest
+      // available (13.0) binary, which is ABI-compatible. No-op for SDXL/LTX, which
+      // don't use bitsandbytes.
+      env: { ...process.env, HF_HOME, CUDA_VISIBLE_DEVICES: GEN_GPU, BNB_CUDA_VERSION: '130' },
     });
     let out = '';
     let err = '';
     child.stdout.on('data', (d: Buffer) => { out += d.toString('utf8'); });
     child.stderr.on('data', (d: Buffer) => { err += d.toString('utf8'); });
-    child.on('close', () => {
+    child.on('close', (code: number | null, signal: string | null) => {
       const marker = out.lastIndexOf('###RESULT###');
-      const jsonPart = marker !== -1 ? out.slice(marker + '###RESULT###'.length) : out;
+      if (marker === -1) {
+        // Process exited (crashed, OOM-killed, or timed out) without ever printing a
+        // result marker — never silently fall through to a success-shaped {}, since
+        // that leaves result.output undefined and crashes the caller downstream.
+        resolve({
+          error: `Generator process exited without output (code=${code}, signal=${signal})`,
+          trace: (err || out).slice(-1500),
+        });
+        return;
+      }
+      const jsonPart = out.slice(marker + '###RESULT###'.length);
       try {
-        resolve(JSON.parse(jsonPart.trim() || '{}'));
+        resolve(JSON.parse(jsonPart.trim()));
       } catch {
         resolve({ error: 'Failed to parse generator output', raw: (jsonPart || err).slice(-1500) });
       }
@@ -37,19 +82,20 @@ function runVenvPython(script: string, timeoutMs: number): Promise<any> {
 }
 
 // ---------------------------------------------------------------------------
-// image_generate — local Stable Diffusion 1.5 text-to-image
+// image_generate — local SDXL text-to-image
 // ---------------------------------------------------------------------------
-const SD15_SCRIPT = (p: Record<string, any>) => `
+const SDXL_SCRIPT = (p: Record<string, any>) => `
 import os, json, sys
 try:
     import torch
-    from diffusers import StableDiffusionPipeline
+    from diffusers import StableDiffusionXLPipeline
 
-    pipe = StableDiffusionPipeline.from_pretrained(
-        "stable-diffusion-v1-5/stable-diffusion-v1-5",
-        torch_dtype=torch.float16, safety_checker=None,
+    pipe = StableDiffusionXLPipeline.from_pretrained(
+        "stabilityai/stable-diffusion-xl-base-1.0",
+        torch_dtype=torch.float16, variant="fp16", use_safetensors=True,
     )
     pipe = pipe.to("cuda")
+    pipe.enable_vae_slicing()
 
     kwargs = dict(
         prompt=${JSON.stringify(p.prompt)},
@@ -73,20 +119,87 @@ except Exception as e:
     print("###RESULT###" + json.dumps({"error": str(e), "trace": traceback.format_exc()[-800:]}))
 `;
 
+// ---------------------------------------------------------------------------
+// image_generate (quality: 'high') — local FLUX.1-schnell, 4-bit quantized
+// ---------------------------------------------------------------------------
+// Community mirror of the (gated) official repo — same Apache-2.0 weights,
+// just rehosted without the HF license-gate. Needed because neither FLUX.1-dev
+// nor -schnell can be pulled from black-forest-labs/* without an HF token tied
+// to an account that has clicked through the license on huggingface.co.
+const FLUX_MODEL_ID = 'Niansuh/FLUX.1-schnell';
+const FLUX_SCRIPT = (p: Record<string, any>) => `
+import os, json, sys
+try:
+    import torch, shutil
+    from diffusers import FluxPipeline, FluxTransformer2DModel, BitsAndBytesConfig as DBnBConfig
+    from transformers import T5EncoderModel, BitsAndBytesConfig as TBnBConfig
+    from huggingface_hub import snapshot_download
+
+    MODEL_ID = "${FLUX_MODEL_ID}"
+
+    # This mirror ships scheduler/config.json instead of the scheduler_config.json
+    # filename diffusers expects — patch it once per cache (idempotent).
+    snap_dir = snapshot_download(MODEL_ID, allow_patterns=["scheduler/config.json"])
+    sched_cfg = os.path.join(snap_dir, "scheduler", "scheduler_config.json")
+    if not os.path.exists(sched_cfg):
+        shutil.copy(os.path.join(snap_dir, "scheduler", "config.json"), sched_cfg)
+
+    transformer_4bit = FluxTransformer2DModel.from_pretrained(
+        MODEL_ID, subfolder="transformer",
+        quantization_config=DBnBConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16),
+        torch_dtype=torch.bfloat16,
+    )
+    text_encoder_2_4bit = T5EncoderModel.from_pretrained(
+        MODEL_ID, subfolder="text_encoder_2",
+        quantization_config=TBnBConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16),
+        torch_dtype=torch.bfloat16,
+    )
+    pipe = FluxPipeline.from_pretrained(
+        MODEL_ID,
+        transformer=transformer_4bit,
+        text_encoder_2=text_encoder_2_4bit,
+        torch_dtype=torch.bfloat16,
+    )
+    pipe.enable_model_cpu_offload()
+
+    kwargs = dict(
+        prompt=${JSON.stringify(p.prompt)},
+        guidance_scale=0.0,
+        num_inference_steps=int(${p.steps}),
+        max_sequence_length=256,
+        width=int(${p.width}), height=int(${p.height}),
+    )
+    ${p.seed != null ? `kwargs["generator"] = torch.Generator("cpu").manual_seed(int(${p.seed}))` : ''}
+
+    image = pipe(**kwargs).images[0]
+    dst = ${JSON.stringify(p.dst)}
+    os.makedirs(os.path.dirname(os.path.abspath(dst)), exist_ok=True)
+    image.save(dst)
+    print("###RESULT###" + json.dumps({
+        "output": dst, "width": image.width, "height": image.height,
+        "vram_peak_mb": torch.cuda.max_memory_allocated() / 1024**2,
+    }))
+except Exception as e:
+    import traceback
+    print("###RESULT###" + json.dumps({"error": str(e), "trace": traceback.format_exc()[-800:]}))
+`;
+
 export const imageGenerateTool = {
   name: 'image_generate',
   description: [
-    'Generate an image from a text prompt using a local Stable Diffusion 1.5 model (runs on-machine GPU, no external API).',
-    'Best for quick, casual illustrations at up to ~768px. Takes a few seconds.',
+    'Generate an image from a text prompt using a local diffusion model (runs on-machine GPU, no external API).',
+    'quality="fast" (default): SDXL, ~10-20 seconds, default resolution 1024×1024, good for casual/quick illustrations.',
+    'quality="high": FLUX.1-schnell (4-bit quantized), ~40-60 seconds total (model load + generation), noticeably more photorealistic detail and prompt accuracy. Use only when the user explicitly asks for higher quality/detail/photorealism, or for a "고품질" request — otherwise default to fast.',
     'Returns the generated image inline in the chat.',
   ].join('\n'),
   schema: {
-    prompt: 'Text description of the image to generate (English works best for SD1.5)',
-    negative_prompt: 'Things to avoid in the image (optional, e.g. "blurry, low quality, deformed")',
-    width: 'Image width in pixels, multiple of 8 (default 512)',
-    height: 'Image height in pixels, multiple of 8 (default 512)',
-    steps: 'Denoising steps — more = higher quality but slower (default 25, range 10–50)',
-    guidance_scale: 'How closely to follow the prompt (default 7.5, range 1–20)',
+    prompt: 'Text description of the image to generate (English works best)',
+    quality: '"fast" (SDXL, default) or "high" (FLUX.1-schnell, much slower but noticeably better detail/realism)',
+    negative_prompt: 'Things to avoid in the image (optional, e.g. "blurry, low quality, deformed") — ignored in quality="high" mode (FLUX.1-schnell does not support it)',
+    width: 'Image width in pixels, multiple of 8 (default 1024)',
+    height: 'Image height in pixels, multiple of 8 (default 1024)',
+    steps: 'Denoising steps — more = higher quality but slower (fast mode: default 30, range 15–50; high mode: default 4, range 1–8)',
+    guidance_scale: 'How closely to follow the prompt (default 7.0, range 1–20) — fast mode only, ignored in quality="high"',
     seed: 'Random seed for reproducibility (optional)',
     output: 'Output file path (optional; defaults to a timestamped file in the workspace)',
   },
@@ -94,6 +207,7 @@ export const imageGenerateTool = {
     type: 'object',
     properties: {
       prompt: { type: 'string' },
+      quality: { type: 'string', enum: ['fast', 'high'] },
       negative_prompt: { type: 'string' },
       width: { type: 'number' },
       height: { type: 'number' },
@@ -108,36 +222,59 @@ export const imageGenerateTool = {
   execute: async (args: any): Promise<ToolResult> => {
     const prompt = String(args?.prompt || '').trim();
     if (!prompt) return { success: false, error: 'prompt is required' };
+    const isHighQuality = args?.quality === 'high';
 
     const workspacePath = getWorkspacePath(args);
     let outPath = String(args?.output || '').trim();
     if (!outPath) {
-      outPath = path.join(workspacePath, `sd15_${Date.now()}.png`);
+      outPath = path.join(workspacePath, `${isHighQuality ? 'flux' : 'sdxl'}_${Date.now()}.png`);
     } else if (!path.isAbsolute(outPath)) {
       outPath = path.resolve(workspacePath, outPath);
     }
 
-    const params = {
-      prompt,
-      negative_prompt: args?.negative_prompt || '',
-      width: Math.round((args?.width ?? 512) / 8) * 8,
-      height: Math.round((args?.height ?? 512) / 8) * 8,
-      steps: Math.min(50, Math.max(10, args?.steps ?? 25)),
-      guidance_scale: args?.guidance_scale ?? 7.5,
-      seed: args?.seed,
-      dst: outPath,
-    };
+    const negativePromptRaw = args?.negative_prompt || '';
+    const [translatedPrompt, translatedNegative] = await Promise.all([
+      translatePromptToEnglish(prompt),
+      isHighQuality ? Promise.resolve(null) : translatePromptToEnglish(negativePromptRaw),
+    ]);
 
-    const result = await runVenvPython(SD15_SCRIPT(params), 180_000);
-    if (result.error) return { success: false, error: result.error, stderr: result.trace || result.raw };
+    const width = Math.round((args?.width ?? 1024) / 16) * 16;
+    const height = Math.round((args?.height ?? 1024) / 16) * 16;
+
+    let result: any;
+    if (isHighQuality) {
+      const params = {
+        prompt: translatedPrompt || prompt,
+        width, height,
+        steps: Math.min(8, Math.max(1, args?.steps ?? 4)),
+        seed: args?.seed,
+        dst: outPath,
+      };
+      result = await runVenvPython(FLUX_SCRIPT(params), 180_000);
+    } else {
+      const params = {
+        prompt: translatedPrompt || prompt,
+        negative_prompt: translatedNegative || negativePromptRaw,
+        width, height,
+        steps: Math.min(50, Math.max(15, args?.steps ?? 30)),
+        guidance_scale: args?.guidance_scale ?? 7.0,
+        seed: args?.seed,
+        dst: outPath,
+      };
+      result = await runVenvPython(SDXL_SCRIPT(params), 180_000);
+    }
+    if (result.error || !result.output) {
+      return { success: false, error: result.error || 'Generator returned no output file', stderr: result.trace || result.raw };
+    }
 
     return {
       success: true,
       stdout: [
+        translatedPrompt ? `(translated prompt: ${translatedPrompt})` : null,
         `Generated: ${result.width} × ${result.height} px`,
         '',
         buildImageMarkdown(result.output, workspacePath),
-      ].join('\n'),
+      ].filter((line): line is string => line !== null).join('\n'),
       data: { ...result, rel_path: path.relative(workspacePath, result.output).replace(/\\/g, '/') },
     };
   },
@@ -162,6 +299,7 @@ try:
         width=int(${p.width}), height=int(${p.height}),
         num_frames=int(${p.num_frames}),
         num_inference_steps=int(${p.steps}),
+        guidance_scale=float(${p.guidance_scale}),
     ).frames[0]
 
     dst = ${JSON.stringify(p.dst)}
@@ -194,7 +332,8 @@ export const videoGenerateTool = {
     height: 'Video height in pixels, multiple of 32 (default 480)',
     num_frames: 'Number of frames — duration = num_frames / fps (default 65, ~2.7s at 24fps)',
     fps: 'Output frame rate (default 24)',
-    steps: 'Denoising steps — more = higher quality but slower (default 30, range 15–50)',
+    steps: 'Denoising steps — more = higher quality but slower (default 40, range 15–50)',
+    guidance_scale: 'How closely to follow the prompt (default 3.0, range 1–10 — LTX-Video responds best to a narrower range than SDXL)',
     output: 'Output file path (optional; defaults to a timestamped .mp4 in the workspace)',
   },
   jsonSchema: {
@@ -207,6 +346,7 @@ export const videoGenerateTool = {
       num_frames: { type: 'number' },
       fps: { type: 'number' },
       steps: { type: 'number' },
+      guidance_scale: { type: 'number' },
       output: { type: 'string' },
     },
     required: ['prompt'],
@@ -224,19 +364,28 @@ export const videoGenerateTool = {
       outPath = path.resolve(workspacePath, outPath);
     }
 
+    const negativePromptRaw = args?.negative_prompt || 'worst quality, blurry, distorted, deformed';
+    const [translatedPrompt, translatedNegative] = await Promise.all([
+      translatePromptToEnglish(prompt),
+      translatePromptToEnglish(negativePromptRaw),
+    ]);
+
     const params = {
-      prompt,
-      negative_prompt: args?.negative_prompt || 'worst quality, blurry, distorted, deformed',
+      prompt: translatedPrompt || prompt,
+      negative_prompt: translatedNegative || negativePromptRaw,
       width: Math.round((args?.width ?? 704) / 32) * 32,
       height: Math.round((args?.height ?? 480) / 32) * 32,
       num_frames: args?.num_frames ?? 65,
       fps: args?.fps ?? 24,
-      steps: Math.min(50, Math.max(15, args?.steps ?? 30)),
+      steps: Math.min(50, Math.max(15, args?.steps ?? 40)),
+      guidance_scale: Math.min(10, Math.max(1, args?.guidance_scale ?? 3.0)),
       dst: outPath,
     };
 
     const result = await runVenvPython(LTX_SCRIPT(params), 600_000);
-    if (result.error) return { success: false, error: result.error, stderr: result.trace || result.raw };
+    if (result.error || !result.output) {
+      return { success: false, error: result.error || 'Generator returned no output file', stderr: result.trace || result.raw };
+    }
 
     const relOut = path.relative(workspacePath, result.output).replace(/\\/g, '/');
     const sizeMB = (result.size_bytes / 1024 / 1024).toFixed(2);
@@ -245,10 +394,11 @@ export const videoGenerateTool = {
     return {
       success: true,
       stdout: [
+        translatedPrompt ? `(translated prompt: ${translatedPrompt})` : null,
         `Generated: ${result.width} × ${result.height} px | ${durationSec}s (${result.num_frames}f @ ${result.fps}fps) | ${sizeMB} MB`,
         '',
         `[${path.basename(result.output)}](/api/files/${relOut})`,
-      ].join('\n'),
+      ].filter((line): line is string => line !== null).join('\n'),
       data: { ...result, rel_path: relOut },
     };
   },

src/tools/web.ts

@@ -992,7 +992,7 @@ export const webFetchTool = {
   },
 };
 
-function getOllamaConfig(): { endpoint: string; model: string } {
+export function getOllamaConfig(): { endpoint: string; model: string } {
   try {
     const cm = getConfig();
     const data = cm.getConfig();

web-ui/index.html

@@ -65,6 +65,7 @@
     <button class="mode-btn" onclick="window.open('/pptx-wizard.html','_blank')" title="슬라이드 제작">📊 슬라이드</button>
     <button class="mode-btn" onclick="window.open('/language-app.html','_blank')" title="우즈베크어 학습">🇺🇿 언어</button>
     <button class="mode-btn" onclick="window.open('/music-app.html','_blank')" title="음악 스튜디오">🎵 음악</button>
+    <button class="mode-btn" onclick="window.open('/studio-app.html','_blank')" title="이미지·동영상 생성">🎨 스튜디오</button>
     <button class="mode-btn" onclick="window.open('/accountant-app.html','_blank')" title="회계 어시스턴트">📊 회계</button>
     <button class="mode-btn" onclick="window.open('/mind-app.html','_blank')" title="마음 상담실">🧠 상담</button>
     <button class="mode-btn" onclick="window.open('/lawyer-app.html','_blank')" title="법률 어시스턴트">⚖️ 법률</button>

web-ui/studio-app.html

@@ -0,0 +1,286 @@
+<!DOCTYPE html>
+<html lang="ko">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>🎨 이미지·동영상 스튜디오</title>
+<link rel="icon" type="image/png" sizes="64x64" href="cherry_logo.png">
+<link rel="stylesheet" href="styles.css">
+<script>try{var _t=localStorage.getItem('cherryclaw_theme')||'dark';document.documentElement.setAttribute('data-theme',_t);}catch(e){}</script>
+<style>
+:root{--brand:#a78bfa;}
+html,body{height:100%;margin:0;padding:0;}
+body{background:var(--bg);color:var(--text);font-family:system-ui,sans-serif;font-weight:500;display:flex;flex-direction:column;overflow:hidden;height:100dvh;}
+
+.st-hdr{display:flex;align-items:center;gap:8px;padding:6px 12px;background:var(--panel);border-bottom:1px solid var(--line);flex-shrink:0;}
+.st-hdr h1{font-size:13px;font-weight:700;margin:0;}
+.st-hdr-sp{flex:1;}
+.hdr-btn{background:none;border:1.5px solid var(--line);border-radius:6px;padding:3px 9px;font-size:11px;color:var(--muted);cursor:pointer;font-family:inherit;transition:.15s;}
+.hdr-btn:hover{border-color:var(--brand);color:var(--brand);}
+
+.st-tabs{display:flex;gap:4px;padding:8px 12px 0;flex-shrink:0;}
+.st-tab{background:none;border:none;border-bottom:2px solid transparent;padding:8px 14px;font-size:13px;font-weight:700;color:var(--muted);cursor:pointer;font-family:inherit;}
+.st-tab.active{color:var(--brand);border-bottom-color:var(--brand);}
+
+.st-layout{flex:1;display:flex;min-height:0;overflow:hidden;}
+.st-form{width:320px;flex-shrink:0;display:flex;flex-direction:column;gap:10px;padding:14px;border-right:1px solid var(--line);background:var(--panel);overflow-y:auto;}
+.st-field{display:flex;flex-direction:column;gap:4px;}
+.st-field label{font-size:11px;font-weight:700;color:var(--muted);}
+.st-field textarea,.st-field input,.st-field select{background:var(--panel-2);border:1px solid var(--line);border-radius:6px;padding:7px 8px;font-size:12px;color:var(--text);font-family:inherit;outline:none;resize:vertical;}
+.st-field textarea:focus,.st-field input:focus,.st-field select:focus{border-color:var(--brand);}
+.st-row{display:flex;gap:8px;}
+.st-row .st-field{flex:1;}
+.st-gen-btn{background:var(--brand);border:none;border-radius:8px;padding:10px;font-size:13px;font-weight:700;color:#1e1b2e;cursor:pointer;font-family:inherit;margin-top:4px;transition:.15s;}
+.st-gen-btn:hover{opacity:.9;}
+.st-gen-btn:disabled{opacity:.5;cursor:not-allowed;}
+.st-hint{font-size:10px;color:var(--muted);line-height:1.4;}
+
+.st-main{flex:1;display:flex;flex-direction:column;min-width:0;overflow:hidden;}
+.st-result{flex:1;display:flex;align-items:center;justify-content:center;padding:20px;overflow:auto;min-height:0;}
+.st-result img,.st-result video{max-width:100%;max-height:100%;border-radius:10px;box-shadow:0 4px 24px rgba(0,0,0,.25);}
+.st-placeholder{color:var(--muted);font-size:13px;text-align:center;}
+.st-spinner{width:32px;height:32px;border:3px solid var(--line);border-top-color:var(--brand);border-radius:50%;animation:st-spin 0.8s linear infinite;margin:0 auto 10px;}
+@keyframes st-spin{to{transform:rotate(360deg);}}
+.st-status{font-size:12px;color:var(--muted);margin-top:6px;}
+.st-error{color:#ef4444;font-size:12px;}
+
+.st-gallery{flex-shrink:0;border-top:1px solid var(--line);background:var(--panel);padding:8px 12px;max-height:130px;overflow-y:auto;}
+.st-gallery-hdr{font-size:10px;font-weight:700;color:var(--muted);text-transform:uppercase;letter-spacing:.05em;margin-bottom:6px;}
+.st-gallery-grid{display:flex;gap:8px;flex-wrap:wrap;}
+.st-gitem{position:relative;width:76px;height:76px;border-radius:6px;overflow:hidden;border:1px solid var(--line);cursor:pointer;flex-shrink:0;background:var(--panel-2);}
+.st-gitem img,.st-gitem video{width:100%;height:100%;object-fit:cover;}
+.st-gitem .st-gtype{position:absolute;top:2px;left:2px;font-size:9px;background:rgba(0,0,0,.6);color:#fff;padding:1px 4px;border-radius:3px;}
+.st-gitem .st-gdel{position:absolute;top:2px;right:2px;background:rgba(0,0,0,.6);color:#fff;border:none;border-radius:3px;font-size:10px;width:16px;height:16px;line-height:1;cursor:pointer;display:none;}
+.st-gitem:hover .st-gdel{display:block;}
+</style>
+</head>
+<body>
+
+<header class="st-hdr">
+  <div style="font-weight:700;font-size:13px">🎨 이미지·동영상 스튜디오</div>
+  <div class="st-hdr-sp"></div>
+  <button class="hdr-btn" onclick="toggleTheme()">🌓</button>
+  <button class="hdr-btn" onclick="window.close()">닫기</button>
+</header>
+
+<div class="st-tabs">
+  <button class="st-tab active" id="tab-image" onclick="switchTab('image')">🖼️ 이미지</button>
+  <button class="st-tab" id="tab-video" onclick="switchTab('video')">🎬 동영상</button>
+</div>
+
+<div class="st-layout">
+  <div class="st-form">
+    <div class="st-field">
+      <label>프롬프트</label>
+      <textarea id="f-prompt" rows="4" placeholder="예: a red panda reading a book, cozy lighting"></textarea>
+    </div>
+    <div class="st-field">
+      <label>네거티브 프롬프트 (선택)</label>
+      <textarea id="f-neg" rows="2" placeholder="blurry, low quality, deformed"></textarea>
+    </div>
+    <div class="st-row">
+      <div class="st-field">
+        <label>가로</label>
+        <input type="number" id="f-width" value="1024" step="8">
+      </div>
+      <div class="st-field">
+        <label>세로</label>
+        <input type="number" id="f-height" value="1024" step="8">
+      </div>
+    </div>
+    <div class="st-row">
+      <div class="st-field">
+        <label>스텝수 (품질 ↔ 속도)</label>
+        <input type="number" id="f-steps" value="30" min="10" max="50">
+      </div>
+      <div class="st-field">
+        <label>가이던스 스케일</label>
+        <input type="number" id="f-guidance" value="7.0" step="0.5">
+      </div>
+    </div>
+    <div class="st-row" id="row-image-only">
+      <div class="st-field">
+        <label>품질</label>
+        <select id="f-quality" onchange="onQualityChange()">
+          <option value="fast">빠름 (SDXL · 10~20초)</option>
+          <option value="high">고품질 (FLUX.1 · 40~60초)</option>
+        </select>
+      </div>
+      <div class="st-field">
+        <label>시드 (선택)</label>
+        <input type="number" id="f-seed" placeholder="랜덤">
+      </div>
+    </div>
+    <div class="st-row" id="row-video-only" style="display:none">
+      <div class="st-field">
+        <label>프레임수</label>
+        <input type="number" id="f-frames" value="65" step="8">
+      </div>
+      <div class="st-field">
+        <label>FPS</label>
+        <input type="number" id="f-fps" value="24">
+      </div>
+    </div>
+    <button class="st-gen-btn" id="gen-btn" onclick="generate()">생성</button>
+    <div class="st-hint" id="hint-text">SDXL 로컬 생성 · 보통 10~20초</div>
+  </div>
+
+  <div class="st-main">
+    <div class="st-result" id="result-area">
+      <div class="st-placeholder">프롬프트를 입력하고 생성 버튼을 눌러주세요</div>
+    </div>
+    <div class="st-gallery">
+      <div class="st-gallery-hdr">최근 생성 이력</div>
+      <div class="st-gallery-grid" id="gallery-grid"></div>
+    </div>
+  </div>
+</div>
+
+<script>
+const TOKEN_KEY='smallclaw_token';
+function getToken(){try{return sessionStorage.getItem(TOKEN_KEY)||localStorage.getItem(TOKEN_KEY)||'';}catch{return '';}}
+function authH(extra){const t=getToken();const h={...(extra||{})};if(t)h['Authorization']='Bearer '+t;return h;}
+async function checkAuth(){
+  if(!getToken()){location.href='/login.html?redirect='+encodeURIComponent(location.pathname);return false;}
+  try{const r=await fetch('/api/auth/status',{headers:authH()});if(!r.ok){location.href='/login.html?redirect='+encodeURIComponent(location.pathname);return false;}const d=await r.json();if(!d.authenticated){location.href='/login.html?redirect='+encodeURIComponent(location.pathname);return false;}return true;}
+  catch{location.href='/login.html?redirect='+encodeURIComponent(location.pathname);return false;}
+}
+function toggleTheme(){const d=document.documentElement;const n=d.getAttribute('data-theme')==='dark'?'light':'dark';d.setAttribute('data-theme',n);try{localStorage.setItem('cherryclaw_theme',n);}catch{}}
+
+let currentTab='image';
+function switchTab(tab){
+  currentTab=tab;
+  document.getElementById('tab-image').classList.toggle('active', tab==='image');
+  document.getElementById('tab-video').classList.toggle('active', tab==='video');
+  document.getElementById('row-image-only').style.display = tab==='image' ? 'flex' : 'none';
+  document.getElementById('row-video-only').style.display = tab==='video' ? 'flex' : 'none';
+  if(tab==='image'){
+    document.getElementById('f-width').value=1024; document.getElementById('f-height').value=1024;
+    document.getElementById('f-quality').value='fast';
+    onQualityChange();
+  } else {
+    document.getElementById('f-width').value=704; document.getElementById('f-height').value=480;
+    document.getElementById('f-steps').value=40; document.getElementById('f-guidance').value=3.0;
+    document.getElementById('hint-text').textContent='LTX-Video 로컬 생성 · 보통 30~90초, 시간이 걸립니다';
+  }
+}
+
+function onQualityChange(){
+  const q=document.getElementById('f-quality').value;
+  const stepsEl=document.getElementById('f-steps');
+  if(q==='high'){
+    stepsEl.min=1; stepsEl.max=8; stepsEl.value=4;
+    document.getElementById('f-guidance').disabled=true;
+    document.getElementById('hint-text').textContent='FLUX.1-schnell 로컬 생성(4비트 양자화) · 보통 40~60초, 시간이 걸립니다';
+  } else {
+    stepsEl.min=10; stepsEl.max=50; stepsEl.value=30;
+    document.getElementById('f-guidance').disabled=false;
+    document.getElementById('f-guidance').value=7.0;
+    document.getElementById('hint-text').textContent='SDXL 로컬 생성 · 보통 10~20초';
+  }
+}
+
+let genStartTime=0, genTimer=null;
+function startTimer(){
+  genStartTime=Date.now();
+  const btn=document.getElementById('gen-btn');
+  btn.disabled=true;
+  genTimer=setInterval(()=>{
+    const s=((Date.now()-genStartTime)/1000).toFixed(0);
+    btn.textContent='생성 중... ('+s+'초)';
+  },500);
+}
+function stopTimer(){
+  clearInterval(genTimer);
+  document.getElementById('gen-btn').disabled=false;
+  document.getElementById('gen-btn').textContent='생성';
+}
+
+async function generate(){
+  const prompt=document.getElementById('f-prompt').value.trim();
+  if(!prompt){alert('프롬프트를 입력해주세요');return;}
+  const resultArea=document.getElementById('result-area');
+  const isHighQuality = currentTab==='image' && document.getElementById('f-quality').value==='high';
+  const statusMsg = currentTab==='video' ? '동영상 생성 중... 최대 1~2분 정도 걸릴 수 있어요'
+    : isHighQuality ? '고품질(FLUX.1) 이미지 생성 중... 최대 1분 정도 걸릴 수 있어요'
+    : '이미지 생성 중...';
+  resultArea.innerHTML='<div style="text-align:center"><div class="st-spinner"></div><div class="st-status">'+statusMsg+'</div></div>';
+  startTimer();
+
+  const body={
+    prompt,
+    negative_prompt: document.getElementById('f-neg').value.trim(),
+    width: parseInt(document.getElementById('f-width').value)||undefined,
+    height: parseInt(document.getElementById('f-height').value)||undefined,
+    steps: parseInt(document.getElementById('f-steps').value)||undefined,
+    guidance_scale: parseFloat(document.getElementById('f-guidance').value)||undefined,
+  };
+  if(currentTab==='image'){
+    body.quality=document.getElementById('f-quality').value;
+    const seed=document.getElementById('f-seed').value;
+    if(seed) body.seed=parseInt(seed);
+  } else {
+    body.num_frames=parseInt(document.getElementById('f-frames').value)||undefined;
+    body.fps=parseInt(document.getElementById('f-fps').value)||undefined;
+  }
+
+  try{
+    const endpoint = currentTab==='image' ? '/api/imagegen/generate-image' : '/api/imagegen/generate-video';
+    const r=await fetch(endpoint,{method:'POST',headers:authH({'Content-Type':'application/json'}),body:JSON.stringify(body)});
+    const d=await r.json();
+    stopTimer();
+    if(!r.ok || !d.success){
+      resultArea.innerHTML='<div class="st-error">생성 실패: '+(d.error||'알 수 없는 오류')+'</div>';
+      return;
+    }
+    if(currentTab==='image'){
+      resultArea.innerHTML='<img src="'+d.url+'" alt="generated">';
+    } else {
+      resultArea.innerHTML='<video src="'+d.url+'" controls autoplay loop></video>';
+    }
+    loadGallery();
+  }catch(e){
+    stopTimer();
+    resultArea.innerHTML='<div class="st-error">요청 실패: '+e.message+'</div>';
+  }
+}
+
+async function loadGallery(){
+  try{
+    const r=await fetch('/api/imagegen/gallery',{headers:authH()});
+    const d=await r.json();
+    const grid=document.getElementById('gallery-grid');
+    grid.innerHTML='';
+    for(const item of (d.items||[])){
+      const el=document.createElement('div');
+      el.className='st-gitem';
+      const media = item.type==='image'
+        ? '<img src="'+item.url+'">'
+        : '<video src="'+item.url+'" muted></video>';
+      el.innerHTML=media+'<span class="st-gtype">'+(item.type==='image'?'🖼️':'🎬')+'</span><button class="st-gdel" title="삭제">✕</button>';
+      el.querySelector('.st-gitem > img, .st-gitem > video')?.addEventListener('click',()=>{
+        document.getElementById('result-area').innerHTML = item.type==='image'
+          ? '<img src="'+item.url+'">' : '<video src="'+item.url+'" controls autoplay loop></video>';
+      });
+      el.addEventListener('click',(ev)=>{ if(ev.target.tagName!=='BUTTON'){
+        document.getElementById('result-area').innerHTML = item.type==='image'
+          ? '<img src="'+item.url+'">' : '<video src="'+item.url+'" controls autoplay loop></video>';
+      }});
+      el.querySelector('.st-gdel').addEventListener('click', async (ev)=>{
+        ev.stopPropagation();
+        if(!confirm(item.name+' 삭제할까요?'))return;
+        await fetch('/api/imagegen/gallery/'+encodeURIComponent(item.name),{method:'DELETE',headers:authH()});
+        loadGallery();
+      });
+      grid.appendChild(el);
+    }
+  }catch(e){}
+}
+
+(async function init(){
+  const ok=await checkAuth();if(!ok)return;
+  loadGallery();
+})();
+</script>
+</body>
+</html>