"""Remote tool calls to public Hugging Face Spaces via gradio_client. ZeroGPU billing: when this Space forwards the visitor's `x-ip-token` header, the remote ZeroGPU Space bills the *visitor's* free daily GPU minutes, not the owner's (see https://huggingface.co/docs/hub/spaces-zerogpu). No HF token of ours is used. Endpoint signatures were read with Client.view_api() on Oct 8, 2026 and saved in ../sources/hf-api/*.json. Community Spaces can change; update the wrappers if they do. MiniMax-H3 is intentionally NOT included (its license bans use in the US, incl. outputs). """ from __future__ import annotations import json, os, tempfile, time, urllib.request from typing import Any try: from gradio_client import Client, handle_file except Exception: # pragma: no cover Client = None handle_file = None # tool key -> ordered list of Spaces (primary first, then fallbacks) TOOLS: dict[str, list[str]] = { "edit": ["black-forest-labs/FLUX.1-Kontext-Dev", "Qwen/Qwen-Image-Edit-2509"], "same_face": ["yanze/PuLID-FLUX", "InstantX/InstantID"], "outpaint": ["fffiloni/diffusers-image-outpaint"], "i2v": ["zerogpu-aoti/wan2-2-fp8da-aoti-faster", "Lightricks/ltx-video-distilled"], "flf_loop": ["linoyts/LTX-2-3-First-Last-Frame"], "upscale": ["finegrain/finegrain-image-enhancer"], } LICENSE_NOTES = { "black-forest-labs/FLUX.1-Kontext-Dev": "FLUX.1 [dev] Non-Commercial License for the model; its terms allow commercial use of outputs (verify before selling)", "Qwen/Qwen-Image-Edit-2509": "Apache-2.0", "yanze/PuLID-FLUX": "PuLID Apache-2.0 on a FLUX.1 [dev] base (same output terms as above)", "InstantX/InstantID": "Apache-2.0 (SDXL base, OpenRAIL++)", "fffiloni/diffusers-image-outpaint": "community Space (SDXL ControlNet Union; OpenRAIL++)", "zerogpu-aoti/wan2-2-fp8da-aoti-faster": "Wan2.2: Apache-2.0", "Lightricks/ltx-video-distilled": "LTXV Open Weights License (free under $10M revenue)", "linoyts/LTX-2-3-First-Last-Frame": "LTX-2 Community License (free under $10M revenue)", "finegrain/finegrain-image-enhancer": "community Space (SD1.5 + ESRGAN; check model licenses)", } DEFAULT_TIMEOUT = 600 _status_cache: dict[str, tuple[float, str]] = {} class ToolError(RuntimeError): pass def _friendly(e: Exception) -> str: msg = str(e) low = msg.lower() if any(k in low for k in ("quota", "zerogpu", "gpu task aborted", "exceeded", "you have exceeded")): return "Daily free GPU minutes used up for this tool. Log in to Hugging Face for more minutes, or try again tomorrow." if "timeout" in low or "timed out" in low: return "The remote tool took too long. It may be busy; try again in a minute." if "runtime_error" in low or "could not fetch config" in low or "space is not running" in low: return "That remote tool is down right now." return msg[:400] def space_status(space: str) -> str: now = time.time() if space in _status_cache and now - _status_cache[space][0] < 60: return _status_cache[space][1] try: with urllib.request.urlopen(f"https://huggingface.co/api/spaces/{space}/runtime", timeout=6) as r: stage = json.load(r).get("stage", "UNKNOWN") except Exception: stage = "UNKNOWN" _status_cache[space] = (now, stage) return stage def status_table() -> list[list[str]]: rows = [] for tool, spaces in TOOLS.items(): for i, sp in enumerate(spaces): st = space_status(sp) light = "🟢" if st == "RUNNING" else ("🟡" if st in ("SLEEPING", "BUILDING", "APP_STARTING", "PAUSED") else "🔴") rows.append([tool, "primary" if i == 0 else "fallback", sp, f"{light} {st}", LICENSE_NOTES.get(sp, "")]) return rows def _client(space: str, request=None): if Client is None: raise ToolError("gradio_client is not installed.") headers = {} if request is not None: try: tok = request.headers.get("x-ip-token") if tok: headers["x-ip-token"] = tok except Exception: pass return Client(space, headers=headers or None, verbose=False, download_files=tempfile.mkdtemp(prefix="aiking_dl_")) def _first_path(x: Any) -> str | None: """Pull a local file path out of whatever a Space returned.""" if x is None: return None if isinstance(x, str): return x if isinstance(x, dict): for k in ("path", "video", "image", "name", "url"): if k in x and x[k]: return _first_path(x[k]) if isinstance(x, (list, tuple)): for item in x: p = _first_path(item) if p: return p return None def _call(space, api_name, request=None, timeout=DEFAULT_TIMEOUT, **kwargs): job = _client(space, request).submit(api_name=api_name, **kwargs) return job.result(timeout=timeout) def _with_fallback(tool: str, runners: dict, request=None): errors = [] for sp in TOOLS[tool]: if space_status(sp) in ("RUNTIME_ERROR", "BUILD_ERROR", "CONFIG_ERROR", "NO_APP_FILE", "STOPPED"): errors.append(f"{sp}: down") continue try: return runners[sp](sp), sp except Exception as e: errors.append(f"{sp}: {_friendly(e)}") raise ToolError(" | ".join(errors)) # ---------- tool wrappers ---------- def edit_image(image_path, prompt, request=None): runners = { "black-forest-labs/FLUX.1-Kontext-Dev": lambda sp: _first_path(_call( sp, "/infer", request, input_image=handle_file(image_path), prompt=prompt, seed=0, randomize_seed=True, guidance_scale=2.5, steps=28)), "Qwen/Qwen-Image-Edit-2509": lambda sp: _first_path(_call( sp, "/infer", request, images=[{"image": handle_file(image_path), "caption": None}], prompt=prompt, seed=0, randomize_seed=True, true_guidance_scale=4.0, num_inference_steps=40, height=1024, width=1024, rewrite_prompt=False)), } return _with_fallback("edit", runners, request) def same_face(face_path, prompt, negative, width=896, height=1152, request=None): runners = { "yanze/PuLID-FLUX": lambda sp: _first_path(_call( sp, "/generate_image", request, prompt=prompt, id_image=handle_file(face_path), start_step=0, guidance=4, seed="-1", true_cfg=1, width=width, height=height, num_steps=20, id_weight=1, neg_prompt=negative, timestep_to_start_cfg=1, max_sequence_length=128)), "InstantX/InstantID": lambda sp: _first_path(_call( sp, "/generate_image", request, face_image_path=handle_file(face_path), pose_image_path=None, prompt=prompt, negative_prompt=negative, style_name="(No style)", num_steps=30, identitynet_strength_ratio=0.8, adapter_strength_ratio=0.8, canny_strength=0.4, depth_strength=0.4, controlnet_selection=["depth"], guidance_scale=5.0, seed=42, scheduler="EulerDiscreteScheduler", enable_LCM=False, enhance_face_region=True)), } return _with_fallback("same_face", runners, request) def outpaint(image_path, width, height, prompt="", alignment="Middle", request=None): runners = { "fffiloni/diffusers-image-outpaint": lambda sp: _first_path(list(reversed(list(_call( sp, "/infer", request, image=handle_file(image_path), width=int(width), height=int(height), overlap_percentage=10, num_inference_steps=8, resize_option="Full", custom_resize_percentage=50, prompt_input=prompt or "", alignment=alignment, overlap_left=True, overlap_right=True, overlap_top=True, overlap_bottom=True) or [])))), } return _with_fallback("outpaint", runners, request) def image_to_video(image_path, prompt, negative="", seconds=5.0, request=None): runners = { "zerogpu-aoti/wan2-2-fp8da-aoti-faster": lambda sp: _first_path(_call( sp, "/generate_video", request, input_image=handle_file(image_path), prompt=prompt, steps=6, negative_prompt=negative or "worst quality, blurry, distorted face, extra limbs, text, watermark", duration_seconds=float(seconds), guidance_scale=1, guidance_scale_2=1, seed=42, randomize_seed=True)), "Lightricks/ltx-video-distilled": lambda sp: _first_path(_call( sp, "/image_to_video", request, prompt=prompt, negative_prompt=negative or "worst quality, inconsistent motion, blurry, jittery, distorted", input_image_filepath=handle_file(image_path), input_video_filepath=None, height_ui=512, width_ui=704, mode="image-to-video", duration_ui=min(float(seconds), 8.5), ui_frames_to_use=9, seed_ui=42, randomize_seed=True, ui_guidance_scale=1, improve_texture_flag=True)), } return _with_fallback("i2v", runners, request) def first_last_loop(first_path, last_path, prompt, seconds=4.0, request=None): def run(sp): w, h = 1536, 1024 try: w, h = _call(sp, "/on_image_upload", request, timeout=60, first_image=handle_file(first_path), last_image=handle_file(last_path or first_path), high_res=False) except Exception: pass return _first_path(_call( sp, "/generate_video", request, first_image=handle_file(first_path), last_image=handle_file(last_path or first_path), input_audio=None, prompt=prompt, duration=float(seconds), enhance_prompt=False, seed=10, randomize_seed=True, height=int(h), width=int(w))) return _with_fallback("flf_loop", {"linoyts/LTX-2-3-First-Last-Frame": run}, request) def upscale(image_path, factor=2, prompt="masterpiece, best quality, highres", request=None): runners = { "finegrain/finegrain-image-enhancer": lambda sp: _first_path(list(reversed(list(_call( sp, "/process", request, input_image=handle_file(image_path), prompt=prompt, negative_prompt="worst quality, low quality, blurry, watermark", seed=42, upscale_factor=float(factor), controlnet_scale=0.6, controlnet_decay=1.0, condition_scale=6, tile_width=112, tile_height=144, denoise_strength=0.35, num_inference_steps=18, solver="DDIM") or [])))), } return _with_fallback("upscale", runners, request)