FSF-Studio / app.py
Aimee
Survive a server restart instead of dying on a stale token
943fe07
Raw History Blame Contribute Delete
43.2 kB
""" Marshall's Plasma Design Studio. Pipeline: client reference image -> auto-trace -> dimensioned vector -> DXF/SVG for the plasma table. """ from __future__ import annotations import base64 import io import math import os import uuid from collections import OrderedDict from pathlib import Path import ezdxf import numpy as np from flask import Flask, jsonify, request, send_file, send_from_directory from PIL import Image, ImageOps from potrace import Bitmap, POTRACE_TURNPOLICY_MINORITY from scipy.ndimage import distance_transform_edt, label STATIC_DIR = Path(__file__).parent / "static" app = Flask(__name__, static_folder=str(STATIC_DIR), static_url_path="/static") MAX_IMAGE_DIM = 4200 # pixels; downscale large uploads before tracing/segmentation # 1200 was sized for a client photo traced onto a two-foot sign. On a ten # foot fence panel it lands at ~9 px/inch, which is one pixel per 0.105" and # destroys fine linework (snow cracks, treeline needles, a fly line) on # import, before the trace ever runs. MAX_PANEL_PX = 6000 # cap composed-panel pixel dimensions to keep tracing snappy # Raising MAX_IMAGE_DIM alone does nothing: this cap independently throttles # a 120" panel to 18 px/inch. At 6000 a 120" panel holds the full 40 px/inch # (0.025"/px), which is finer than any plasma kerf. MASK_CACHE_LIMIT = 8 # most recent composed/traced masks kept for /analyze _mask_cache: "OrderedDict[str, np.ndarray]" = OrderedDict() IMAGE_CACHE_LIMIT = 32 # most recent uploaded images, keyed by token _image_cache: "OrderedDict[str, Image.Image]" = OrderedDict() LAYER_MASK_CACHE_LIMIT = 64 # cache per-layer masks, keyed by (token + mode + params hash) _layer_mask_cache: "OrderedDict[str, np.ndarray]" = OrderedDict() def _cache_mask(mask: np.ndarray) -> str: token = uuid.uuid4().hex _mask_cache[token] = mask while len(_mask_cache) > MASK_CACHE_LIMIT: _mask_cache.popitem(last=False) return token def _cache_image(img: Image.Image) -> str: token = uuid.uuid4().hex _image_cache[token] = img while len(_image_cache) > IMAGE_CACHE_LIMIT: _image_cache.popitem(last=False) return token def _cache_layer_mask(key: str, mask: np.ndarray) -> None: _layer_mask_cache[key] = mask _layer_mask_cache.move_to_end(key) while len(_layer_mask_cache) > LAYER_MASK_CACHE_LIMIT: _layer_mask_cache.popitem(last=False) def _otsu_threshold(arr: np.ndarray) -> int: """Otsu's method: pick the threshold that maximizes inter-class variance.""" hist, _ = np.histogram(arr, bins=256, range=(0, 256)) total = arr.size if total == 0: return 128 sum_total = float((np.arange(256) * hist).sum()) sum_b = 0.0 w_b = 0 var_max = 0.0 threshold = 128 for t in range(256): w_b += int(hist[t]) if w_b == 0: continue w_f = total - w_b if w_f == 0: break sum_b += t * float(hist[t]) m_b = sum_b / w_b m_f = (sum_total - sum_b) / w_f var_between = w_b * w_f * (m_b - m_f) ** 2 if var_between > var_max: var_max = var_between threshold = t return threshold def _open_resized(file_storage, mode: str = "RGB") -> Image.Image: img = Image.open(file_storage.stream).convert(mode) img = ImageOps.exif_transpose(img) if max(img.size) > MAX_IMAGE_DIM: scale = MAX_IMAGE_DIM / max(img.size) new_size = (int(img.size[0] * scale), int(img.size[1] * scale)) img = img.resize(new_size, Image.LANCZOS) return img def load_and_prepare(file_storage, threshold: int, invert: bool, blur: float, auto_threshold: bool = False) -> np.ndarray: img = _open_resized(file_storage, mode="L") if blur and blur > 0: from PIL import ImageFilter img = img.filter(ImageFilter.GaussianBlur(radius=blur)) arr = np.array(img) if auto_threshold: threshold = _otsu_threshold(arr) # potracer's Bitmap traces False regions when True touches the array edge. # For a dark subject on light background, mark light pixels True so the # dark subject is what gets outlined (no phantom outer hull). mask = arr >= threshold if not invert else arr < threshold return mask, int(threshold) def subject_mask(file_storage, alpha_cutoff: int = 50, invert: bool = False) -> np.ndarray: """Use rembg (U-2-Net) to extract foreground; return mask in our convention (True = background, False = subject) so trace outlines the subject.""" img = _open_resized(file_storage, mode="RGBA") return subject_mask_from_image(img, alpha_cutoff=alpha_cutoff, invert=invert) def subject_mask_from_image(img: Image.Image, alpha_cutoff: int = 50, invert: bool = False) -> np.ndarray: from rembg import remove rgba = img.convert("RGBA") buf = io.BytesIO() rgba.save(buf, format="PNG") out_bytes = remove(buf.getvalue()) out = Image.open(io.BytesIO(out_bytes)) arr = np.array(out) alpha = arr[:, :, 3] is_subject = alpha > alpha_cutoff if invert: is_subject = ~is_subject return ~is_subject # True for background def threshold_mask_from_image(img: Image.Image, threshold: int, invert: bool, blur: float, auto_threshold: bool) -> tuple[np.ndarray, int]: gray = img.convert("L") if blur and blur > 0: from PIL import ImageFilter gray = gray.filter(ImageFilter.GaussianBlur(radius=blur)) arr = np.array(gray) if auto_threshold: threshold = _otsu_threshold(arr) mask = arr >= threshold if not invert else arr < threshold return mask, int(threshold) def _crop_tuple(params: dict) -> tuple[int, int, int, int] | None: crop = params.get("crop") if not crop: return None try: x0 = int(crop["x0"]); y0 = int(crop["y0"]) x1 = int(crop["x1"]); y1 = int(crop["y1"]) except (KeyError, TypeError, ValueError): return None if x1 <= x0 or y1 <= y0: return None return (x0, y0, x1, y1) def _layer_mask_key(image_token: str, mode: str, params: dict) -> str: crop = _crop_tuple(params) crop_str = f"c={crop}" if crop else "c=none" canon = ( f"{image_token}|{mode}|" f"t={int(params.get('threshold', 128))}|" f"b={float(params.get('blur', 0)):.3f}|" f"i={bool(params.get('invert', False))}|" f"a={int(params.get('alpha_cutoff', 50))}|" f"auto={bool(params.get('auto_threshold', False))}|" f"{crop_str}" ) return canon def compute_layer_mask(image_token: str, mode: str, params: dict) -> np.ndarray: """Return mask in our convention: True=background, False=subject.""" if image_token not in _image_cache: raise KeyError(f"unknown image_token: {image_token}") key = _layer_mask_key(image_token, mode, params) if key in _layer_mask_cache: _layer_mask_cache.move_to_end(key) return _layer_mask_cache[key] img = _image_cache[image_token] crop = _crop_tuple(params) if crop: # Clamp the crop to the image bounds and crop the source. x0, y0, x1, y1 = crop x0 = max(0, min(img.width, x0)) y0 = max(0, min(img.height, y0)) x1 = max(x0 + 1, min(img.width, x1)) y1 = max(y0 + 1, min(img.height, y1)) img = img.crop((x0, y0, x1, y1)) if mode == "subject": mask = subject_mask_from_image( img, alpha_cutoff=int(params.get("alpha_cutoff", 50)), invert=bool(params.get("invert", False)), ) elif mode in ("auto", "manual"): mask, _ = threshold_mask_from_image( img, threshold=int(params.get("threshold", 128)), invert=bool(params.get("invert", False)), blur=float(params.get("blur", 0)), auto_threshold=(mode == "auto"), ) else: raise ValueError(f"unknown mask mode: {mode}") _cache_layer_mask(key, mask) return mask def compose_panel(panel_w_in: float, panel_h_in: float, layers: list, intent: str = "cutouts", px_per_inch: float = 40.0) -> tuple[np.ndarray, float]: """Rasterize each visible layer onto a panel-sized binary canvas. Panel convention: True = kept steel (panel/background), False = cut. Starting state depends on global intent: - "cutouts": panel starts all-steel; layers default to SUBTRACT. - "silhouette": panel starts all-cut; layers default to ADD. Each layer can override its default polarity via mask_params.polarity in {"add", "subtract"}, which is how Marshall composes things like "wolf-as-steel with text holes punched inside" (wolf ADD, text SUBTRACT) or "panel with moon hole and eagle steel inside the moon" (moon SUBTRACT, eagle ADD). Later layers override earlier ones on the pixels they touch. Returns (mask, in_per_px). """ panel_w_px = int(round(panel_w_in * px_per_inch)) panel_h_px = int(round(panel_h_in * px_per_inch)) longest = max(panel_w_px, panel_h_px) if longest > MAX_PANEL_PX: scale = MAX_PANEL_PX / longest panel_w_px = int(round(panel_w_px * scale)) panel_h_px = int(round(panel_h_px * scale)) px_per_inch = px_per_inch * scale in_per_px = 1.0 / px_per_inch # Panel always starts as all-steel (True). Per-layer polarity does the work: # SUBTRACT cuts the layer's subject out, ADD asserts steel (useful for # restoring steel inside an already-cut region — e.g. an eagle inside a # moon hole). Global intent is kept for the analyzer only. panel = np.ones((panel_h_px, panel_w_px), dtype=bool) default_polarity = "subtract" skipped: list = [] for layer in layers: if not layer.get("visible", True): continue token = layer.get("image_token") mode = layer.get("mask_mode", "subject") params = layer.get("mask_params", {}) or {} polarity = layer.get("polarity") or params.get("polarity") or default_polarity if polarity not in ("add", "subtract"): polarity = default_polarity try: mask = compute_layer_mask(token, mode, params) except (KeyError, ValueError) as e: # Silently skipping renders an empty panel and explains nothing. # The in-memory image cache dies on every restart, so this is the # normal failure, not an exotic one. skipped.append({"image_token": token, "reason": str(e)}) continue # mask True=background; subject pixels are False. subject = (~mask).astype(np.uint8) * 255 # 0 or 255 layer_img = Image.fromarray(subject, mode="L") layer_w_in = float(layer.get("scale_in", min(panel_w_in, panel_h_in) * 0.5)) if layer_w_in <= 0: continue aspect = layer_img.height / max(layer_img.width, 1) layer_w_px = max(int(round(layer_w_in * px_per_inch)), 1) layer_h_px = max(int(round(layer_w_px * aspect)), 1) if layer_w_px < 1 or layer_h_px < 1: continue layer_img_r = layer_img.resize((layer_w_px, layer_h_px), Image.NEAREST) cx_in = float(layer.get("x_in", panel_w_in / 2)) cy_in = float(layer.get("y_in", panel_h_in / 2)) cx_px = int(round(cx_in * px_per_inch)) cy_px = int(round(cy_in * px_per_inch)) x0 = cx_px - layer_w_px // 2 y0 = cy_px - layer_h_px // 2 x1 = x0 + layer_w_px y1 = y0 + layer_h_px sx0, sy0 = max(0, -x0), max(0, -y0) dx0, dy0 = max(0, x0), max(0, y0) dx1, dy1 = min(panel_w_px, x1), min(panel_h_px, y1) if dx0 >= dx1 or dy0 >= dy1: continue sub = np.array(layer_img_r)[sy0:sy0 + (dy1 - dy0), sx0:sx0 + (dx1 - dx0)] is_subject = sub > 127 # ADD sets subject -> True (steel); SUBTRACT sets subject -> False (cut). if polarity == "add": panel[dy0:dy1, dx0:dx1] |= is_subject else: # subtract panel[dy0:dy1, dx0:dx1] &= ~is_subject return panel, in_per_px, skipped def trace_to_paths(mask: np.ndarray, turdsize: int, alphamax: float, opttolerance: float): """Return list of paths. Each path is a list of subpaths (curves). Each subpath has a 'start' [x, y] and 'segments' list with type 'C' (corner) or 'B' (bezier) plus their control/end points. """ bmp = Bitmap(mask) path = bmp.trace( turdsize=turdsize, turnpolicy=POTRACE_TURNPOLICY_MINORITY, alphamax=alphamax, opticurve=True, opttolerance=opttolerance, ) out = [] for curve in path: sp = curve.start_point sub = {"start": [float(sp.x), float(sp.y)], "segments": []} for seg in curve.segments: ep = seg.end_point if seg.is_corner: c = seg.c sub["segments"].append({ "type": "C", "c": [float(c.x), float(c.y)], "end": [float(ep.x), float(ep.y)], }) else: c1, c2 = seg.c1, seg.c2 sub["segments"].append({ "type": "B", "c1": [float(c1.x), float(c1.y)], "c2": [float(c2.x), float(c2.y)], "end": [float(ep.x), float(ep.y)], }) out.append(sub) return out def paths_to_svg_d(paths) -> list[str]: """Convert internal path data into SVG 'd' attribute strings (one per closed subpath).""" out = [] for sub in paths: sx, sy = sub["start"] d = [f"M {sx:.3f} {sy:.3f}"] for seg in sub["segments"]: ex, ey = seg["end"] if seg["type"] == "C": cx, cy = seg["c"] d.append(f"L {cx:.3f} {cy:.3f} L {ex:.3f} {ey:.3f}") else: c1x, c1y = seg["c1"] c2x, c2y = seg["c2"] d.append(f"C {c1x:.3f} {c1y:.3f} {c2x:.3f} {c2y:.3f} {ex:.3f} {ey:.3f}") d.append("Z") out.append(" ".join(d)) return out def flatten_paths(paths, bezier_steps: int = 24): """Convert paths to lists of (x, y) polylines for DXF export.""" polylines = [] for sub in paths: pts = [] x, y = sub["start"] pts.append((x, y)) for seg in sub["segments"]: if seg["type"] == "C": cx, cy = seg["c"] ex, ey = seg["end"] pts.append((cx, cy)) pts.append((ex, ey)) x, y = ex, ey else: c1x, c1y = seg["c1"] c2x, c2y = seg["c2"] ex, ey = seg["end"] px, py = x, y for i in range(1, bezier_steps + 1): t = i / bezier_steps bx = (1 - t) ** 3 * px + 3 * (1 - t) ** 2 * t * c1x + 3 * (1 - t) * t ** 2 * c2x + t ** 3 * ex by = (1 - t) ** 3 * py + 3 * (1 - t) ** 2 * t * c1y + 3 * (1 - t) * t ** 2 * c2y + t ** 3 * ey pts.append((bx, by)) x, y = ex, ey polylines.append(pts) return polylines def fit_transform(paths, target_w: float, target_h: float, units_in_inches: bool): """Compute scale and offset so the path bbox fits target dimensions, preserving aspect. Returns (scale, dx, dy, fitted_w, fitted_h).""" xs, ys = [], [] for sub in paths: xs.append(sub["start"][0]); ys.append(sub["start"][1]) for seg in sub["segments"]: if seg["type"] == "C": xs.append(seg["c"][0]); ys.append(seg["c"][1]) else: xs.append(seg["c1"][0]); ys.append(seg["c1"][1]) xs.append(seg["c2"][0]); ys.append(seg["c2"][1]) xs.append(seg["end"][0]); ys.append(seg["end"][1]) if not xs: return 1.0, 0.0, 0.0, 0.0, 0.0 minx, maxx = min(xs), max(xs) miny, maxy = min(ys), max(ys) src_w = maxx - minx src_h = maxy - miny if src_w == 0 or src_h == 0: return 1.0, -minx, -miny, src_w, src_h scale = min(target_w / src_w, target_h / src_h) fitted_w = src_w * scale fitted_h = src_h * scale dx = -minx * scale + (target_w - fitted_w) / 2 dy = -miny * scale + (target_h - fitted_h) / 2 return scale, dx, dy, fitted_w, fitted_h def transform_polylines(polylines, scale, dx, dy, flip_y_height=None): out = [] for pts in polylines: new = [] for x, y in pts: nx = x * scale + dx ny = y * scale + dy if flip_y_height is not None: ny = flip_y_height - ny new.append((nx, ny)) out.append(new) return out @app.get("/") def root(): return send_from_directory(str(STATIC_DIR), "index.html") # --- Gemini image generation --- GEMINI_FLASH_MODEL = "gemini-3.1-flash-image-preview" # "Nano Banana 2" PLASMA_STYLE_SUFFIX = ( " High-contrast black silhouette on a pure white background. " "Bold simple shapes, clean outlines, no shading or gradients, no text. " "Designed to be cut from steel on a plasma table." ) _gemini_client = None def get_gemini_client(): global _gemini_client if _gemini_client is None: api_key = os.environ.get("GEMINI_API_KEY") if not api_key: raise RuntimeError( "GEMINI_API_KEY is not set. On HF Spaces, add it under " "Settings → Variables and secrets. Locally, set it in your shell." ) from google import genai _gemini_client = genai.Client(api_key=api_key) return _gemini_client def _gemini_generate_image(prompt: str, size: str = "1K") -> Image.Image: from google.genai import types client = get_gemini_client() resp = client.models.generate_content( model=GEMINI_FLASH_MODEL, contents=[prompt], config=types.GenerateContentConfig( image_config=types.ImageConfig(image_size=size), ), ) # Find the first image part in the response for cand in (resp.candidates or []): for part in (cand.content.parts or []): inline = getattr(part, "inline_data", None) if inline and inline.data: return Image.open(io.BytesIO(inline.data)).convert("RGBA") raise RuntimeError("Gemini returned no image data") @app.post("/generate") def generate(): data = request.get_json(force=True) or {} prompt = (data.get("prompt") or "").strip() if not prompt: return jsonify({"error": "prompt required"}), 400 plasma_style = bool(data.get("plasma_style", True)) size = data.get("size", "1K") if size not in ("1K", "2K", "4K"): size = "1K" full_prompt = prompt + (PLASMA_STYLE_SUFFIX if plasma_style else "") try: img = _gemini_generate_image(full_prompt, size=size) except RuntimeError as e: return jsonify({"error": str(e)}), 503 except Exception as e: return jsonify({"error": f"generation failed: {e}"}), 502 # Downscale to fit our pipeline expectations if max(img.size) > MAX_IMAGE_DIM: scale = MAX_IMAGE_DIM / max(img.size) img = img.resize((int(img.width * scale), int(img.height * scale)), Image.LANCZOS) token = _cache_image(img) thumb = img.copy() thumb.thumbnail((160, 160), Image.LANCZOS) buf = io.BytesIO() thumb.convert("RGB").save(buf, format="JPEG", quality=80) thumb_b64 = base64.b64encode(buf.getvalue()).decode("ascii") # Echo the source PNG as a base64 data URL so the frontend can use it for # crop modal display (same as if the user had uploaded the file directly). src_buf = io.BytesIO() img.save(src_buf, format="PNG") src_b64 = base64.b64encode(src_buf.getvalue()).decode("ascii") return jsonify({ "image_token": token, "width": img.width, "height": img.height, "thumbnail": f"data:image/jpeg;base64,{thumb_b64}", "source_data_url": f"data:image/png;base64,{src_b64}", "name": prompt[:40] + ("…" if len(prompt) > 40 else ""), "used_prompt": full_prompt, }) @app.post("/upload") def upload(): if "image" not in request.files: return jsonify({"error": "no image uploaded"}), 400 f = request.files["image"] img = _open_resized(f, mode="RGBA") token = _cache_image(img) # Encode a thumbnail dataURL for the layer panel. thumb = img.copy() thumb.thumbnail((160, 160), Image.LANCZOS) buf = io.BytesIO() thumb.convert("RGB").save(buf, format="JPEG", quality=80) thumb_b64 = base64.b64encode(buf.getvalue()).decode("ascii") return jsonify({ "image_token": token, "width": img.width, "height": img.height, "thumbnail": f"data:image/jpeg;base64,{thumb_b64}", "name": getattr(f, "filename", "image") or "image", "is_line_art": bool(is_line_art(img)), }) @app.post("/layer/preview") def layer_preview(): """Return the layer's binary mask as a transparent PNG, sized to the source image. Used for rendering a clean cutout preview on the canvas while the user adjusts position/scale.""" data = request.get_json(force=True) token = data.get("image_token") mode = data.get("mask_mode", "subject") params = data.get("mask_params", {}) or {} if not token or token not in _image_cache: return jsonify({"error": "unknown image_token"}), 400 try: mask = compute_layer_mask(token, mode, params) except Exception as e: return jsonify({"error": str(e)}), 500 # mask: True = background (transparent), False = subject (cream) h, w = mask.shape rgba = np.zeros((h, w, 4), dtype=np.uint8) subject = ~mask rgba[subject] = [247, 243, 234, 255] # cream rgba[~subject] = [0, 0, 0, 0] img = Image.fromarray(rgba, mode="RGBA") buf = io.BytesIO() img.save(buf, format="PNG") b64 = base64.b64encode(buf.getvalue()).decode("ascii") # Subject bbox in source-image pixels — frontend converts to inches # using the layer's panel width and aspect ratio. bbox = None ys, xs = np.where(subject) if xs.size: bbox = { "x0": int(xs.min()), "y0": int(ys.min()), "x1": int(xs.max()), "y1": int(ys.max()), } return jsonify({ "mask_png": f"data:image/png;base64,{b64}", "width": int(w), "height": int(h), "subject_bbox": bbox, "subject_pixels": int(subject.sum()), }) @app.post("/compose") def compose(): """Compose all visible layers into a single panel mask, trace it, and cache the result for /analyze and /export.""" data = request.get_json(force=True) try: panel_w_in = float(data["panel_w_in"]) panel_h_in = float(data["panel_h_in"]) except (KeyError, ValueError, TypeError): return jsonify({"error": "panel_w_in and panel_h_in required"}), 400 if panel_w_in <= 0 or panel_h_in <= 0: return jsonify({"error": "panel dimensions must be > 0"}), 400 layers = data.get("layers") or [] if not isinstance(layers, list) or not layers: return jsonify({"error": "layers array required"}), 400 turdsize = int(data.get("turdsize", 8)) alphamax = float(data.get("alphamax", 1.0)) opttolerance = float(data.get("opttolerance", 0.4)) intent = data.get("intent", "cutouts") if intent not in ("cutouts", "silhouette"): intent = "cutouts" panel_mask, in_per_px, skipped_layers = compose_panel( panel_w_in, panel_h_in, layers, intent=intent) if (~panel_mask).sum() == 0: return jsonify({ "paths": [], "svg_d": [], "width": int(panel_mask.shape[1]), "height": int(panel_mask.shape[0]), "token": None, "in_per_px": in_per_px, "panel_w_in": panel_w_in, "panel_h_in": panel_h_in, "empty": True, "skipped_layers": skipped_layers, }) paths = trace_to_paths(panel_mask, turdsize=turdsize, alphamax=alphamax, opttolerance=opttolerance) h, w = panel_mask.shape token = _cache_mask(panel_mask) return jsonify({ "width": int(w), "height": int(h), "paths": paths, "svg_d": paths_to_svg_d(paths), "token": token, "in_per_px": in_per_px, "panel_w_in": panel_w_in, "panel_h_in": panel_h_in, "empty": False, "skipped_layers": skipped_layers, }) @app.post("/trace") def trace(): if "image" not in request.files: return jsonify({"error": "no image uploaded"}), 400 f = request.files["image"] threshold = int(request.form.get("threshold", 128)) invert = request.form.get("invert", "false").lower() == "true" blur = float(request.form.get("blur", 0)) turdsize = int(request.form.get("turdsize", 8)) alphamax = float(request.form.get("alphamax", 1.0)) opttolerance = float(request.form.get("opttolerance", 0.4)) extract_subject = request.form.get("extract_subject", "false").lower() == "true" auto_threshold = request.form.get("auto_threshold", "false").lower() == "true" used_threshold = threshold if extract_subject: try: mask = subject_mask(f, invert=invert) except Exception as e: return jsonify({"error": f"subject extraction failed: {e}"}), 500 else: mask, used_threshold = load_and_prepare( f, threshold=threshold, invert=invert, blur=blur, auto_threshold=auto_threshold, ) paths = trace_to_paths(mask, turdsize=turdsize, alphamax=alphamax, opttolerance=opttolerance) h, w = mask.shape token = _cache_mask(mask) return jsonify({ "width": int(w), "height": int(h), "paths": paths, "svg_d": paths_to_svg_d(paths), "token": token, "used_threshold": int(used_threshold), "mode": "subject" if extract_subject else ("auto" if auto_threshold else "manual"), }) def is_line_art(img: Image.Image, extreme_frac: float = 0.88) -> bool: """True when an image is already black-and-white artwork rather than a photo. This matters more than it looks. The default mask mode is rembg subject extraction, which is a photo segmentation net: fed a line drawing it fills the fine white gaps between the lines and hands back a few large blobs. Measured on a real 120x28 panel drawing, subject extraction composed to 28 contours where a plain threshold gave 322. The drawing is destroyed before anything downstream gets a look at it. Test: a line drawing puts nearly all of its pixels at the two ends of the histogram. A photograph does not. """ g = img.convert("L") if max(g.size) > 900: # sampling is plenty for this g = g.copy() g.thumbnail((900, 900), Image.NEAREST) h = g.histogram() total = sum(h) or 1 dark = sum(h[:31]) light = sum(h[225:]) return (dark + light) / total >= extreme_frac def _bbox_of_mask_component(component_mask: np.ndarray) -> list[int]: ys, xs = np.where(component_mask) return [int(xs.min()), int(ys.min()), int(xs.max()), int(ys.max())] @app.post("/analyze") def analyze(): data = request.get_json(force=True) token = data.get("token") if not token or token not in _mask_cache: return jsonify({"error": "missing or expired token; re-trace first"}), 400 mask = _mask_cache[token] try: width_in = float(data["width_in"]) height_in = float(data["height_in"]) except (KeyError, TypeError, ValueError): return jsonify({"error": "width_in and height_in required"}), 400 if width_in <= 0 or height_in <= 0: return jsonify({"error": "panel dimensions must be > 0"}), 400 intent = data.get("intent", "cutouts") if intent not in ("cutouts", "silhouette"): intent = "cutouts" kerf_in = float(data.get("kerf_in", 0.06)) min_feature_in = float(data.get("min_feature_in", max(kerf_in * 1.5, 0.1))) # Largest ball allowed through any opening. 0 disables the check. max_gap_in = float(data.get("max_gap_in", 0.0)) gap_min_in2 = float(data.get("gap_min_in2", 0.05)) # Below these real-world thresholds, regions are considered noise (not real # cuttability problems). Tunable via request. island_min_in2 = float(data.get("island_min_in2", 0.25)) thin_cluster_min_in2 = float(data.get("thin_cluster_min_in2", 0.005)) h, w = mask.shape # mask = True for light pixels (background); design = dark shapes = ~mask. # Apply binary opening to suppress single-pixel antialiasing jaggies that # would otherwise dominate the thin-feature count on photo-traced images. from scipy.ndimage import binary_opening design_raw = ~mask design = binary_opening(design_raw, iterations=1) xs, ys = np.where(design) if xs.size == 0: return jsonify({"warnings": [], "islands": [], "overlay_png_b64": None, "summary": {"design_pixels": 0}}) # Real-world scale: prefer caller-provided in_per_px (composed panel) over # bbox-fit derivation (single-image trace). explicit_in_per_px = data.get("in_per_px") if explicit_in_per_px is not None and float(explicit_in_per_px) > 0: in_per_px = float(explicit_in_per_px) else: bb_x0, bb_x1 = int(xs.min()), int(xs.max()) bb_y0, bb_y1 = int(ys.min()), int(ys.max()) bbox_w_px = max(bb_x1 - bb_x0, 1) bbox_h_px = max(bb_y1 - bb_y0, 1) in_per_px = min(width_in / bbox_w_px, height_in / bbox_h_px) # --- Minimum feature width: distance transform of the design --- # distance_transform_edt: distance from each True pixel to nearest False pixel. # Local feature width at a pixel = 2 * distance. distance_px = distance_transform_edt(design) feature_width_in = 2.0 * distance_px * in_per_px too_thin = design & (feature_width_in < min_feature_in) & (feature_width_in > 0) too_thin_count = int(too_thin.sum()) warnings = [] if too_thin_count > 0: # cluster too-thin pixels into connected components for reporting thin_labeled, thin_n = label(too_thin) thin_components = [] cluster_min_px = max(int(thin_cluster_min_in2 / max(in_per_px ** 2, 1e-9)), 3) for i in range(1, thin_n + 1): comp = thin_labeled == i csize = int(comp.sum()) if csize < cluster_min_px: continue min_w = float((feature_width_in[comp]).min()) thin_components.append({ "size_px": csize, "area_in2": round(csize * (in_per_px ** 2), 4), "min_width_in": round(min_w, 4), "bbox": _bbox_of_mask_component(comp), }) thin_components.sort(key=lambda c: c["min_width_in"]) if thin_components: narrowest = thin_components[0]["min_width_in"] warnings.append({ "kind": "thin_feature", "severity": "high" if narrowest < min_feature_in * 0.6 else "medium", "count": len(thin_components), "narrowest_in": round(narrowest, 4), "min_required_in": round(min_feature_in, 4), "message": ( f"{len(thin_components)} thin region(s); narrowest is " f"{narrowest:.3f}\" (minimum {min_feature_in:.3f}\")" ), "regions": thin_components[:20], }) # --- Floating islands (silhouette mode only) --- islands_payload = [] if intent == "silhouette": labeled, n = label(design) if n > 1: # treat the largest component as the "main" piece sizes = np.bincount(labeled.ravel()) sizes[0] = 0 # background main_id = int(sizes.argmax()) island_min_px = max(int(island_min_in2 / max(in_per_px ** 2, 1e-9)), 1) for comp_id in range(1, n + 1): if comp_id == main_id: continue comp = labeled == comp_id csize = int(comp.sum()) if csize < island_min_px: continue area_in2 = csize * (in_per_px ** 2) islands_payload.append({ "size_px": csize, "area_in2": round(area_in2, 3), "bbox": _bbox_of_mask_component(comp), }) islands_payload.sort(key=lambda i: -i["area_in2"]) if islands_payload: warnings.append({ "kind": "floating_island", "severity": "high", "count": len(islands_payload), "message": ( f"{len(islands_payload)} steel piece(s) not connected to the " f"main piece; will fall out unless bridged" ), }) # --- Gap test: will a ball of max_gap_in pass through any opening? --- # The guard/pool rule Marsh's client is working to. A sphere is stopped by # the NARROWEST clearance, so the test is the largest circle that fits # inside an opening: distance transform of the CUT area, doubled. # # Only openings enclosed by steel are measured. On a composed panel every # cut is enclosed by the border, which is what we want. On a bare traced # image the surrounding background touches the image edge and is not a cut, # so excluding edge-touching regions keeps it from being reported. gaps_payload = [] gap_mask = np.zeros_like(design, dtype=bool) if max_gap_in > 0: cut = ~design cut_lab, cut_n = label(cut) if cut_n: edge_ids = set(cut_lab[0, :]) | set(cut_lab[-1, :]) \ | set(cut_lab[:, 0]) | set(cut_lab[:, -1]) edge_ids.discard(0) enclosed = cut & ~np.isin(cut_lab, list(edge_ids)) if enclosed.any(): gap_dist_in = 2.0 * distance_transform_edt(cut) * in_per_px over = enclosed & (gap_dist_in >= max_gap_in) min_px = max(int(gap_min_in2 / max(in_per_px ** 2, 1e-9)), 1) if over.any(): over_lab, over_n = label(over) for i in range(1, over_n + 1): comp = over_lab == i if int(comp.sum()) < min_px: continue widest = float(gap_dist_in[comp].max()) yy, xx = np.where(comp) gaps_payload.append({ "bbox": _bbox_of_mask_component(comp), "center_px": [int(xx.mean()), int(yy.mean())], "widest_in": round(widest, 3), "area_in2": round(float(comp.sum()) * in_per_px ** 2, 3), }) gap_mask |= comp gaps_payload.sort(key=lambda g: -g["widest_in"]) if gaps_payload: worst = gaps_payload[0]["widest_in"] warnings.append({ "kind": "gap_too_large", "severity": "high", "count": len(gaps_payload), "worst_in": worst, "max_allowed_in": round(max_gap_in, 3), "message": ( f"{len(gaps_payload)} opening(s) pass a {max_gap_in:.2f}\" ball; " f"the widest fits {worst:.2f}\"" ), "regions": gaps_payload[:20], }) # --- Build a warning overlay PNG for the canvas --- overlay = np.zeros((h, w, 4), dtype=np.uint8) if too_thin_count > 0: overlay[too_thin] = [255, 60, 60, 220] if gap_mask.any(): # openings that pass the ball, in blue overlay[gap_mask] = [40, 130, 255, 190] if islands_payload: # mark islands in orange labeled, _ = label(design) sizes = np.bincount(labeled.ravel()) sizes[0] = 0 main_id = int(sizes.argmax()) for island in islands_payload: x0, y0, x1, y1 = island["bbox"] sub_labels = labeled[y0:y1 + 1, x0:x1 + 1] island_pixels = (sub_labels != main_id) & (sub_labels != 0) yy, xx = np.where(island_pixels) overlay[yy + y0, xx + x0] = [255, 145, 0, 200] overlay_b64 = None if too_thin_count > 0 or islands_payload or gap_mask.any(): img = Image.fromarray(overlay, mode="RGBA") buf = io.BytesIO() img.save(buf, format="PNG") overlay_b64 = base64.b64encode(buf.getvalue()).decode("ascii") summary = { "design_pixels": int(design.sum()), "in_per_px": round(in_per_px, 5), "min_feature_in": min_feature_in, "kerf_in": kerf_in, "intent": intent, "max_gap_in": max_gap_in, "px_per_inch": round(1.0 / in_per_px, 2) if in_per_px else None, "gap_count": len(gaps_payload), "worst_gap_in": (gaps_payload[0]["widest_in"] if gaps_payload else None), } return jsonify({ "warnings": warnings, "islands": islands_payload, "gaps": gaps_payload, "overlay_png_b64": overlay_b64, "summary": summary, }) def simplify_polyline(pts, tol): """Douglas-Peucker. Drops collinear-enough vertices without moving the outline more than `tol`.""" n = len(pts) if n < 3 or tol <= 0: return pts keep = [False] * n keep[0] = keep[-1] = True stack = [(0, n - 1)] while stack: i, j = stack.pop() if j <= i + 1: continue x1, y1 = pts[i] x2, y2 = pts[j] ddx, ddy = x2 - x1, y2 - y1 seg = (ddx * ddx + ddy * ddy) ** 0.5 best, bi = -1.0, -1 for k in range(i + 1, j): x0, y0 = pts[k] if seg > 0: d = abs(ddy * x0 - ddx * y0 + x2 * y1 - y2 * x1) / seg else: d = ((x0 - x1) ** 2 + (y0 - y1) ** 2) ** 0.5 if d > best: best, bi = d, k if best > tol and bi > 0: keep[bi] = True stack.append((i, bi)) stack.append((bi, j)) return [q for q, k in zip(pts, keep) if k] def _parse_export_payload(): data = request.get_json(force=True) paths = data.get("paths") or [] width_in = float(data.get("width_in", 0)) height_in = float(data.get("height_in", 0)) units = data.get("units", "inches") in_per_px = data.get("in_per_px") in_per_px = float(in_per_px) if in_per_px else None if width_in <= 0 or height_in <= 0: raise ValueError("width_in and height_in must be > 0") return paths, width_in, height_in, units, in_per_px def _export_transform(paths, w_in, h_in, in_per_px): """Return (scale, dx, dy). When in_per_px is given (composed panel), the paths are already in panel pixel coords; just scale by in_per_px. Otherwise fall back to bbox-fit-into-panel.""" if in_per_px and in_per_px > 0: return in_per_px, 0.0, 0.0, w_in, h_in scale, dx, dy, fw, fh = fit_transform(paths, w_in, h_in, True) return scale, dx, dy, fw, fh @app.post("/export/svg") def export_svg(): try: paths, w_in, h_in, units, in_per_px = _parse_export_payload() except ValueError as e: return jsonify({"error": str(e)}), 400 scale, dx, dy, fw, fh = _export_transform(paths, w_in, h_in, in_per_px) parts = [ f'<?xml version="1.0" encoding="UTF-8"?>', f'<svg xmlns="http://www.w3.org/2000/svg" ' f'width="{w_in}in" height="{h_in}in" ' f'viewBox="0 0 {w_in} {h_in}">', f'<g fill="black" stroke="none" fill-rule="evenodd">', ] for sub in paths: sx = sub["start"][0] * scale + dx sy = sub["start"][1] * scale + dy d = [f"M {sx:.4f} {sy:.4f}"] for seg in sub["segments"]: ex = seg["end"][0] * scale + dx ey = seg["end"][1] * scale + dy if seg["type"] == "C": cx = seg["c"][0] * scale + dx cy = seg["c"][1] * scale + dy d.append(f"L {cx:.4f} {cy:.4f} L {ex:.4f} {ey:.4f}") else: c1x = seg["c1"][0] * scale + dx c1y = seg["c1"][1] * scale + dy c2x = seg["c2"][0] * scale + dx c2y = seg["c2"][1] * scale + dy d.append(f"C {c1x:.4f} {c1y:.4f} {c2x:.4f} {c2y:.4f} {ex:.4f} {ey:.4f}") d.append("Z") parts.append(f'<path d="{" ".join(d)}"/>') parts.append("</g></svg>") svg = "\n".join(parts).encode("utf-8") return send_file( io.BytesIO(svg), mimetype="image/svg+xml", as_attachment=True, download_name="plasma-design.svg", ) @app.post("/export/dxf") def export_dxf(): try: paths, w_in, h_in, units, in_per_px = _parse_export_payload() except ValueError as e: return jsonify({"error": str(e)}), 400 scale, dx, dy, fw, fh = _export_transform(paths, w_in, h_in, in_per_px) polylines = flatten_paths(paths, bezier_steps=32) # DXF uses Y-up; bitmap traces are Y-down. Flip so the design isn't mirrored. polylines = transform_polylines(polylines, scale, dx, dy, flip_y_height=h_in) doc = ezdxf.new(dxfversion="R2010", units=ezdxf.units.IN if units == "inches" else ezdxf.units.MM) doc.header["$INSUNITS"] = 1 if units == "inches" else 4 msp = doc.modelspace() doc.layers.add(name="CUT", color=1) # Flattening beziers at 32 steps is far finer than a torch can follow, and # at 40 px/inch on a 120" panel it produced a 34MB DXF that CAM software # will not thank anyone for. 0.01" is well inside a plasma kerf, so this is # invisible on the part. simplify_in = float(request.get_json(force=True).get("simplify_in", 0.01)) kept = dropped = 0 for pts in polylines: if len(pts) < 2: continue before = len(pts) if simplify_in > 0: pts = simplify_polyline(pts, simplify_in) kept += len(pts) dropped += before - len(pts) if len(pts) < 2: continue msp.add_lwpolyline(pts, close=True, dxfattribs={"layer": "CUT"}) app.logger.info("dxf export: %d vertices kept, %d dropped (simplify %.3fin)", kept, dropped, simplify_in) buf = io.StringIO() doc.write(buf) data = buf.getvalue().encode("utf-8") return send_file( io.BytesIO(data), mimetype="application/dxf", as_attachment=True, download_name="plasma-design.dxf", ) if __name__ == "__main__": import os port = int(os.environ.get("PORT", 5174)) host = os.environ.get("HOST", "127.0.0.1") app.run(host=host, port=port, debug=False)
"""
Marshall's Plasma Design Studio.
Pipeline: client reference image -> auto-trace -> dimensioned vector -> DXF/SVG
for the plasma table.
"""
from __future__ import annotations
import base64
import io
import math
import os
import uuid
from collections import OrderedDict
from pathlib import Path
import ezdxf
import numpy as np
from flask import Flask, jsonify, request, send_file, send_from_directory
from PIL import Image, ImageOps
from potrace import Bitmap, POTRACE_TURNPOLICY_MINORITY
from scipy.ndimage import distance_transform_edt, label
STATIC_DIR = Path(__file__).parent / "static"
app = Flask(__name__, static_folder=str(STATIC_DIR), static_url_path="/static")
MAX_IMAGE_DIM = 4200 # pixels; downscale large uploads before tracing/segmentation
# 1200 was sized for a client photo traced onto a two-foot sign. On a ten
# foot fence panel it lands at ~9 px/inch, which is one pixel per 0.105" and
# destroys fine linework (snow cracks, treeline needles, a fly line) on
# import, before the trace ever runs.
MAX_PANEL_PX = 6000 # cap composed-panel pixel dimensions to keep tracing snappy
# Raising MAX_IMAGE_DIM alone does nothing: this cap independently throttles
# a 120" panel to 18 px/inch. At 6000 a 120" panel holds the full 40 px/inch
# (0.025"/px), which is finer than any plasma kerf.
MASK_CACHE_LIMIT = 8 # most recent composed/traced masks kept for /analyze
_mask_cache: "OrderedDict[str, np.ndarray]" = OrderedDict()
IMAGE_CACHE_LIMIT = 32 # most recent uploaded images, keyed by token
_image_cache: "OrderedDict[str, Image.Image]" = OrderedDict()
LAYER_MASK_CACHE_LIMIT = 64 # cache per-layer masks, keyed by (token + mode + params hash)
_layer_mask_cache: "OrderedDict[str, np.ndarray]" = OrderedDict()
def _cache_mask(mask: np.ndarray) -> str:
token = uuid.uuid4().hex
_mask_cache[token] = mask
while len(_mask_cache) > MASK_CACHE_LIMIT:
_mask_cache.popitem(last=False)
return token
def _cache_image(img: Image.Image) -> str:
token = uuid.uuid4().hex
_image_cache[token] = img
while len(_image_cache) > IMAGE_CACHE_LIMIT:
_image_cache.popitem(last=False)
return token
def _cache_layer_mask(key: str, mask: np.ndarray) -> None:
_layer_mask_cache[key] = mask
_layer_mask_cache.move_to_end(key)
while len(_layer_mask_cache) > LAYER_MASK_CACHE_LIMIT:
_layer_mask_cache.popitem(last=False)
def _otsu_threshold(arr: np.ndarray) -> int:
"""Otsu's method: pick the threshold that maximizes inter-class variance."""
hist, _ = np.histogram(arr, bins=256, range=(0, 256))
total = arr.size
if total == 0:
return 128
sum_total = float((np.arange(256) * hist).sum())
sum_b = 0.0
w_b = 0
var_max = 0.0
threshold = 128
for t in range(256):
w_b += int(hist[t])
if w_b == 0:
continue
w_f = total - w_b
if w_f == 0:
break
sum_b += t * float(hist[t])
m_b = sum_b / w_b
m_f = (sum_total - sum_b) / w_f
var_between = w_b * w_f * (m_b - m_f) ** 2
if var_between > var_max:
var_max = var_between
threshold = t
return threshold
def _open_resized(file_storage, mode: str = "RGB") -> Image.Image:
img = Image.open(file_storage.stream).convert(mode)
img = ImageOps.exif_transpose(img)