unsloth/studio/backend/core/inference/diffusion_controlnet.py
Daniel Han 060fac0a9d ControlNet: reject filesystem-like ids and do not cache a model past an unload race
Two review findings on the ControlNet path:
- resolve_controlnet's bare-repo fallback accepted any id with a slash, so a
  path-shaped id (/tmp/x, ../x) reached from_pretrained as a local directory.
  Restrict the fallback to a strict owner/name HF repo id shape.
- _controlnet_pipe now re-checks the cancel event after the blocking
  from_pretrained: an unload that raced the download had already cleared the
  caches, so caching the late module would pin it past the unload.
2026-07-01 23:52:44 +00:00

224 lines
8.8 KiB
Python

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
"""Diffusion ControlNet support: family-gated discovery of ControlNet models, resolution to
a loadable diffusers repo/dir, control-image preprocessing, and a capability gate.
Mirrors ``diffusion_lora.py``. Two differences from LoRA: (1) a ControlNet is a full diffusers
repo (loaded via ``from_pretrained``), not a single-file adapter, so resolution yields a repo id
or local directory rather than a file path; (2) ControlNet needs a spatial *control image*, which
is either supplied already-preprocessed ("passthrough", as in ComfyUI where preprocessing is a
separate step) or derived here ("canny", a dependency-free edge map).
ControlNet models are architecture-specific (a FLUX ControlNet cannot drive a Qwen base), so
discovery is family-gated exactly like the LoRA picker. The request never carries a filesystem
path -- only a discovery id or a public ``owner/name`` repo id -- so a client cannot make the
backend read an arbitrary location.
"""
from __future__ import annotations
import re
import threading
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Optional
from utils.paths.storage_roots import studio_root
# Control map types. "passthrough": the supplied image IS the control map (a depth/pose/etc.
# map produced elsewhere). "canny": derive an edge map here (no heavy detector dependency).
CONTROL_TYPES = ("passthrough", "canny")
# Families whose diffusers pipeline supports ControlNet (declared in diffusion_families via
# controlnet_pipeline_class). Native sd.cpp ControlNet is a follow-up. Torchao fp8/int8 dense
# and GGUF-via-diffusers are gated off, same rule as LoRA.
_DIFFUSERS_BLOCKED_QUANT = ("int8", "fp8", "nvfp4", "mxfp8")
@dataclass(frozen = True)
class ControlNetCatalogEntry:
"""One discoverable ControlNet model."""
id: str
display_name: str
source: str # "local" | "hub"
families: tuple[str, ...] = () # compatible family names (empty = shown, not gated)
repo_id: Optional[str] = None # for source == "hub"
local_path: Optional[str] = None # for source == "local"
control_types: tuple[str, ...] = ("passthrough",) # recommended control types
is_union: bool = False # a single model covering many control modes
@dataclass(frozen = True)
class ResolvedControlNet:
"""A ControlNet resolved to something ``from_pretrained`` can load."""
id: str
path: str # repo id (hub) or local directory
is_local: bool
# Curated, family-tagged catalog. Union models (one model, many control modes) dominate real
# usage, so they are the default picks. Extend as more are curated; local dirs + a bare public
# ``owner/name`` repo id also work.
_CURATED: tuple[ControlNetCatalogEntry, ...] = (
ControlNetCatalogEntry(
id = "flux-union-pro",
display_name = "FLUX.1 ControlNet Union Pro (Shakker-Labs)",
source = "hub",
families = ("flux.1",),
repo_id = "Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro",
control_types = ("canny", "depth", "pose", "passthrough"),
is_union = True,
),
ControlNetCatalogEntry(
id = "qwen-union",
display_name = "Qwen-Image ControlNet Union (InstantX)",
source = "hub",
families = ("qwen-image",),
repo_id = "InstantX/Qwen-Image-ControlNet-Union",
control_types = ("canny", "depth", "pose", "passthrough"),
is_union = True,
),
)
def controlnets_dir() -> Path:
"""Local directory Studio scans for user-provided ControlNet model folders."""
d = studio_root() / "controlnets" / "diffusion"
d.mkdir(parents = True, exist_ok = True)
return d
def sanitize_id(raw: str) -> str:
"""Filesystem-safe id from a repo id / folder name."""
stem = raw.rsplit("/", 1)[-1]
stem = re.sub(r"[^A-Za-z0-9._-]+", "_", stem).strip("._-")
return stem or "controlnet"
def _scan_local() -> list[ControlNetCatalogEntry]:
"""A local ControlNet is a directory containing a diffusers config + weights."""
entries: list[ControlNetCatalogEntry] = []
root = controlnets_dir()
try:
children = sorted(root.iterdir())
except OSError:
return entries
for p in children:
if not p.is_dir():
continue
if not (p / "config.json").exists():
continue
entries.append(
ControlNetCatalogEntry(
id = p.name,
display_name = p.name,
source = "local",
local_path = str(p),
control_types = CONTROL_TYPES,
)
)
return entries
def list_controlnets(*, family: Optional[str] = None) -> list[ControlNetCatalogEntry]:
"""Merged catalog (curated + local), optionally family-filtered. Cheap: one dir scan
plus the in-memory curated list. Network is only touched on resolve()."""
merged = list(_CURATED) + _scan_local()
if family:
fam = family.strip().lower()
merged = [e for e in merged if not e.families or fam in {f.lower() for f in e.families}]
merged.sort(key = lambda e: (e.source != "local", e.display_name.lower()))
return merged
def _catalog_by_id() -> dict[str, ControlNetCatalogEntry]:
return {e.id: e for e in (list(_CURATED) + _scan_local())}
def resolve_controlnet(
spec_id: str,
*,
family: Optional[str] = None,
hf_token: Optional[str] = None,
cancel_event: Optional[threading.Event] = None,
) -> ResolvedControlNet:
"""Resolve a ControlNet id to a loadable repo id / local dir.
Accepts a catalog/local id, or a bare public HF repo id (``owner/name``). The backend
loads the result with ``ControlNetModelClass.from_pretrained(path)`` (download + cache
handled there, like the base pipeline). Raises on an unknown id -> the caller maps to 400.
"""
entry = _catalog_by_id().get(spec_id)
if entry is not None:
if entry.source == "local":
path = entry.local_path or ""
if not path or not Path(path).is_dir():
raise FileNotFoundError(f"ControlNet '{spec_id}' is no longer present on disk")
return ResolvedControlNet(spec_id, path, is_local = True)
if not entry.repo_id:
raise ValueError(f"ControlNet '{spec_id}' has no repo")
return ResolvedControlNet(spec_id, entry.repo_id, is_local = False)
# A bare public HF repo id (owner/name). STRICT shape -- exactly one slash and
# alphanumeric-leading segments -- so a filesystem-looking id (/tmp/x, ../x, ~/x,
# C:\x) can never reach from_pretrained, which would happily treat it as a local
# directory and bypass the controlnets_dir() no-raw-path contract.
if re.fullmatch(r"[A-Za-z0-9][A-Za-z0-9_.-]*/[A-Za-z0-9][A-Za-z0-9_.-]*", spec_id):
return ResolvedControlNet(spec_id, spec_id, is_local = False)
raise FileNotFoundError(
f"unknown ControlNet '{spec_id}': not a local model, catalog entry, or HF repo id"
)
def preprocess_control(image: Any, control_type: str) -> Any:
"""Turn a source image into a control map.
``passthrough`` returns the image unchanged (it is already a depth/pose/edge map made
elsewhere). ``canny`` derives a dependency-free gradient edge map (a rough stand-in for a
true Canny; a cv2/kornia detector and depth/pose detectors are a follow-up). Unknown types
pass through so a new type never hard-fails generation.
"""
ct = (control_type or "passthrough").strip().lower()
if ct != "canny":
return image
import numpy as np
from PIL import Image
gray = np.asarray(image.convert("L"), dtype = np.float32)
gy, gx = np.gradient(gray)
mag = np.hypot(gx, gy)
peak = float(mag.max())
if peak <= 1e-6:
return image # flat image -> nothing to trace; don't emit a black map
mag = mag / peak * 255.0
edges = (mag > 40.0).astype(np.uint8) * 255 # white edges on black, the ControlNet convention
return Image.fromarray(edges).convert("RGB")
def supports_controlnet(
*,
engine: str,
family: Optional[str],
has_controlnet_pipeline: bool,
model_kind: Optional[str],
transformer_quant: Optional[str],
) -> bool:
"""Whether the loaded model can apply a ControlNet.
diffusers only for now (native sd.cpp is a follow-up). Requires the family to declare a
ControlNet pipeline. Blocked for the diffusers GGUF path and torchao fp8/int8 dense
(same constraints as LoRA): those transformers cannot host the extra conditioning cleanly.
"""
if not family or not has_controlnet_pipeline:
return False
if engine != "diffusers":
return False
if model_kind == "gguf":
return False
if transformer_quant and str(transformer_quant).strip().lower() in _DIFFUSERS_BLOCKED_QUANT:
return False
return True