TEED edges
TEED (Soria et al., 2023): a tiny edge detector (58 k parameters) that draws soft edge maps for the soft-edge ControlNet, quantized to a few levels by default. MIT code and weights (trained on BIPED, a research-use dataset; see the card).
experimental Fork of huggingface/controlnet_aux@b1aac31 · backends: portable, kernels
- Paper Tiny and Efficient Model for The Edge Detection Generalization ICCV 2023 workshops · first released 12 Aug 2023
Checkpoint teed
Paper 2023 Added in 0.2.0
Use it
pip install fastforks Lean install with only what this model needs: pip install fastforks-core
from fastforks import edges
model = edges.load("teed") # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.edges
result.backend # what ran: mode, device, dtype, kernels fastforks bench teed Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.
Speed and memory
Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | not measured | – | – | – | – |
| Original fp16torch.autocast(float16) around the call; huggingface/controlnet_aux@b1aac318b15c teed, host OpenCV resize, float post-processing | fp16 | 29.6 | – | 149 MiB/184 MiB | 0.1 | 0.09 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 4.09 | 7.2× | 122 MiB/146 MiB | 0.2 | 0.01 |
| Kernels | not measured | – | – | – | – | |
| Portable + torch.compile | not measured here | – | – | – | – | |
| TensorRT | not measured here | – | – | – | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-14. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two
times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.
Other devices
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14 | Original | fp32 | 25.8 | 1,434 MiB process | – | 1024x1024 |
| Portable | fp16 | 16.7 | 1,313 MiB process | matches | 1024x1024 | |
| Portable | fp32 | 20.1 | 1,314 MiB process | matches | 1024x1024 | |
| Kernels | fp32 | not implemented | – | 1024x1024 | ||
CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.
Licence and weights
Commercial use allowed
- Code
- MIT LICENSE
- Weights
- MIT source
- Training data
-
- BIPED (Barcelona Images for Perceptual Edge Detection) research-only, non-commercial
Our reading of the licences, not legal advice.
Weights
teed_5_model.pth 249 KB· hosted on raw.githubusercontent.comfastforks downloads and checks these files on first use (fastforks download teed).
Accuracy
The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.
Accuracy not measured for teed.
Closeness to the original
Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 and p99 <= 0.25 (absolute error of the output edges).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.002 | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.02 | 0.00023 |
| GPU default dtypes, p99 |diff| | ≤ 0.25 | 0.00137 |
Recorded for teed on RTX 5060 Ti, torch 2.10.0+cu130, 2026-11-17, with these kernels: smish
Details
Tiny and efficient edge detector (Soria et al., ICCV 2023 workshops; controlnet_aux `TEEDdetector`, ComfyUI's "TEED Soft-Edge Lines"): a 58 k-parameter network whose four outputs are averaged into a soft edge map, quantized to a few levels by default.
- Input / output
- Input: one RGB image. Output: a line / edge map
[H, W]in [0, 1] at the input size (1 = line), the ControlNet preprocessors (line art, HED, PiDiNet, TEED / Anyline); M-LSD also returns straight line segments. - Input size
- any H x W >= 32, padded to a multiple of 32; tested up to 1080x1920 (122 MiB allocated)
- Options
short_side=512>= 64· changes the outputsafe_steps=20 to 255· changes the output
- Outputs
edges- CUDA kernels
smish
- Apple (MPS)
- fp32, fp16
Full reference, generated from the repository
TEED (Soria et al., 2023): a tiny edge detector (58 k parameters) that draws soft edge maps for the soft-edge ControlNet, quantized to a few levels by default. MIT code and weights (trained on BIPED, a research-use dataset; see the card).
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
teed | synthetic (1080x1920) | original portable kernels | 29.6 4.09 – | 149 / 184 122 / 146 – |
teed | synthetic (1024x1024) | original portable kernels portable + torch.compile TensorRT fp16 | 15.8 2.54 – 1.10 0.799 | 89 / 110 68 / 88 – 71 / 104 106 / 120 |
Details
- Original code: https://github.com/huggingface/controlnet_aux at
b1aac318b15c - Code licence: MIT (permissive); files:
LICENSE - Weights licence: MIT
- Licence evidence: https://github.com/xavysp/TEED/blob/40fa4b1391dc6424f88989d0ca75d5b592c8681d/LICENSE ; https://github.com/xavysp/TEED/tree/40fa4b1391dc6424f88989d0ca75d5b592c8681d/checkpoints/BIPED/5
- Training data: BIPED (Barcelona Images for Perceptual Edge Detection) (BIPED (the repository's checkpoints/BIPED folder and README); the dataset is for non-commercial use.)
- Options:
short_side= 512 (>= 64; alsoresolution; alsodetect_resolution),safe_steps= 2 (0 to 255) - Kernel libraries:
smish - Variants:
teed - Outputs: edges
- Input size: any H x W >= 32, padded to a multiple of 32
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
teed1080x1920 (0.1 GiB peak allocated, 0.1 reserved) teed1024x1024, first call: portable + torch.compile 9 s (cold Inductor cache), TensorRT fp16 12 s (ONNX + engine).teedoriginal 1080x1920, original 1024x1024: torch.autocast(float16) around the call; huggingface/controlnet_aux@b1aac318b15c teed, host OpenCV resize, float post-processingteedload / first call: original 0.13 s / 0.083 s, portable 0.16 s / 0.005 s- Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5
Parity
CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 and p99 <= 0.25 (absolute error of the output edges).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
teed | 0.00023 / 0.00137 | RTX 5060 Ti, torch 2.10.0+cu130 | smish@c9aaf63e6ef7:sm_120 |
Example output
No example rendered for this checkpoint yet.