Models · Feature backbones

TIPSv2

Google DeepMind's TIPSv2 image encoders (B, L): image features aligned with text, with patch features meant for dense tasks. Apache-2.0.

experimental Fork of google-deepmind/tips@bf10a73 · backends: portable, kernels

Checkpoint

Paper 2026 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 1.49 ms 224x224, fp16
vs the faster original 4.7× original fp32: 7.03 ms
Peak memory 206 MiB original: 373 MiB
Accuracy vs original fp32 −0.18 kernels top-1 %, ImageNet-1k val (59.34 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install fastforks-core

from fastforks import features

model = features.load("tipsv2-b14")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.last_hidden_state  # also: pooler_output, hidden_states
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench tipsv2-b14

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.

Input: synthetic, 224x224 (the task’s standard size)

BackenddtypeTime per call (ms)vs original fp32Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32tips image_encoder first cls tokenfp32 7.03–373 MiB/386 MiB1.00.27
Original fp16tips image_encoder first cls tokenfp16 7.210.98×529 MiB/548 MiB0.90.07
Portable fp32fp32not measured ––––
Portablefp16 2.932.4×204 MiB/224 MiB0.20.18
Kernelsfp16 1.494.7×206 MiB/226 MiB1.30.11
Portable + torch.compile 2.183.2×239 MiB/266 MiB9 (build)–
TensorRT fp16fp16 1.335.3×198 MiBenginenot comparable42 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-10 to 2026-11-16. 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

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-13Originalfp3231.7 930 MiB process–224x224
Portablefp1623.1 775 MiB processmatches224x224
Portablefp3224.6 895 MiB processmatches224x224
Kernelsfp16not implemented–224x224

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
Apache-2.0 LICENSE
Weights
Apache-2.0 source
Training data
image-text pairs; the dataset is not named in the repository README or on the model cards at the pinned revisions

Our reading of the licences, not legal advice.

Weights

tipsv2-b14/model.safetensors 784 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download tipsv2-b14).

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.

ImageNet-1k val

kNN top-1 % (k = 20, 10000 labelled / 5000 test images)

Rundtypetop-1 % ↑Δ top-1 % vs fp32 originalimagesms / image
Original fp32fp3259.34 reference 5000 5.20
Original fp16fp1659.30 −0.04 5000 2.12
Portable fp32not measured
Portablefp1659.36 +0.02 5000 1.36
Kernelsfp1659.16 −0.18 5000 0.968

Two disjoint seeded subsets of the val folders; each image shorter side 256 (bicubic), centre crop 224, the CLS / pooled feature; cosine kNN, vote weighted by similarity. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-10; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 8b2f2df5, 2026-11-10.

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| <= 6.9e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 1.9 % and p99 <= 6.9 % of the reference's median |value| (absolute error of the output last_hidden_state).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.000069 (relative)not recorded
GPU default dtypes, mean |diff|≤ 0.019 (relative)0.01252 (relative)
GPU default dtypes, p99 |diff|≤ 0.069 (relative)0.046 (relative)

Recorded for tipsv2-b14 on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12, with these kernels: dpt_fusedfastbb

Details

TIPSv2 (Google DeepMind, CVPR 2026) image encoders: ViT-B / L at patch 14 from text-image pretraining with patch-text alignment; dense patch tokens and two global tokens at any size (native 448 px).

Input / output
Input: one RGB image. Output: dense patch features and a pooled image feature.
Input size
any H x W >= 1, resized or padded inside the model; tested up to 3000x4000 (1,467 MiB allocated)
Options
  • fp16_accumulate=false
Outputs
last_hidden_state, pooler_output, hidden_states
Checkpoint
arch="vitb"
CUDA kernels
  • fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbones
  • dpt_fusedfused fp16 matrix multiplies and convolutions for DINOv2 backbones and DPT depth heads
Apple (MPS)
fp32, fp16

Limitations

  • Image encoder only: the text encoder in the checkpoint is not loaded (the features task has no text input), although the file is downloaded whole.
  • last_hidden_state holds 2 prefix tokens, then the patches: upstream's first cls token (pooler_output, the global image embedding aligned with text) and its second cls token (index 1, the spatially aware global embedding).
  • kernels replays a CUDA graph per small input shape without hidden states (up to about 520k activations: ViT-B ~360 px, ViT-L ~310 px), captured on the shape's second call, at most 4 per model; `model.release_graphs()` frees them.
Full reference, generated from the repository

Google DeepMind's TIPSv2 image encoders (B, L): image features aligned with text, with patch features meant for dense tasks. Apache-2.0.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
tipsv2-b14synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
7.03
7.21
2.93
1.49
2.18
1.33
373 / 386
529 / 548
204 / 224
206 / 226
239 / 266
398 / 422
tipsv2-l14synthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
21.9
14.2
6.15
4.58
6.04
4.28
1207 / 1226
1776 / 1786
621 / 624
624 / 648
659 / 666
1236 / 1264
Details
  • Original code: https://github.com/google-deepmind/tips at bf10a73
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0
  • Licence evidence: https://huggingface.co/google/tipsv2-b14/blob/ed1e4dc6b74bf3935ae099e9d5eb30fa96528454/README.md (license: apache-2.0); code: https://github.com/google-deepmind/tips/blob/bf10a73765b048edfbfcdf1f7e6ca8292542baae/LICENSE
  • Training data: image-text pairs; the dataset is not named in the repository README or on the model cards at the pinned revisions
  • Options: fp16_accumulate = False (precision only)
  • Kernel libraries: fastbb, dpt_fused
  • Variants: tipsv2-b14, tipsv2-l14
  • Outputs: last_hidden_state, pooler_output, hidden_states
  • Input size: any H x W >= 1, resized or padded inside the model
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): tipsv2-b14 3000x4000 (1.4 GiB peak allocated, 2.2 reserved); tipsv2-l14 3000x4000 (2.2 GiB peak allocated, 2.2 reserved)
  • tipsv2-b14 224x224, first call: portable + torch.compile 9 s (cold Inductor cache), TensorRT fp16 42 s (ONNX + engine).
  • tipsv2-l14 224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 48 s (ONNX + engine).
  • tipsv2-b14 original (fp32) 224x224, original 224x224: tips image_encoder first cls token
  • tipsv2-l14 original (fp32) 224x224, original 224x224: tips image_encoder first cls token
  • tipsv2-b14 load / first call: original 0.9 s / 0.066 s, portable 0.2 s / 0.182 s, kernels 1.42 s / 0.085 s
  • tipsv2-l14 load / first call: original 3.1 s / 0.019 s, portable 2.91 s / 0.083 s, kernels 1.44 s / 0.108 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Two disjoint seeded subsets of the val folders; each image shorter side 256 (bicubic), centre crop 224, the CLS / pooled feature; cosine kNN, vote weighted by similarity. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portablekernelsnote
tipsv2-b14ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)59.3459.3059.3659.16
tipsv2-l14ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)63.4863.4463.4263.42

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 8b2f2df5, 2026-11-10.

Parity

CPU, float32: max |diff| <= 6.9e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 1.9 % and p99 <= 6.9 % of the reference's median |value| (absolute error of the output last_hidden_state).

variantkernels error vs original, mean / p99GPU, torchkernels used
tipsv2-b140.01252 / 0.046RTX 5060 Ti, torch 2.14.1+cu130dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120
tipsv2-l140.00988 / 0.03882RTX 5060 Ti, torch 2.14.1+cu130dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120

Limitations

  • Image encoder only: the text encoder in the checkpoint is not loaded (the features task has no text input), although the file is downloaded whole.
  • last_hidden_state holds 2 prefix tokens, then the patches: upstream's first cls token (pooler_output, the global image embedding aligned with text) and its second cls token (index 1, the spatially aware global embedding).
  • kernels replays a CUDA graph per small input shape without hidden states (up to about 520k activations: ViT-B ~360 px, ViT-L ~310 px), captured on the shape's second call, at most 4 per model; model.release_graphs() frees them.

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Input
Output
Chelsea the cat, input
Input · 100%

Input: Chelsea the cat, Stéfan van der Walt (CC0).