dphn/Dolphin3.0-Mistral-24B is the ungated mirror of the Dolphin 3.0 Mistral 24B — exactly what you asked for. It's ~48GB fp16, which needs GPU+CPU split (device_map="auto" with 32GB on GPU, ~16GB in RAM). Let me kick off the download and update the service in parallel.

tour-comfy/bootstrap.sh

@@ -1,64 +0,0 @@
-#!/bin/bash
-# One-time (idempotent) host setup for the Qwen-Image-Edit service.
-# Runs as the service user (NO sudo). Safe to re-run: existing pieces are skipped.
-#
-# Builds, under the project BASE (the parent of this api/ dir):
-#   venv/        torch 2.3.1+rocm5.7 + ComfyUI deps   (gfx906 / ROCm 5.7)
-#   ComfyUI/     pinned to v0.3.77 + ComfyUI-GGUF custom node
-#   ComfyUI/models/{unet,text_encoders,vae}/  the v23 Q8 GGUF + encoder + vae
-set -e
-
-source "$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )/env.sh"
-GGUF_NODE="$COMFY/custom_nodes/ComfyUI-GGUF"
-COMFY_TAG="v0.3.77"   # newest tag that runs on torch 2.3.1 (no comfy_kitchen)
-
-echo "[bootstrap] BASE=$BASE  VENV=$VENV"
-
-# --- ComfyUI (pinned) -------------------------------------------------------
-if [ ! -d "$COMFY/.git" ]; then
-  echo "[bootstrap] cloning ComfyUI @ $COMFY_TAG ..."
-  git clone --depth 1 --branch "$COMFY_TAG" \
-      https://github.com/comfyanonymous/ComfyUI.git "$COMFY"
-fi
-
-# --- venv + python deps -----------------------------------------------------
-if [ ! -d "$VENV" ]; then
-  echo "[bootstrap] creating venv at $VENV ..."
-  python3 -m venv "$VENV"
-fi
-source "$VENV/bin/activate"
-python -m pip install --upgrade pip wheel
-echo "[bootstrap] installing torch (rocm5.7) ..."
-pip install torch==2.3.1+rocm5.7 torchvision==0.18.1+rocm5.7 \
-    --index-url https://download.pytorch.org/whl/rocm5.7
-echo "[bootstrap] installing ComfyUI requirements ..."
-pip install -r "$COMFY/requirements.txt"
-
-# --- ComfyUI-GGUF custom node ----------------------------------------------
-if [ ! -d "$GGUF_NODE" ]; then
-  echo "[bootstrap] cloning ComfyUI-GGUF ..."
-  git clone --depth 1 https://github.com/city96/ComfyUI-GGUF.git "$GGUF_NODE"
-fi
-pip install -r "$GGUF_NODE/requirements.txt" || pip install gguf
-
-# --- API deps ---------------------------------------------------------------
-pip install fastapi "uvicorn[standard]" websocket-client python-multipart pillow requests
-
-# --- models (resume-safe; skipped if already complete) ----------------------
-M="$COMFY/models"
-mkdir -p "$M/unet" "$M/text_encoders" "$M/vae"
-dl() {  # url  dest
-  if [ -s "$2" ]; then echo "[bootstrap] have $(basename "$2")"; return; fi
-  echo "[bootstrap] downloading $(basename "$2") ..."
-  wget -c -q -O "$2" "$1"
-}
-dl "https://huggingface.co/Novice25/Qwen-Image-Edit-Rapid-AIO-GGUF/resolve/main/v23/Qwen-Rapid-NSFW-v23_Q8_0.gguf" \
-   "$M/unet/Qwen-Rapid-NSFW-v23_Q8_0.gguf"
-dl "https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors" \
-   "$M/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors"
-dl "https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors" \
-   "$M/vae/qwen_image_vae.safetensors"
-
-echo "[bootstrap] verifying torch + GPU ..."
-python -c "import torch; print('torch', torch.__version__, 'cuda', torch.cuda.is_available(), torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'NO GPU')"
-echo "[bootstrap] BOOTSTRAP_DONE"

tour-comfy/car.html

@@ -1459,6 +1459,13 @@
             font-size: 9px; font-weight: 700; border-radius: 50%; width: 15px; height: 15px;
             display: flex; align-items: center; justify-content: center; line-height: 1;
         }
+        .sb-gen-refs .sb-gen-ref-x {
+            position: absolute; top: -4px; right: -4px; background: #dc2626; color: #fff;
+            font-size: 11px; font-weight: 700; border-radius: 50%; width: 15px; height: 15px;
+            display: none; align-items: center; justify-content: center; line-height: 1;
+        }
+        .sb-gen-refs .sb-gen-ref:hover .sb-gen-ref-x { display: flex; }
+        .sb-gen-refs .sb-gen-ref:hover img { border-color: #dc2626; }
         #studioAngleBar {
             position: absolute; bottom: 8px; left: 50%; transform: translateX(-50%);
             display: flex; gap: 4px; opacity: 0; transition: opacity 0.2s;
@@ -1490,7 +1497,9 @@
         }
         .sb-template-card:hover { border-color: #7c3aed; transform: scale(1.02); }
         .sb-template-card.selected { border-color: #22c55e; }
-        .sb-template-card video { width:100%; height:100%; object-fit:cover; pointer-events:none; }
+        .sb-template-card video,
+        .sb-template-card img { width:100%; height:100%; object-fit:cover; pointer-events:none; }
+        .sb-template-card video { position:absolute; inset:0; }
         .sb-template-label {
             position: absolute; bottom: 0; left: 0; right: 0;
             background: linear-gradient(transparent,rgba(0,0,0,0.82));
@@ -1845,11 +1854,13 @@
                             <div class="sb-actions" style="margin-bottom:8px">
                                 <button class="sb-btn" id="lbExtract" style="display:none" onclick="extractCurrentImage()">Extract</button>
                                 <button class="sb-btn" id="lbNoBgBtn" onclick="lbRemoveBg()">No BG</button>
+                                <button class="sb-btn" id="lbInvertAlphaBtn" onclick="lbInvertAlpha()" title="Invert transparency — fixes when the wrong segment was kept">Invert α</button>
                                 <button class="sb-btn" id="lbCropBtn" style="display:none" onclick="lbAutoCrop()" title="Crop away transparent border">Crop</button>
                                 <button class="sb-btn" id="lbDuplicateBtn" onclick="lbDuplicate()" title="Duplicate image into same group">Duplicate</button>
                                 <button class="sb-btn" id="lbManualCropBtn" onclick="startManualCrop()" title="Drag to crop image">Crop…</button>
                                 <button class="sb-btn" id="lbRotateLeftBtn" onclick="lbRotate(-90)" title="Rotate 90° counter-clockwise">⟲ 90°</button>
                                 <button class="sb-btn" id="lbRotateRightBtn" onclick="lbRotate(90)" title="Rotate 90° clockwise">⟳ 90°</button>
+                                <button class="sb-btn" id="lbPoseBtn" onclick="lbTogglePose()" title="Preview body pose skeleton">Pose</button>
                                 <a      class="sb-btn" id="lbDownloadBtn" download style="text-decoration:none">Download</a>
                                 <button class="sb-btn" id="lbArchiveBtn" onclick="lbArchive()" style="color:#f59e0b" title="Move to archive (recoverable)">Archive</button>
                                 <button class="sb-btn danger" id="lbDeleteBtn" onclick="lbDeleteArm()" title="Click once to arm, then again to confirm">Delete</button>
@@ -2330,6 +2341,13 @@
             document.getElementById('studio').classList.remove('open');
             document.body.style.overflow = '';
             stopSlideshow();
+            // Release every streaming/decoding video so closing the studio frees memory
+            // and stops background network streams (template clips + scenery preview).
+            _tplStopAll();
+            ['sceneVideoEl', 'lbVideo', 'cltVideoPreview'].forEach(id => {
+                const v = document.getElementById(id);
+                if (v) { v.pause(); v.removeAttribute('src'); v.load(); }
+            });
         }
 
         function toggleSlideshow() {
@@ -2391,6 +2409,8 @@
 
         function updateStudio() {
             const fname = lbNames[lbIdx];
+            if (document.getElementById('poseCanvas')) _removePoseOverlay();  // skeleton is per-image
+            document.getElementById('poseResults')?.remove();
             _fsModelFilename = fname;  // keep faceswap/scenery/segment in sync
             const isVid = isVideo(fname) || fileContentType[fname] === 'video';
             const lbImgEl   = document.getElementById('lbImg');
@@ -2442,6 +2462,11 @@
             if (faceswapBtn)faceswapBtn.style.display = !isVid ? '' : 'none';
             const cropBtn = document.getElementById('lbCropBtn');
             if (cropBtn) cropBtn.style.display = (!isVid && fileHasBg[fname] === false) ? '' : 'none';
+            // Invert α is only meaningful on a transparent image (after background removal).
+            const invAlphaBtn = document.getElementById('lbInvertAlphaBtn');
+            if (invAlphaBtn) invAlphaBtn.style.display = (!isVid && fileHasBg[fname] === false) ? '' : 'none';
+            const poseBtn = document.getElementById('lbPoseBtn');
+            if (poseBtn) poseBtn.style.display = isVid ? 'none' : '';
             const archiveBtn = document.getElementById('lbArchiveBtn');
             if (archiveBtn) archiveBtn.textContent = fileArchived[fname] ? 'Restore' : 'Archive';
 
@@ -2767,7 +2792,6 @@
                     showToast('Rotated', 'success');
                     lbUrls[lbIdx] = IMAGE_FOLDER + fname + '?t=' + Date.now();
                     updateStudio();
-                    refreshNow();
                 } else {
                     showToast('Rotate failed: ' + await r.text(), 'error');
                 }
@@ -2775,6 +2799,239 @@
             btns.forEach(b => { if (b) b.disabled = false; });
         }
 
+        // --- body-pose skeleton preview (interactive) ---
+        const POSE_MIN_SCORE = 0.3;
+        const POSE_KP_COLORS = ['#ef4444','#f97316','#eab308','#22c55e','#06b6d4','#3b82f6','#a855f7'];
+
+        function _removePoseOverlay() {
+            document.getElementById('poseCanvas')?.remove();
+            document.getElementById('poseToolbar')?.remove();
+            window._poseState = null;
+            const btn = document.getElementById('lbPoseBtn');
+            if (btn) btn.classList.remove('primary');
+        }
+
+        async function lbTogglePose() {
+            if (document.getElementById('poseCanvas')) { _removePoseOverlay(); return; }
+            const fname = lbNames[lbIdx];
+            if (!fname || !/\.(png|jpg|jpeg|webp)$/i.test(fname)) { showToast('Select an image', 'info'); return; }
+            const btn = document.getElementById('lbPoseBtn');
+            if (btn) { btn.disabled = true; btn.textContent = '…'; }
+            try {
+                const r = await fetch(`${API}/images/${encodeURIComponent(fname)}/pose`, { method: 'POST' });
+                if (r.status === 501) {
+                    showToast('Pose estimator not installed on the server (pip install rtmlib onnxruntime)', 'info', 6000);
+                } else if (r.ok) {
+                    const d = await r.json();
+                    if (!d.people || !d.people.length) showToast('No person detected', 'info');
+                    else { _initPoseOverlay(d, fname); showToast(`Pose (${d.backend}) — drag joints to edit`, 'success'); }
+                } else {
+                    showToast('Pose failed: ' + await r.text(), 'error');
+                }
+            } catch (e) { showToast('Pose failed: ' + e, 'error'); }
+            if (btn) { btn.disabled = false; btn.textContent = 'Pose'; }
+        }
+
+        function _initPoseOverlay(data, fname) {
+            _removePoseOverlay();
+            const viewer = document.getElementById('studioViewer');
+            const img    = document.getElementById('lbImg');
+            if (!viewer || !img) return;
+            const canvas = document.createElement('canvas');
+            canvas.id = 'poseCanvas';
+            canvas.style.cssText = 'position:absolute;top:0;left:0;width:100%;height:100%;cursor:grab;z-index:90;';
+            viewer.appendChild(canvas);
+            canvas.width  = viewer.clientWidth;
+            canvas.height = viewer.clientHeight;
+
+            // object-fit:contain letterbox mapping image px ↔ canvas px.
+            const iW = data.width, iH = data.height;
+            const scale = Math.min(canvas.width / iW, canvas.height / iH);
+            const st = window._poseState = {
+                fname, canvas, ctx: canvas.getContext('2d'),
+                people: data.people.map(p => p.map(k => k.slice())),  // deep copy (editable)
+                skeleton: data.skeleton, width: iW, height: iH,
+                scale, offX: (canvas.width - iW * scale) / 2, offY: (canvas.height - iH * scale) / 2,
+                drag: null,
+            };
+
+            // Floating toolbar (top-left so it doesn't fight the crop bar on the right).
+            const bar = document.createElement('div');
+            bar.id = 'poseToolbar';
+            bar.style.cssText = 'position:absolute;top:8px;left:8px;display:flex;gap:6px;z-index:102;';
+            bar.innerHTML =
+                '<button class="sb-btn" onclick="lbPoseReset()" title="Re-detect the original pose">Reset</button>'
+              + '<button class="sb-btn primary" onclick="lbFindSimilarPose()" title="Find library images in this pose">Similar pose</button>';
+            viewer.appendChild(bar);
+
+            canvas.addEventListener('mousedown', _posePointerDown);
+            canvas.addEventListener('mousemove', _posePointerMove);
+            window.addEventListener('mouseup', _posePointerUp);
+            _drawPose();
+            const btn = document.getElementById('lbPoseBtn');
+            if (btn) btn.classList.add('primary');
+        }
+
+        function _drawPose() {
+            const st = window._poseState; if (!st) return;
+            const { ctx, canvas, scale, offX, offY } = st;
+            const cx = x => offX + x * scale, cy = y => offY + y * scale;
+            ctx.clearRect(0, 0, canvas.width, canvas.height);
+            st.people.forEach((kpts, pi) => {
+                const color = POSE_KP_COLORS[pi % POSE_KP_COLORS.length];
+                ctx.lineWidth = 3; ctx.strokeStyle = color;
+                st.skeleton.forEach(([a, b]) => {
+                    const pa = kpts[a], pb = kpts[b];
+                    if (!pa || !pb || pa[2] < POSE_MIN_SCORE || pb[2] < POSE_MIN_SCORE) return;
+                    ctx.beginPath(); ctx.moveTo(cx(pa[0]), cy(pa[1])); ctx.lineTo(cx(pb[0]), cy(pb[1])); ctx.stroke();
+                });
+                kpts.forEach((p, ki) => {
+                    if (!p || p[2] < POSE_MIN_SCORE) return;
+                    const hot = st.drag && st.drag.pi === pi && st.drag.ki === ki;
+                    ctx.beginPath(); ctx.arc(cx(p[0]), cy(p[1]), hot ? 7 : 5, 0, Math.PI * 2);
+                    ctx.fillStyle = '#fff'; ctx.fill();
+                    ctx.lineWidth = 2; ctx.strokeStyle = color; ctx.stroke();
+                });
+            });
+        }
+
+        function _poseHitTest(mx, my) {
+            const st = window._poseState; if (!st) return null;
+            const cx = x => st.offX + x * st.scale, cy = y => st.offY + y * st.scale;
+            let best = null, bestD = 14 * 14;  // ~14px grab radius
+            st.people.forEach((kpts, pi) => kpts.forEach((p, ki) => {
+                if (!p || p[2] < POSE_MIN_SCORE) return;
+                const dx = mx - cx(p[0]), dy = my - cy(p[1]), d = dx * dx + dy * dy;
+                if (d < bestD) { bestD = d; best = { pi, ki }; }
+            }));
+            return best;
+        }
+
+        function _poseEvtXY(e) {
+            const r = window._poseState.canvas.getBoundingClientRect();
+            return [e.clientX - r.left, e.clientY - r.top];
+        }
+        function _posePointerDown(e) {
+            const st = window._poseState; if (!st) return;
+            const [mx, my] = _poseEvtXY(e);
+            const hit = _poseHitTest(mx, my);
+            if (hit) { st.drag = hit; st.canvas.style.cursor = 'grabbing'; _drawPose(); e.preventDefault(); }
+        }
+        function _posePointerMove(e) {
+            const st = window._poseState; if (!st) return;
+            if (!st.drag) {
+                const [mx, my] = _poseEvtXY(e);
+                st.canvas.style.cursor = _poseHitTest(mx, my) ? 'grab' : 'default';
+                return;
+            }
+            const [mx, my] = _poseEvtXY(e);
+            const p = st.people[st.drag.pi][st.drag.ki];
+            p[0] = Math.max(0, Math.min(st.width,  (mx - st.offX) / st.scale));
+            p[1] = Math.max(0, Math.min(st.height, (my - st.offY) / st.scale));
+            if (p[2] < POSE_MIN_SCORE) p[2] = 1.0;  // dragging a hidden joint makes it visible
+            _drawPose();
+        }
+        function _posePointerUp() {
+            const st = window._poseState; if (!st || !st.drag) return;
+            st.drag = null; st.canvas.style.cursor = 'grab'; _drawPose();
+        }
+
+        async function lbPoseReset() {
+            const st = window._poseState; if (!st) return;
+            const fname = st.fname;
+            try {
+                const r = await fetch(`${API}/images/${encodeURIComponent(fname)}/pose`, { method: 'POST' });
+                if (r.ok) { const d = await r.json(); if (d.people?.length) _initPoseOverlay(d, fname); }
+            } catch (e) { showToast('Reset failed: ' + e, 'error'); }
+        }
+
+        async function lbFindSimilarPose() {
+            const st = window._poseState; if (!st) return;
+            // Use the (possibly edited) primary skeleton.
+            const kpts = st.people[0];
+            const btn = event?.target;
+            if (btn) btn.disabled = true;
+            try {
+                const r = await fetch(`${API}/pose/similar`, {
+                    method: 'POST', headers: { 'Content-Type': 'application/json' },
+                    body: JSON.stringify({ keypoints: kpts, width: st.width, height: st.height, limit: 12 }),
+                });
+                if (r.ok) {
+                    const d = await r.json();
+                    _renderPoseResults(d.similar || []);
+                } else if (r.status === 400) {
+                    showToast('Pose too sparse — drag more joints into place', 'info');
+                } else {
+                    showToast('Similar failed: ' + await r.text(), 'error');
+                }
+            } catch (e) { showToast('Similar failed: ' + e, 'error'); }
+            if (btn) btn.disabled = false;
+        }
+
+        function _renderPoseResults(items) {
+            document.getElementById('poseResults')?.remove();
+            const viewer = document.getElementById('studioViewer');
+            if (!viewer) return;
+            const panel = document.createElement('div');
+            panel.id = 'poseResults';
+            panel.style.cssText = 'position:absolute;left:0;right:0;bottom:0;z-index:103;background:rgba(0,0,0,0.85);'
+                + 'padding:8px;display:flex;gap:8px;overflow-x:auto;align-items:center;';
+            if (!items.length) {
+                panel.innerHTML = '<span style="color:#aaa;font-size:12px;flex:1">No similar poses found — try building the pose index.</span>'
+                    + '<button class="sb-btn" onclick="lbBuildPoseIndex()">Build index</button>'
+                    + '<button class="sb-btn" onclick="document.getElementById(\'poseResults\')?.remove()">Close</button>';
+            } else {
+                panel.innerHTML = '<span style="color:#aaa;font-size:11px;flex-shrink:0">Similar poses:</span>'
+                    + items.map(it => {
+                        const u = IMAGE_FOLDER + it.filename + '?t=' + Date.now();
+                        return `<div title="dist ${it.distance}" style="flex-shrink:0;cursor:pointer;text-align:center"
+                            onclick="openPoseResult('${(it.group_id||'').replace(/'/g,"\\'")}','${it.filename.replace(/'/g,"\\'")}')">
+                            <img src="${u}" loading="lazy" style="width:64px;height:64px;object-fit:cover;border-radius:5px;border:1px solid #444" onerror="this.style.opacity='0.3'">
+                            <div style="font-size:9px;color:#777">${it.distance}</div></div>`;
+                    }).join('')
+                    + '<button class="sb-btn" style="flex-shrink:0" onclick="document.getElementById(\'poseResults\')?.remove()">Close</button>';
+            }
+            viewer.appendChild(panel);
+        }
+
+        function openPoseResult(gid, fname) {
+            _removePoseOverlay();
+            document.getElementById('poseResults')?.remove();
+            if (gid && groupData.has(gid)) {
+                const i = groupData.get(gid).names.indexOf(fname);
+                openStudio(gid, i >= 0 ? i : 0);   // openStudio sets lbUrls/lbNames/lbIdx + updateStudio
+            } else {
+                const i = lbNames.indexOf(fname);
+                if (i >= 0) { lbIdx = i; updateStudio(); }
+                else showToast('Open its group from the gallery: ' + fname, 'info', 5000);
+            }
+        }
+
+        async function lbBuildPoseIndex() {
+            try {
+                const r = await fetch(`${API}/pose/index`, { method: 'POST' });
+                if (r.status === 501) { showToast('Pose estimator not installed', 'info'); return; }
+                if (!r.ok) { showToast('Index failed: ' + await r.text(), 'error'); return; }
+                showToast('Building pose index…', 'info');
+                _pollPoseIndex();
+            } catch (e) { showToast('Index failed: ' + e, 'error'); }
+        }
+
+        async function _pollPoseIndex() {
+            try {
+                const r = await fetch(`${API}/pose/index/status`);
+                if (r.ok) {
+                    const s = await r.json();
+                    if (s.running) {
+                        showToast(`Pose index: ${s.done}/${s.total}…`, 'info', 2500);
+                        setTimeout(_pollPoseIndex, 2000);
+                    } else {
+                        showToast(`Pose index ready (${s.indexed} images)`, 'success');
+                    }
+                }
+            } catch (e) { /* stop polling */ }
+        }
+
         function startManualCrop() {
             const fname = lbNames[lbIdx];
             if (!fname || !/\.(png|jpg|jpeg|webp)$/i.test(fname)) { showToast('Select an image to crop', 'info'); return; }
@@ -2930,23 +3187,54 @@
 
         async function lbRemoveBg() {
             const fname = lbNames[lbIdx];
-            showToast(`Removing background for ${fname}...`);
+            const btn = document.getElementById('lbNoBgBtn');
+            if (btn) { btn.disabled = true; btn.textContent = 'Removing…'; }
+            showToast('Removing background…');
             try {
                 const r = await fetch(`${API}/remove-background/${fname}`, { method: 'POST' });
                 if (r.ok) {
-                    showToast(`Background removed for ${fname}`, 'success');
-                    // Refresh current image in lightbox by adding timestamp
-                    const img = document.getElementById('lbImg');
-                    img.src = img.src.split('?')[0] + '?t=' + Date.now();
-                    refreshNow();
+                    showToast('Background removed', 'success');
+                    fileHasBg[fname] = false;
+                    _bustStudioImage(fname);  // cache-busts viewer + calls updateStudio()
                 } else {
                     showToast(`Failed to remove background for ${fname}`, 'error');
                 }
             } catch (e) {
                 showToast(`Failed to remove background: ${e}`, 'error');
+            } finally {
+                if (btn) { btn.disabled = false; btn.textContent = 'No BG'; }
             }
         }
 
+        async function lbInvertAlpha() {
+            const fname = lbNames[lbIdx];
+            if (!fname || !/\.(png|webp)$/i.test(fname)) { showToast('Invert needs a transparent PNG', 'info'); return; }
+            const btn = document.getElementById('lbInvertAlphaBtn');
+            if (btn) btn.disabled = true;
+            try {
+                const r = await fetch(`${API}/images/${encodeURIComponent(fname)}/invert-alpha`, { method: 'POST' });
+                if (r.ok) {
+                    showToast('Transparency inverted', 'success');
+                    fileHasBg[fname] = false;
+                    _bustStudioImage(fname);
+                    refreshNow();
+                } else {
+                    showToast('Invert failed: ' + await r.text(), 'error');
+                }
+            } catch (e) { showToast('Invert failed: ' + e, 'error'); }
+            if (btn) btn.disabled = false;
+        }
+
+        // Force the studio image (filmstrip + main viewer) to reload from disk.
+        function _bustStudioImage(fname) {
+            const t = Date.now();
+            const i = lbNames.indexOf(fname);
+            if (i >= 0) lbUrls[i] = IMAGE_FOLDER + fname + '?t=' + t;
+            const img = document.getElementById('lbImg');
+            if (img && lbNames[lbIdx] === fname) img.src = IMAGE_FOLDER + fname + '?t=' + t;
+            updateStudio();
+        }
+
         async function deleteGroup(el, event) {
             if (event) event.stopPropagation();
             const gid = el.closest('.image-card').dataset.group;
@@ -3018,8 +3306,9 @@
                     showToast(`Error uploading ${fileName}: ${e}`, 'error');
                 }
             }
-            // Trigger refresh
-            refreshNow();
+            // The source image is registered by a background task slightly after /upload
+            // returns; one debounced refresh picks it up without piling up concurrent loads.
+            setTimeout(refreshNow, 800);
         }
 
         // --- clipboard paste support ---
@@ -3131,6 +3420,8 @@
         }
 
         async function loadImages() {
+            if (loadImages._inFlight) return;
+            loadImages._inFlight = true;
             const gallery = document.getElementById('gallery');
             const statusDot = document.getElementById('statusDot');
             statusDot.classList.add('updating');
@@ -3204,6 +3495,7 @@
                         </div>`;
                 }
                 statusDot.classList.remove('updating');
+                loadImages._inFlight = false;
                 return;
             }
 
@@ -3311,14 +3603,27 @@
                 if (fresh) {
                     const curName = lbNames[lbIdx];
                     const newIdx  = fresh.names.indexOf(curName);
+                    // Snapshot ref filenames before overwriting lbNames
+                    const oldRefNames = _sbRefIndices.map(i => lbNames[i]).filter(Boolean);
                     lbUrls  = fresh.urls;
                     lbNames = fresh.names;
-                    lbIdx   = newIdx >= 0 ? newIdx : Math.min(lbIdx, fresh.names.length - 1);
+                    // Keep lbIdx stable when current image not yet in fresh data
+                    // (e.g. duplicate just created — server JSON may lag by the debounce window).
+                    if (newIdx >= 0) {
+                        lbIdx = newIdx;
+                    } else if (lbIdx >= fresh.names.length) {
+                        lbIdx = fresh.names.length - 1;
+                    }
+                    // Re-map multi-ref selections to new positions by filename
+                    _sbRefIndices = oldRefNames
+                        .map(n => fresh.names.indexOf(n))
+                        .filter(i => i >= 0);
                     updateStudio();
                 }
             }
 
             statusDot.classList.remove('updating');
+            loadImages._inFlight = false;
         }
 
         // Discover images by trying common patterns (fallback for file:// protocol)
@@ -3365,8 +3670,13 @@
             if (!_isStudioOpen()) refreshNow();
         }
 
+        // Debounced refresh — coalesces multiple rapid calls into one loadImages() run.
+        // Direct callers that need a response (pollJob completion, etc.) still use
+        // loadImages() directly; everything else should go through refreshNow().
+        let _refreshTimer = null;
         function refreshNow() {
-            loadImages();
+            clearTimeout(_refreshTimer);
+            _refreshTimer = setTimeout(loadImages, 80);
         }
 
         const API = 'http://127.0.0.1:8500';
@@ -3711,7 +4021,7 @@
             const data = groupData.get(lbCurrentGid);
             if (!data) return;
             const filename = data.names[lbIdx];
-            const btn = document.querySelector('#lightbox .lb-btn[onclick="tagCurrentImage()"]');
+            const btn = document.querySelector('button[onclick="tagCurrentImage()"]');
             if (btn) btn.textContent = '…';
             try {
                 const r = await fetch(`${API}/tag`, {
@@ -3732,7 +4042,7 @@
                     showToast(`Tag failed: ${err}`, 'info');
                 }
             } catch (e) { showToast('API not reachable', 'info'); }
-            if (btn) btn.textContent = 'Tag';
+            if (btn) btn.textContent = 'Re-tag';
         }
 
         async function extractCurrentImage() {
@@ -3840,17 +4150,24 @@
                 }
                 if (!document.getElementById('studio').classList.contains('open')) return;
                 if (tag === 'INPUT' || tag === 'TEXTAREA') return;
-                if (e.key === 'ArrowLeft')  { lbNav(-1); e.preventDefault(); }
-                if (e.key === 'ArrowRight') { lbNav(1);  e.preventDefault(); }
-                if (e.key === 'Home') { 
-                    lbIdx = 0; 
-                    updateStudio(); 
-                    e.preventDefault(); 
+                // Throttle filmstrip navigation — ignore if last nav was < 80 ms ago.
+                if (e.key === 'ArrowLeft' || e.key === 'ArrowRight') {
+                    const now = Date.now();
+                    if (now - (document._lbNavLast || 0) >= 80) {
+                        document._lbNavLast = now;
+                        lbNav(e.key === 'ArrowLeft' ? -1 : 1);
+                    }
+                    e.preventDefault();
                 }
-                if (e.key === 'End') { 
-                    lbIdx = lbUrls.length - 1; 
-                    updateStudio(); 
-                    e.preventDefault(); 
+                if (e.key === 'Home') {
+                    lbIdx = 0;
+                    updateStudio();
+                    e.preventDefault();
+                }
+                if (e.key === 'End') {
+                    lbIdx = lbUrls.length - 1;
+                    updateStudio();
+                    e.preventDefault();
                 }
                 if (e.key === ' ') {
                     const vid = document.getElementById('lbVideo');
@@ -3859,22 +4176,24 @@
                         e.preventDefault();
                     }
                 }
-                if ((e.key === 'h' || e.key === 'H') && tag !== 'INPUT') { lbToggleHidden(); }
-                if (e.key === 'F' || e.key === 'f') { 
-                    lbSetPreferred(); 
-                    e.preventDefault(); 
+                // Guard async actions against concurrent executions via a simple inflight flag.
+                if ((e.key === 'h' || e.key === 'H') && tag !== 'INPUT') {
+                    if (!document._lbHideInFlight) { document._lbHideInFlight = true; lbToggleHidden().finally(() => { document._lbHideInFlight = false; }); }
                 }
-                if (e.key === 'Delete') { 
-                    lbArchive(); 
-                    e.preventDefault(); 
+                if (e.key === 'F' || e.key === 'f') {
+                    lbSetPreferred();
+                    e.preventDefault();
                 }
-                if (e.key === 'ArrowUp') { 
-                    lbMoveInGroup(-1); 
-                    e.preventDefault(); 
+                if (e.key === 'Delete') {
+                    lbArchive();
+                    e.preventDefault();
                 }
-                if (e.key === 'ArrowDown') { 
-                    lbMoveInGroup(1); 
-                    e.preventDefault(); 
+                // ArrowUp/Down move images within the group — debounce rapid keypresses.
+                if (e.key === 'ArrowUp' || e.key === 'ArrowDown') {
+                    clearTimeout(document._lbMoveTimer);
+                    const d = e.key === 'ArrowUp' ? -1 : 1;
+                    document._lbMoveTimer = setTimeout(() => lbMoveInGroup(d), 120);
+                    e.preventDefault();
                 }
             });
         });
@@ -3955,7 +4274,7 @@
 
         async function setAsPreferred(filename, btn) {
             try {
-                const r = await fetch(`/images/${encodeURIComponent(filename)}/set-preferred`, {method:'POST'});
+                const r = await fetch(`${API}/images/${encodeURIComponent(filename)}/set-preferred`, {method:'POST'});
                 if (!r.ok) { console.error('set-preferred failed', r.status); return; }
                 const data = await r.json();
                 // Update sort orders locally
@@ -4008,6 +4327,7 @@
         }
 
         function switchSidebarTab(tab, initial) {
+            _tplStopAll();  // release any hovered template video before re-rendering panels
             _activeSidebarTab = tab;
             localStorage.setItem('studioSidebarTab', tab);
             document.querySelectorAll('#studioSidebar .sb-tab').forEach(t =>
@@ -4143,9 +4463,9 @@
                     const fn = lbNames[idx] || '';
                     const u  = lbUrls[idx] || '';
                     const thumb = isVideo(fn) ? posterFor(u) : u;
-                    return `<div class="sb-gen-ref" title="${escHtml(fn)}"><img src="${thumb}" loading="lazy" onerror="this.style.opacity='0.3'"><div class="sb-gen-ref-badge">${pos+1}</div></div>`;
+                    return `<div class="sb-gen-ref" title="Click to remove · ${escHtml(fn)}" style="cursor:pointer" onclick="sbRemoveRef(${idx})"><img src="${thumb}" loading="lazy" onerror="this.style.opacity='0.3'"><div class="sb-gen-ref-badge">${pos+1}</div><div class="sb-gen-ref-x">×</div></div>`;
                 }).join('');
-                html += `<div class="sb-label" style="margin-bottom:4px">References <span style="color:#555;font-weight:400;font-size:9px">· Shift+Click filmstrip to change</span></div>
+                html += `<div class="sb-label" style="margin-bottom:4px">References <span style="color:#555;font-weight:400;font-size:9px">· Shift+Click filmstrip or click a thumb to remove</span></div>
                 <div class="sb-gen-refs">${refThumbs}</div>`;
                 // With 2–3 refs, let the user choose how the prompt is applied.
                 if (_sbRefIndices.length >= 2) {
@@ -4326,6 +4646,15 @@
             renderSidebarGenerate();
         }
 
+        // Remove a reference by clicking its thumbnail above the custom prompt — no page
+        // refresh needed (previously refs could only be cleared via Shift+Click filmstrip).
+        function sbRemoveRef(idx) {
+            const pos = _sbRefIndices.indexOf(idx);
+            if (pos >= 0) _sbRefIndices.splice(pos, 1);
+            updateStudio();
+            if (_activeSidebarTab === 'generate') renderSidebarGenerate();
+        }
+
         function sbSelectWireframe(val) {
             _sbWireframeRef = val;
             renderSidebarGenerate();
@@ -4543,6 +4872,53 @@
             }
         }
 
+        // ---- template video cards (faceswap / scenery) ----
+        // The wireframe library is large (dozens of files, >1 GB total) and lives on a
+        // network mount. Rendering every card as an autoplaying <video> spun up that many
+        // simultaneous decoders + range-streams, which OOM-crashed the renderer. Instead
+        // each card shows a static poster frame (a small server-rendered PNG) and only
+        // mounts a single real <video> while hovered — at most one decoder alive at a time.
+
+        // HTML for one template card. `extraHTML` lets faceswap add its trim button.
+        function _tplCardHTML(name, selected, onclickExpr, extraHTML = '') {
+            const stem  = name.replace(/\.[^.]+$/, '');
+            const sel   = selected ? ' selected' : '';
+            const nSafe = name.replace(/'/g, "\\'");
+            return `<div class="sb-template-card${sel}" onclick="${onclickExpr}"
+                onmouseenter="_tplPlay(this,'${nSafe}')" onmouseleave="_tplStop(this)">
+                <img src="${API}/wireframe/frame/${encodeURIComponent(name)}?t=0" loading="lazy" alt="">
+                <div class="sb-template-label">${escHtml(stem)}</div>${extraHTML}
+            </div>`;
+        }
+
+        // Mount + play a single <video> over the poster while hovered.
+        function _tplPlay(card, name) {
+            if (card._vid) return;
+            const v = document.createElement('video');
+            v.src = `${API}/wireframe/${encodeURIComponent(name)}`;
+            v.muted = true; v.loop = true; v.playsInline = true; v.preload = 'auto';
+            card.appendChild(v);
+            card._vid = v;
+            v.play().catch(() => {});
+        }
+
+        // Tear the video down completely so the decoder + network stream are released.
+        function _tplStop(card) {
+            const v = card && card._vid;
+            if (!v) return;
+            v.pause();
+            v.removeAttribute('src');
+            v.load();      // aborts the in-flight stream and frees the decoder
+            v.remove();
+            card._vid = null;
+        }
+
+        // Defensive: stop every active template video in the sidebar (called on tab switch
+        // so a hovered clip can't keep streaming under a freshly-rendered panel).
+        function _tplStopAll() {
+            document.querySelectorAll('#studioSidebar .sb-template-card').forEach(_tplStop);
+        }
+
         // ---- faceswap sidebar tab ----
 
         async function loadTemplateVideos() {
@@ -4612,17 +4988,10 @@
             } else {
                 html += '<div class="sb-template-grid">';
                 availableVideos.forEach(v => {
-                    const stem = v.replace(/\.[^.]+$/, '');
                     const vSafe = v.replace(/'/g, "\\'");
-                    const sel = _fsSelectedVideo === v ? ' selected' : '';
-                    html += `<div class="sb-template-card${sel}" onclick="sbSelectTemplate('${vSafe}')"
-                        onmouseenter="const v=this.querySelector('video');v&&v.play().catch(()=>{})" onmouseleave="const v=this.querySelector('video');v&&v.pause()">
-                        <video src="${API}/wireframe/${encodeURIComponent(v)}" muted loop playsinline
-                               preload="metadata" style="pointer-events:none;width:100%;height:100%;object-fit:cover"></video>
-                        <div class="sb-template-label">${escHtml(stem)}</div>
-                        <button class="sb-template-trim-btn" title="Trim video"
-                                onclick="event.stopPropagation();openTrimPanel('${vSafe}')">✂</button>
-                    </div>`;
+                    const trimBtn = `<button class="sb-template-trim-btn" title="Trim video"
+                                onclick="event.stopPropagation();openTrimPanel('${vSafe}')">✂</button>`;
+                    html += _tplCardHTML(v, _fsSelectedVideo === v, `sbSelectTemplate('${vSafe}')`, trimBtn);
                 });
                 html += '</div>';
             }
@@ -4658,9 +5027,7 @@
             </div>`;
 
             panel.innerHTML = html;
-            requestAnimationFrame(() => {
-                panel.querySelectorAll('.sb-template-card video').forEach(v => v.play().catch(()=>{}));
-            });
+            // Cards show static poster frames; videos mount only on hover (_tplPlay).
             // Check FaceFusion availability
             fetch(`${API}/faceswap/check`).then(r => r.json()).then(s => {
                 const hairCb  = document.getElementById('sbFsHair');
@@ -4863,15 +5230,8 @@
             if (availableVideos.length) {
                 html += '<div class="sb-template-grid">';
                 availableVideos.forEach(v => {
-                    const stem = v.replace(/\.[^.]+$/, '');
                     const vSafe = v.replace(/'/g, "\\'");
-                    const sel = _sceneVideo === v ? ' selected' : '';
-                    html += `<div class="sb-template-card${sel}" onclick="sceneSelectVideo('${vSafe}')"
-                        onmouseenter="const v=this.querySelector('video');v&&v.play().catch(()=>{})" onmouseleave="const v=this.querySelector('video');v&&v.pause()">
-                        <video src="${API}/wireframe/${encodeURIComponent(v)}" muted loop playsinline
-                               preload="metadata" style="pointer-events:none;width:100%;height:100%;object-fit:cover"></video>
-                        <div class="sb-template-label">${escHtml(stem)}</div>
-                    </div>`;
+                    html += _tplCardHTML(v, _sceneVideo === v, `sceneSelectVideo('${vSafe}')`);
                 });
                 html += '</div>';
             }
@@ -4913,9 +5273,7 @@
                     if (scrubber) { scrubber.max = _sceneDuration; scrubber.step = Math.max(0.05, _sceneDuration / 500); }
                 }
             }
-            requestAnimationFrame(() => {
-                panel.querySelectorAll('.sb-template-card video').forEach(v => v.play().catch(()=>{}));
-            });
+            // Cards show static poster frames; videos mount only on hover (_tplPlay).
         }
 
         async function sceneSelectVideo(v) {

tour-comfy/config.json

@@ -1,13 +1,13 @@
 {
   "api_url": "http://127.0.0.1:8500/edit",
-  "prompt": "high quality. realistic. detailed, female nude. realistic, high quality. realistic. detailed. female nude. realistic",
+  "prompt": "high quality. hyper realistic. detailed, detailed skin, detailed female nude. realistic, high quality. realistic. detailed. female nude. realistic",
   "base_prompts": [
-    "Head-on full-nude-body three-quarter female portrait, realistic, transparent background",
-    "Head-on straight-on full-nude-body female portrait, realistic, transparent background",
-    "Head-on straight-on full-body female portrait, realistic, no background",
+    "Head-on full-nude-body three-quarter female portrait, realistic, black void background",
+    "Head-on straight-on full-nude-body female portrait, realistic, black void background",
+    "Head-on straight-on full-body female portrait, realistic, black void background",
     "high quality, full-nude-body, female, masterpiece, realistic, photo, looking at viewer",
-    "high quality, full-nude-body, female, masterpiece, realistic, photo, detailed skin, professional lighting, transparent background",
-    "high quality, full-nude-body, female, masterpiece, realistic, photo, cinematic lighting, dramatic shadows, sharp focus, transparent background"
+    "high quality, full-nude-body, female, masterpiece, realistic, photo, detailed skin, professional lighting, black void background",
+    "high quality, full-nude-body, female, masterpiece, realistic, photo, cinematic lighting, dramatic shadows, sharp focus, black void background"
   ],
   "seed": -1,
   "max_area": 655360,

tour-comfy/edit_api.py

@@ -993,6 +993,8 @@ def _move_to_trash(filepath: str):
 # --- static data files -------------------------------------------------------
 
 _static_write_lock = threading.Lock()
+_invalidate_timer: "threading.Timer | None" = None
+_invalidate_timer_lock = threading.Lock()
 
 
 def _write_json(path: str, data) -> None:
@@ -1084,8 +1086,17 @@ def _write_all_static() -> None:
 
 
 def _invalidate_static() -> None:
-    """Spawn a daemon thread to regenerate all static data files (non-blocking)."""
-    threading.Thread(target=_write_all_static, daemon=True).start()
+    """Coalesce rapid invalidation calls — restarts a 0.3 s debounce timer each time.
+    At most one _write_all_static() runs per quiet window, preventing thread floods
+    during batch jobs that call this after every single image."""
+    global _invalidate_timer
+    with _invalidate_timer_lock:
+        if _invalidate_timer is not None:
+            _invalidate_timer.cancel()
+        t = threading.Timer(0.3, _write_all_static)
+        t.daemon = True
+        t.start()
+        _invalidate_timer = t
 
 
 # -----------------------------------------------------------------------------
@@ -1253,6 +1264,9 @@ def _multi_ref_worker(job_id: str, filenames: list[str], prompts: list[str], pos
                 print(f"DB error in multi-ref: {db_err}")
 
             jobs[job_id]["done"] += 1
+            # Regenerate static JSON so the frontend's polling picks up the new
+            # image immediately (progressive refresh, matching _batch_worker).
+            _invalidate_static()
         except Exception as e:
             print(f"Error in multi-ref for prompt '{prompt}': {e}")
             jobs[job_id]["failed"] += 1
@@ -1734,7 +1748,9 @@ def merge_groups(req: MergeRequest):
         except Exception as db_err:
             print(f"Database error in merge: {db_err}")
 
-    _invalidate_static()
+    # Write synchronously: the frontend reloads images.json immediately after this
+    # returns, so an async rebuild would race and show the pre-merge grouping.
+    _write_all_static()
     return {"group_id": gid, "files": req.filenames}
 
 
@@ -1750,7 +1766,7 @@ def extract_from_group(req: ExtractRequest):
     except Exception as db_err:
         print(f"Database error in extract: {db_err}")
 
-    _invalidate_static()
+    _write_all_static()
     return {"filename": req.filename}
 
 
@@ -1928,6 +1944,10 @@ def _process_upload(file_path: str, filename: str, prompts: list[str], name: str
             clip_description=clip_desc, tags=tags, embedding=embedding,
             group_id=group_id, sort_order=0, has_clothing=has_clothing,
         )
+        # Surface the new group with its base image right away — the pose/base-prompt
+        # generation below can take a while, and the user shouldn't wait for it to
+        # see the group land on the gallery.
+        _invalidate_static()
 
         # 4. Crop if needed
         cropped_pil = _crop_to_bbox(pil)
@@ -2162,16 +2182,38 @@ def remove_background(filename: str):
     person = database.get_person(filename)
     if not person or not person[5] or not os.path.exists(person[5]):
         raise HTTPException(404, "Image file not found")
-    
+
     path = person[5]
     with open(path, "rb") as f:
         png_bytes = f.read()
-    
+
     transparent_png = _apply_transparency(png_bytes)
-    
+
     with open(path, "wb") as f:
         f.write(transparent_png)
-        
+
+    # Persist the state + refresh static data so the flag (and No-BG/Crop buttons)
+    # survive a page reload instead of reverting to has_background=True.
+    database.upsert_person(filename, has_background=False)
+    _invalidate_static()
+    return {"status": "success", "filename": filename, "has_background": False}
+
+
+@app.post("/images/{filename}/invert-alpha")
+def invert_alpha(filename: str):
+    """Invert the alpha channel in place — recovers cases where background removal
+    kept the background and dropped the subject (the wrong segment)."""
+    import numpy as np
+    person = database.get_person(filename)
+    if not person or not person[5] or not os.path.exists(person[5]):
+        raise HTTPException(404, "Image file not found")
+    path = person[5]
+    img = Image.open(path).convert("RGBA")
+    arr = np.array(img)
+    arr[:, :, 3] = 255 - arr[:, :, 3]
+    Image.fromarray(arr, "RGBA").save(path, format="PNG")
+    database.upsert_person(filename, has_background=False)
+    _invalidate_static()
     return {"status": "success", "filename": filename}
 
 
@@ -2747,6 +2789,358 @@ def sam2_check():
     return {"sam2": predictor is not False and predictor is not None}
 
 
+# --- 2D body-pose preview -----------------------------------------------------
+# Estimates COCO-17 keypoints from the model image so the UI can overlay a
+# posenet-style skeleton. Estimator is feature-detected: rtmlib (ONNX, reuses the
+# already-installed onnxruntime) is preferred, mediapipe is a fallback. If neither
+# is installed the endpoints report unavailable instead of erroring the request.
+
+_pose_estimator = None          # cached (callable, backend_name) or False if unavailable
+_pose_lock = threading.Lock()
+
+# COCO-17 keypoint names (the order rtmlib's Body model returns).
+POSE_KEYPOINT_NAMES = [
+    "nose", "left_eye", "right_eye", "left_ear", "right_ear",
+    "left_shoulder", "right_shoulder", "left_elbow", "right_elbow",
+    "left_wrist", "right_wrist", "left_hip", "right_hip",
+    "left_knee", "right_knee", "left_ankle", "right_ankle",
+]
+# Bone connections (index pairs into COCO-17) for drawing the skeleton.
+POSE_SKELETON = [
+    (5, 7), (7, 9), (6, 8), (8, 10),          # arms
+    (11, 13), (13, 15), (12, 14), (14, 16),   # legs
+    (5, 6), (11, 12), (5, 11), (6, 12),       # torso
+    (0, 1), (0, 2), (1, 3), (2, 4), (0, 5), (0, 6),  # head/neck
+]
+# mediapipe Pose (33 landmarks) → COCO-17 index map.
+_MP_TO_COCO = [0, 2, 5, 7, 8, 11, 12, 13, 14, 15, 16, 23, 24, 25, 26, 27, 28]
+
+
+def _load_pose_estimator():
+    global _pose_estimator
+    if _pose_estimator is not None:
+        return _pose_estimator
+    with _pose_lock:
+        if _pose_estimator is not None:
+            return _pose_estimator
+        # Preferred: rtmlib (RTMPose, ONNX) — returns COCO-17 directly.
+        try:
+            from rtmlib import Body
+            import numpy as np
+            model = Body(mode="balanced", backend="onnxruntime", device="cpu")
+
+            def _infer_rtm(pil):
+                bgr = np.array(pil.convert("RGB"))[:, :, ::-1]
+                kpts, scores = model(bgr)   # (N,17,2), (N,17)
+                people = []
+                for person_kpts, person_scores in zip(kpts, scores):
+                    people.append([[float(x), float(y), float(s)]
+                                   for (x, y), s in zip(person_kpts, person_scores)])
+                return people
+
+            _pose_estimator = (_infer_rtm, "rtmlib")
+            print("[pose] using rtmlib (RTMPose)")
+            return _pose_estimator
+        except Exception as e:
+            print(f"[pose] rtmlib unavailable: {e}")
+
+        # Fallback: mediapipe Pose (single person, normalized landmarks).
+        try:
+            import mediapipe as mp
+            import numpy as np
+            mp_pose = mp.solutions.pose.Pose(static_image_mode=True, model_complexity=2)
+
+            def _infer_mp(pil):
+                rgb = np.array(pil.convert("RGB"))
+                h, w = rgb.shape[:2]
+                res = mp_pose.process(rgb)
+                if not res.pose_landmarks:
+                    return []
+                lm = res.pose_landmarks.landmark
+                kpts = []
+                for mp_idx in _MP_TO_COCO:
+                    p = lm[mp_idx]
+                    kpts.append([float(p.x * w), float(p.y * h), float(p.visibility)])
+                return [kpts]
+
+            _pose_estimator = (_infer_mp, "mediapipe")
+            print("[pose] using mediapipe Pose")
+            return _pose_estimator
+        except Exception as e:
+            print(f"[pose] mediapipe unavailable: {e}")
+
+        _pose_estimator = False
+        return _pose_estimator
+
+
+# --- pose similarity (descriptor + index) -------------------------------------
+# Pose descriptors are normalized (translation + scale invariant) COCO-17 vectors,
+# cached in <output>/_data/poses_index.json so we can rank library images by pose.
+
+_POSE_MIN_SCORE = 0.3
+# Left/right keypoint pairs for the mirror-invariant distance.
+_POSE_MIRROR = [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]
+_pose_index_status = {"running": False, "done": 0, "total": 0}
+
+
+def _pose_descriptor(keypoints):
+    """Normalize one person's COCO-17 keypoints into a translation/scale-invariant
+    descriptor: {"vec": [34 floats], "vis": [17 ints]}. Returns None if too sparse."""
+    vis = [1 if (kp[2] >= _POSE_MIN_SCORE) else 0 for kp in keypoints]
+    if sum(vis) < 6:
+        return None
+
+    def _mid(a, b):
+        if vis[a] and vis[b]:
+            return ((keypoints[a][0] + keypoints[b][0]) / 2.0,
+                    (keypoints[a][1] + keypoints[b][1]) / 2.0)
+        return None
+
+    hip = _mid(11, 12)
+    sho = _mid(5, 6)
+    center = hip or sho
+    if center is None:
+        return None
+    # Scale by torso length; fall back to keypoint spread if torso isn't visible.
+    if hip and sho:
+        scale = ((hip[0] - sho[0]) ** 2 + (hip[1] - sho[1]) ** 2) ** 0.5
+    else:
+        xs = [keypoints[i][0] for i in range(17) if vis[i]]
+        ys = [keypoints[i][1] for i in range(17) if vis[i]]
+        scale = max(max(xs) - min(xs), max(ys) - min(ys))
+    if not scale or scale < 1e-3:
+        return None
+
+    vec = []
+    for i in range(17):
+        if vis[i]:
+            vec.append((keypoints[i][0] - center[0]) / scale)
+            vec.append((keypoints[i][1] - center[1]) / scale)
+        else:
+            vec.extend([0.0, 0.0])
+    return {"vec": vec, "vis": vis}
+
+
+def _pose_distance(a, b):
+    """Weighted L2 between two descriptors over jointly-visible joints, taking the
+    min of the direct and left-right-mirrored comparison. Lower = more similar."""
+    def _dist(av, avis, bv, bvis, mirror):
+        total, n = 0.0, 0
+        for i in range(17):
+            j = _POSE_MIRROR[i] if mirror else i
+            if not (avis[i] and bvis[j]):
+                continue
+            bx = bv[j * 2] * (-1 if mirror else 1)   # flip x when mirrored
+            by = bv[j * 2 + 1]
+            dx = av[i * 2] - bx
+            dy = av[i * 2 + 1] - by
+            total += dx * dx + dy * dy
+            n += 1
+        return (total / n) ** 0.5 if n >= 4 else float("inf")
+    direct = _dist(a["vec"], a["vis"], b["vec"], b["vis"], False)
+    mirror = _dist(a["vec"], a["vis"], b["vec"], b["vis"], True)
+    return min(direct, mirror)
+
+
+def _best_person(people):
+    """Pick the largest-bbox person from an estimator result (most prominent subject)."""
+    best, best_area = None, -1.0
+    for kpts in people:
+        xs = [k[0] for k in kpts if k[2] >= _POSE_MIN_SCORE]
+        ys = [k[1] for k in kpts if k[2] >= _POSE_MIN_SCORE]
+        if len(xs) < 2:
+            continue
+        area = (max(xs) - min(xs)) * (max(ys) - min(ys))
+        if area > best_area:
+            best, best_area = kpts, area
+    return best
+
+
+def _pose_index_path():
+    return os.path.join(_load_output_dir(), "_data", "poses_index.json")
+
+
+def _load_pose_index():
+    try:
+        with open(_pose_index_path(), "r") as f:
+            return json.load(f)
+    except Exception:
+        return {}
+
+
+_pose_index_lock = threading.Lock()
+
+
+def _save_pose_index_entry(filename, desc):
+    with _pose_index_lock:
+        idx = _load_pose_index()
+        idx[filename] = desc
+        os.makedirs(os.path.dirname(_pose_index_path()), exist_ok=True)
+        _write_json(_pose_index_path(), idx)
+
+
+@app.get("/pose/check")
+def pose_check():
+    """Report whether a body-pose estimator is available (and which backend)."""
+    est = _load_pose_estimator()
+    if not est:
+        return {"available": False,
+                "hint": "pip install rtmlib onnxruntime  (or: pip install mediapipe)"}
+    return {"available": True, "backend": est[1]}
+
+
+@app.post("/images/{filename}/pose")
+def estimate_pose(filename: str):
+    """Estimate COCO-17 body keypoints for an image. Returns pixel-space keypoints
+    plus the skeleton edge list so the frontend can overlay a posenet-style preview."""
+    person = database.get_person(filename)
+    if not person or not person[5] or not os.path.exists(person[5]):
+        raise HTTPException(404, "Image file not found")
+    est = _load_pose_estimator()
+    if not est:
+        raise HTTPException(501, "No pose estimator installed. Try: pip install rtmlib onnxruntime")
+    infer, backend = est
+    pil = Image.open(person[5]).convert("RGB")
+    try:
+        people = infer(pil)
+    except Exception as e:
+        raise HTTPException(500, f"Pose estimation failed: {e}")
+    # Cache the descriptor so "find similar pose" can rank this image later.
+    best = _best_person(people)
+    if best is not None:
+        desc = _pose_descriptor(best)
+        if desc is not None:
+            try:
+                _save_pose_index_entry(filename, desc)
+            except Exception as e:
+                print(f"[pose] index save failed for {filename}: {e}")
+    return {
+        "status": "success",
+        "backend": backend,
+        "width": pil.width,
+        "height": pil.height,
+        "names": POSE_KEYPOINT_NAMES,
+        "skeleton": POSE_SKELETON,
+        "people": people,
+    }
+
+
+def _build_pose_index_task():
+    try:
+        est = _load_pose_estimator()
+        if not est:
+            return
+        infer, _ = est
+        output_dir = _load_output_dir()
+        with _pose_index_lock:
+            idx = _load_pose_index()
+        persons = database.list_persons()
+        todo = [p[0] for p in persons
+                if p[0] not in idx
+                and (p[12] or "image") != "video"
+                and p[0].lower().endswith(('.png', '.jpg', '.jpeg', '.webp'))]
+        _pose_index_status.update(running=True, done=0, total=len(todo))
+        print(f"[pose] index build: {len(todo)} images to process")
+        dirty = 0
+        for fn in todo:
+            try:
+                fpath = os.path.join(output_dir, fn)
+                if os.path.exists(fpath):
+                    best = _best_person(infer(Image.open(fpath).convert("RGB")))
+                    desc = _pose_descriptor(best) if best is not None else None
+                    if desc is not None:
+                        idx[fn] = desc
+                        dirty += 1
+            except Exception as e:
+                print(f"[pose] index error for {fn}: {e}")
+            _pose_index_status["done"] += 1
+            # Batch-flush every 50 to avoid O(n^2) full-file rewrites.
+            if dirty >= 50:
+                with _pose_index_lock:
+                    _write_json(_pose_index_path(), idx)
+                dirty = 0
+                print(f"[pose] index progress: {_pose_index_status['done']}/{len(todo)}")
+        with _pose_index_lock:
+            _write_json(_pose_index_path(), idx)
+        print(f"[pose] index build complete: {len(idx)} entries")
+    except Exception as e:
+        print(f"[pose] index build failed: {e}")
+    finally:
+        _pose_index_status["running"] = False
+
+
+@app.post("/pose/index")
+def build_pose_index():
+    """Compute pose descriptors for all library images lacking one (daemon thread)."""
+    if not _load_pose_estimator():
+        raise HTTPException(501, "No pose estimator installed. Try: pip install rtmlib onnxruntime")
+    if _pose_index_status.get("running"):
+        return {"status": "already_running", **_pose_index_status}
+    threading.Thread(target=_build_pose_index_task, daemon=True).start()
+    return {"status": "started"}
+
+
+@app.get("/pose/index/status")
+def pose_index_status():
+    idx = _load_pose_index()
+    return {**_pose_index_status, "indexed": len(idx)}
+
+
+def _rank_similar_poses(query_desc, limit, exclude=None):
+    idx = _load_pose_index()
+    scored = []
+    for fn, desc in idx.items():
+        if fn == exclude or not desc or "vec" not in desc:
+            continue
+        d = _pose_distance(query_desc, desc)
+        if d != float("inf"):
+            scored.append((d, fn))
+    scored.sort(key=lambda x: x[0])
+    groups = get_groups() if scored else {}
+    return [{"filename": fn, "group_id": groups.get(fn), "distance": round(d, 4)}
+            for d, fn in scored[:limit]]
+
+
+@app.get("/pose/similar/{filename}")
+def similar_pose(filename: str, limit: int = 12):
+    """Rank library images by pose similarity to the given image."""
+    idx = _load_pose_index()
+    query = idx.get(filename)
+    if query is None:
+        # Compute on demand (also caches it).
+        person = database.get_person(filename)
+        if not person or not person[5] or not os.path.exists(person[5]):
+            raise HTTPException(404, "Image file not found")
+        est = _load_pose_estimator()
+        if not est:
+            raise HTTPException(501, "No pose estimator installed.")
+        best = _best_person(est[0](Image.open(person[5]).convert("RGB")))
+        query = _pose_descriptor(best) if best is not None else None
+        if query is None:
+            raise HTTPException(404, "No detectable pose in this image")
+        try:
+            _save_pose_index_entry(filename, query)
+        except Exception:
+            pass
+    return {"filename": filename, "similar": _rank_similar_poses(query, limit, exclude=filename)}
+
+
+class PoseSimilarRequest(BaseModel):
+    keypoints: list[list[float]]   # [[x,y,score], ...17] in image pixels
+    width: int = 0
+    height: int = 0
+    limit: int = 12
+
+
+@app.post("/pose/similar")
+def similar_pose_from_keypoints(req: PoseSimilarRequest):
+    """Rank library images by similarity to a supplied (e.g. hand-edited) skeleton."""
+    query = _pose_descriptor(req.keypoints)
+    if query is None:
+        raise HTTPException(400, "Supplied pose is too sparse to match")
+    return {"similar": _rank_similar_poses(query, req.limit)}
+
+
 if __name__ == "__main__":
     import uvicorn
     uvicorn.run(app, host=os.environ.get("HOST", "0.0.0.0"),

tour-comfy/systemd/comfyui-api.service

@@ -1,18 +0,0 @@
-[Unit]
-Description=Qwen-Image-Edit FastAPI Service
-After=comfyui-backend.service
-Requires=comfyui-backend.service
-
-[Service]
-Type=simple
-User=__USER__
-Group=__GROUP__
-WorkingDirectory=__BASE__/api
-ExecStart=/bin/bash __BASE__/api/start_api.sh
-Restart=always
-RestartSec=5
-StandardOutput=journal
-StandardError=journal
-
-[Install]
-WantedBy=multi-user.target

tour-comfy/systemd/comfyui-backend.service

@@ -1,17 +0,0 @@
-[Unit]
-Description=ComfyUI Backend for Qwen-Image-Edit
-After=network.target
-
-[Service]
-Type=simple
-User=__USER__
-Group=__GROUP__
-WorkingDirectory=__BASE__
-ExecStart=/bin/bash __BASE__/api/run_comfyui.sh
-Restart=always
-RestartSec=5
-StandardOutput=journal
-StandardError=journal
-
-[Install]
-WantedBy=multi-user.target