Best Depth Estimation Models for CPU-Only PCs in 2026: Sapiens2 Normal 0.4B

Quick Answer

facebook’s Sapiens2 Normal 0.4B is the top pick as of September 2026 under “newest release first, then downloads,” winning the download tiebreak against the equally recent Sapiens2 Normal 1B; no public benchmark yet establishes a CPU performance winner [9] [7]. In order, the ranking is Sapiens2 Normal 0.4B (facebook), Sapiens2 Normal 1B (facebook), TIPSv2 B14 DPT (google), Depth Pro (apple), and DepthCrafter (tencent) [9][7].

Key Takeaways

• facebook’s Sapiens2 Normal 0.4B takes first place because it shares the newest release date, 2026-04-23, with facebook’s Sapiens2 Normal 1B and has more downloads: 2,483 versus 746—not a demonstrated CPU performance advantage. [9][7]
• Sapiens2 Normal 0.4B has 453M parameters and estimated 4-bit weights of 226.5 MB (params x 0.5 bytes); Sapiens2 Normal 1B has 1.5B parameters and an equivalent estimate of 750 MB. Both use sapiens2-license; weight estimates do not establish total RAM requirements. [9][7]
• google’s TIPSv2 B14 DPT has 158M parameters, estimated 4-bit weights of 79 MB (params x 0.5 bytes), and an Apache-2.0 license. Its publication date is 2026-04-09. [3]
• apple’s Depth Pro is an established pick with 952M parameters, estimated 4-bit weights of 476 MB (params x 0.5 bytes), and the apple-amlr license. Its original F16 weights occupy 1.9 GB. [1]
• tencent’s DepthCrafter is an established pick with 3.0 GB of full-precision weights. Its license field reads “license”; a parameter count is unavailable for calculating a comparable quantized-weight estimate. [5]
• Ranking note: newest release first, then downloads; size does not determine placement, and every candidate has “no public benchmark yet.” [9][7][3][1][5] Established picks Depth Pro and DepthCrafter bypass the recency gate and follow the same ordering rule. [1][5] Scope admits only depth estimation models from labs with a published paper or leaderboard record, using the Hugging Face depth-estimation pipeline tag. [1][3][5][7][9]

How do these depth estimation models compare on parameters, memory, licenses and published benchmarks?

Which depth estimation models should you consider for a CPU-only PC?

1. Sapiens2 Normal 0.4B

Sapiens2 Normal 0.4B by facebook ranks first because its April 23, 2026 release ties facebook’s Sapiens2 Normal 1B, while its 2,483 monthly downloads exceed that model’s 746.[9][7] The ordering basis is newest release first, then downloads; download counts serve only as a tiebreak.[9][7] The ranking admits only depth estimation models from labs with a published paper or leaderboard record; facebook documents this family in the Sapiens2 paper.[8]

The model has 453 million parameters, with a 4-bit weight footprint of 226.5 MB, estimated (params x 0.5 bytes).[9] The published F32 weights occupy 3.6 GB, and the license is sapiens2-license.[9] The calculated footprint describes parameter storage alone. Treat it as a planning estimate, not a measured RAM requirement or confirmation that a compatible quantized implementation exists.

For a CPU-only PC, the hardware requirement remains unresolved: the listed specifications do not establish total inference memory, supported CPU execution paths or latency. A practical use is exploratory local evaluation, with deployment contingent on checking runtime compatibility, license terms and actual memory consumption. The central caveat is no public benchmark yet: the ranking position reflects release timing and the download tiebreak, so it does not establish depth accuracy or CPU performance.

2. Sapiens2 Normal 1B

Sapiens2 Normal 1B by facebook ranks second under the ordering rule: newest release first, then downloads.[7][9] Its Hugging Face release date matches facebook’s Sapiens2 Normal 0.4B at April 23, 2026, but its 746 downloads trail that model’s 2,483 over the reported 30-day period.[7][9] The benchmark status is “no public benchmark yet,” so the position reflects release timing and the download tiebreak.

The model lists 1.5B parameters and 12.3 GB of F32 weights under the sapiens2-license.[7] Weight-only memory at 4-bit would be 750 MB, estimated (params x 0.5 bytes) from the cited 1.5B parameters.[7] That calculation does not establish a working quantized implementation or total RAM consumption. CPU compatibility, runtime memory requirements and inference latency remain unspecified, so a concrete PC configuration cannot be recommended from those figures.

Treat the model as a candidate for local evaluation if you can check CPU execution, measure memory consumption and review the license before adoption. The accompanying paper is titled “Sapiens2.”[8] The practical caveat is that the available specifications do not establish suitability for routine CPU-only depth estimation; the weight estimate alone cannot justify a hardware purchase.

3. TIPSv2 B14 DPT

TIPSv2 B14 DPT by google ranks third in this selection [3][7][9]. The ordering is newest release first, then downloads: its Hugging Face publication date of April 9, 2026 [3] places it behind the Sapiens2 releases from facebook, published April 23, 2026 [7][9], and ahead of the established picks from apple and tencent [1][5]. Its benchmark status is “no public benchmark yet”; the position does not establish superior depth accuracy or CPU speed.

The model has 158M parameters [3], giving a theoretical 4-bit weight payload of 79 MB, estimated (params x 0.5 bytes) [3]. The published checkpoint instead contains F32 weights totaling 0.6 GB [3]. Those figures describe weights, not total system RAM requirements. A CPU-only deployment also needs memory for execution, but no supported CPU configuration, runtime memory measurement, or CPU latency is documented here. Treat the estimate as a planning figure, not a demonstrated hardware requirement.

The practical use case is evaluating local depth estimation for a project whose licensing requirements align with Apache-2.0, the listed license [3]. The accompanying paper is google’s “TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment” [4]. The deployment caveat is unverified execution: neither a working quantized CPU path nor input-resolution or sequence-length limits are established, so confirm runtime compatibility and memory use before committing to a PC configuration.

4. Depth Pro

Depth Pro from apple ranks fourth as an established pick, with a Hugging Face release date of November 27, 2024.[1] Its position follows the ordering of newest release first, then downloads; the newer eligible releases therefore precede it.[3][7][9] The established-pick exemption keeps Depth Pro eligible despite its age.[1] Benchmark status: no public benchmark yet. Its position should not be read as a measured comparison of CPU performance.

The apple/DepthPro-hf model has 952M parameters and F16 weights totaling 1.9 GB.[1] Weight storage at four bits would be approximately 476 MB, estimated (params x 0.5 bytes) from that parameter count.[1] That calculation covers weights alone, not total runtime RAM, and does not establish that a compatible quantized implementation is available. For local hardware planning, an exact CPU specification or RAM requirement remains unverified, as does inference latency on a CPU-only PC.

Depth Pro is a candidate for monocular metric depth estimation, the task described in apple’s paper, “Depth Pro: Sharp Monocular Metric Depth in Less Than a Second.”[2] The paper’s title should not be treated as a CPU latency guarantee. The practical caveat is deployment uncertainty: validate the intended runtime and memory use before committing to a workflow. Distribution uses the apple-amlr license; check its terms against the intended application.[1]

5. DepthCrafter

DepthCrafter by tencent ranks fifth as an established pick, with a Hugging Face publication date of September 14, 2024.[5] The ordering uses newest release first, then downloads; its established-pick status permits inclusion outside the recency window.[5] DepthCrafter has no public benchmark yet in this comparison, so its position does not establish relative depth accuracy or CPU performance. Its intended use is generating consistent depth sequences for open-world videos.[6]

The repository lists full-precision weights at 3.0 GB.[5] Weight-file size alone does not establish the system RAM needed for local inference. A parameter count is not provided in the listed specifications, so a defensible 4-bit weight-memory estimate cannot be calculated. CPU latency, peak RAM and a supported quantized CPU execution path remain unverified; there is consequently no substantiated minimum CPU or RAM configuration to recommend.

Consider DepthCrafter when consistency across a video’s depth sequence matters to your workflow.[6] For a CPU-only PC, treat it as a candidate requiring runtime and memory validation before committing to a deployment. A separate caveat is licensing: the repository’s license field is simply “license,” which does not establish specific reuse permissions.[5] Check the actual terms against your intended use.

How can you estimate depth model weight memory from parameter counts?

Estimate depth model weight memory by multiplying the parameter count by the assumed storage per parameter: Google’s google/tipsv2-b14-dpt has 158M parameters, giving 79 MB at 4-bit, estimated (params x 0.5 bytes).[3] Treat that result as a weight-storage calculation, not a measured RAM requirement.

Apply the same calculation consistently when comparing candidates. Facebook’s facebook/sapiens2-normal-0.4b has 453M parameters, giving 226.5 MB at 4-bit, estimated (params x 0.5 bytes).[9] Apple’s Depth Pro (apple/DepthPro-hf) has 952M parameters, giving 476 MB at 4-bit, estimated (params x 0.5 bytes).[1] Those estimates describe hypothetical quantized weights; they do not establish that compatible quantized checkpoints are available.

Keep published weight sizes separate from calculated estimates. Google lists F32 weights of 0.6 GB for google/tipsv2-b14-dpt, while Apple lists F16 weights of 1.9 GB for Depth Pro.[3][1] Label the precision beside each published size so readers can distinguish the downloadable weights from a proposed quantized representation.

Avoid working backward from a repository’s weight size when the parameter count is unavailable. Tencent’s DepthCrafter lists full-precision weights of 3.0 GB; leave its parameter-based estimate unspecified.[5] For CPU deployment, use the calculation as a starting point, then verify actual RAM use and runtime with the intended inference software.

What do the model licenses allow?

Usage rights depend on each model’s license; the listed identifiers alone do not establish permission for commercial deployment, modification or redistribution across the lineup.[1][3][5][7][9]

Google’s google/tipsv2-b14-dpt lists Apache-2.0.[3] Check the linked license text before distributing weights or incorporating the model into a product. Keep permission to use the model separate from any obligations that apply when you redistribute it.

Apple’s apple/DepthPro-hf lists apple-amlr.[1] Review that agreement for your intended use. The identifier alone does not establish whether a particular commercial application, modified model or redistribution arrangement is allowed.

Facebook’s facebook/sapiens2-normal-0.4b and facebook/sapiens2-normal-1b both list sapiens2-license.[9][7] Evaluate that agreement for either checkpoint. A shared license label should direct you to the applicable terms, rather than serve as a substitute for reading them.

Tencent’s DepthCrafter (tencent) lists only license, which does not identify a recognizable set of permissions.[5] Open the repository’s license document before deciding whether deployment or redistribution is permitted.

For a local application, record the applicable license text alongside the downloaded checkpoint. Confirm that your intended use is covered, then check any conditions concerning modification, redistribution, attribution and accompanying notices.

Are published benchmarks available to compare CPU performance?

No published benchmark scores are available to compare CPU performance across the models in this ranking. Their status is “no public benchmark yet”; the ordering therefore does not establish which model runs faster on a CPU.

Facebook’s Sapiens2 variants, facebook/sapiens2-normal-0.4b [9] and facebook/sapiens2-normal-1b [7], and Google’s TIPSv2 model, google/tipsv2-b14-dpt [3], have no public benchmark yet. Apple’s Depth Pro, apple/DepthPro-hf [1], and Tencent’s DepthCrafter, DepthCrafter (tencent) [5], share that status. Published papers accompany each family, but paper publication alone does not establish comparable CPU performance. [2][4][6][8]

Parameter counts and weight-file sizes answer different questions from inference timing. Google’s TIPSv2 model has 158M parameters [3], while Apple’s Depth Pro has 952M parameters [1]. Those counts alone cannot establish CPU latency, peak runtime memory or depth quality. Estimated storage for quantized weights would likewise leave execution speed and runtime overhead unresolved.

A useful CPU comparison would report the processor, thread count, inference backend, numerical precision, input resolution, peak memory and elapsed time under matched conditions. Depth quality should accompany timing so that speed gains can be assessed alongside output accuracy. Until comparable measurements are available, treat the ranking as a release-order shortlist, with downloads breaking release-date ties, rather than a CPU performance recommendation.

Frequently Asked Questions

Which model heads the CPU-only depth estimation ranking?

Facebook’s Sapiens2 normal model (facebook/sapiens2-normal-0.4b) leads because it shares the 2026-04-23 release date with Facebook’s facebook/sapiens2-normal-1b but has 2,483 monthly downloads against 746 [9][7]. The latter ranks next [7][9]. Every candidate has no public benchmark yet [1][3][5][7][9]. The ordering therefore reflects release dates and a download tiebreak; CPU speed and depth accuracy remain unverified.

How is the ranking ordered, and which models qualify?

Ranking note: newest release first, then downloads. Following the Facebook entries are Google’s TIPSv2 (google/tipsv2-b14-dpt) [3][4], Apple’s Depth Pro (apple/DepthPro-hf), an established pick [1][2], and Tencent’s DepthCrafter (DepthCrafter (tencent)), an established pick [5][6]. Established picks bypass the recency gate [1][5]. Scope: this ranking admits only depth estimation models from labs with a published paper or leaderboard record (Hugging Face pipeline tags: depth-estimation).

How much memory might quantized weights require?

Google’s model has 158M parameters: 79 MB estimated (params x 0.5 bytes) for quantized weights [3]. Facebook’s smaller entry has 453M parameters: 226.5 MB estimated (params x 0.5 bytes) [9]. Apple’s model has 952M parameters: 476 MB estimated (params x 0.5 bytes) [1]. Treat these as weight-storage calculations. Actual RAM requirements and compatible CPU quantization implementations need separate verification.

Which licenses should I check before deploying locally?

Google’s TIPSv2 lists Apache-2.0 [3]. Apple’s Depth Pro lists apple-amlr [1], while the Facebook entries list sapiens2-license [7][9]. Tencent’s DepthCrafter exposes the generic label “license” [5]. Review the corresponding terms against your intended use before deployment. Local execution alone does not answer questions about redistribution, modification, or commercial permission.

How fast will Depth Pro run on my CPU?

Treat CPU latency as unverified. Apple’s paper is titled “Depth Pro: Sharp Monocular Metric Depth in Less Than a Second,” but that title does not establish performance on your CPU [2]. Depth Pro has 952M parameters and published F16 weights of 1.9 GB [1]. Measure latency and peak RAM with your intended input images before choosing a deployment configuration.

Can DepthCrafter process video, and how long can inputs be?

Tencent’s DepthCrafter addresses video depth sequences in “DepthCrafter: Generating Consistent Long Depth Sequences for Open-world Videos” [6]. Do not interpret “long” as an unlimited input allowance. Verify supported sequence lengths and CPU execution requirements before deployment. Published full-precision weights occupy 3.0 GB [5]. DepthCrafter is an established pick [5]; a quantized-memory estimate requires a supported parameter count.

Sources

1. apple/DepthPro-hf model card (Hugging Face) — 2026-09-25
2. Depth Pro: Sharp Monocular Metric Depth in Less Than a Second — 2024-10-02
3. google/tipsv2-b14-dpt model card (Hugging Face) — 2026-09-25
4. TIPSv2: Advancing Vision-Language Pretraining with Enhanced Patch-Text Alignment — 2026-04-13
5. tencent/DepthCrafter model card (Hugging Face) — 2026-09-25
6. DepthCrafter: Generating Consistent Long Depth Sequences for Open-world Videos — 2024-09-03
7. facebook/sapiens2-normal-1b model card (Hugging Face) — 2026-09-25
8. Sapiens2 — 2026-04-23
9. facebook/sapiens2-normal-0.4b model card (Hugging Face) — 2026-09-25