
Depth Pro 2 vs Marigold: Metric Depth or a Diffusion Prior
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 186 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 358 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 139 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0737 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 135 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 80 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 117 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 111 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 42 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 210 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Marigold does not predict depth. It generates it, one denoising step at a time, from a latent diffusion model fine-tuned out of Stable Diffusion 2. Depth Pro 2 predicts it, in a single forward pass, from a transformer trained on the task directly. Both produce a depth map that looks right, both are research-licensed, and the difference between them is the difference between a sampler and a regressor — which shows up first as a bill, second as a latency profile, and only third as accuracy.
That ordering is deliberate. The usual comparison puts AbsRel on one side and Boundary F1 on the other and declares a winner, which is not possible here: Marigold's numbers and Depth Pro 2's numbers are drawn from different protocols on different datasets and have never been run against each other by either lab. What is comparable is the cost of getting an answer, the scale convention the answer arrives in, and what each licence permits. Those three questions have real answers.
Two ways to get a number out of pixels
Marigold's model card is explicit about its own type: a generative latent diffusion-based affine-invariant monocular depth estimator, fine-tuned from stable-diffusion-2. The v1.1 checkpoint, prs-eth/marigold-depth-v1-1, was created on 12 December 2024 and carries 34 likes and about 6,600 downloads. It is designed for the DDIM scheduler and between 1 and 50 denoising steps, and it inherits the base diffusion model's effective resolution of roughly 768 pixels, so larger inputs should be resized with the longer side at 768 before feeding them in.
Depth Pro 2 is a discriminative model. The card describes a multi-view transformer over the input frames and a DPT decoder that predicts an inverse depth map per frame, assembling tokens from several blocks into feature maps at one thirty-second to one quarter of the input resolution. One call, one output, at the input's full resolution — the card is specific about that, and it is the opposite of a 768-pixel working resolution.
• Inference — Marigold runs 1 to 50 denoising steps per image; Depth Pro 2 runs one forward pass.
• Working resolution — Marigold's effective resolution is about 768 pixels on the long side; Depth Pro 2 predicts at the input's full resolution.
• Ensembling — Marigold produces an uncertainty map only when multiple predictions are ensembled with a size above two, which multiplies the cost again; Depth Pro 2 returns no uncertainty.
• Scale — Marigold's output is affine-invariant, values between 0 and 1 interpolating between near and far planes of the model's own choosing; Depth Pro 2's output is metric, in metres.
• Input — Depth Pro 2 takes a photo, a video file, or a live camera feed; Marigold v1.1 takes a single image.
Marigold is a sampler, and that has a bill
The step count is not a footnote. A diffusion depth model's quality is a function of how many denoising steps it runs and how many samples it ensembles, and both of those are knobs you set. At the low end you get a fast, rough map; at 50 steps with an ensemble you get a better map and a much larger compute bill for the same image. Marigold's card frames this as a feature — 1 to 50 steps is a range, and you pick where on it you want to sit — and it is honest, but it means "Marigold's accuracy" is not a single number. It is a curve, and the curve has a cost axis.
Depth Pro 2 has no equivalent dial. The size is the dial — S, B, L or H+ — and it changes the weights rather than the number of passes. A prediction is one pass regardless. For a notebook, that difference is academic; for a service that processes a video stream at 30 frames per second, it is the whole design.
Which is why the cost conversation around a depth model does not stop at the model. The stage that consumes a depth map — describing the scene, deciding what to do, writing the result — is language-model work, and that is the part where per-call cost is published and comparable. OrcaRouter passes provider list price through at 0% markup across 200-plus hosted models on a single key, with automatic failover across providers. A diffusion sampler in the vision stage with an ensemble multiplier is exactly the kind of cost that makes you want the rest of the pipeline to be predictable.
Depth Pro 2 answers in one pass
The single-pass design comes with its own constraints, and they are documented rather than hidden. Depth is the distance along the camera's viewing axis — the z coordinate, not radial distance. Inverse depth in Python is normalized by focal length: focallength_px / (width * depth). Depth scales with focal length, so a focal length 10 percent too long places everything 10 percent too far away.

The video behaviour is where the architecture pays off. Buffered mode processes clips of 16 frames overlapping by four, keeping depth steady across clip boundaries; unbuffered mode predicts each frame as it arrives from that frame and the 15 before it, with the documented warm-up penalty that the first frames of a stream have fewer predecessors and therefore less steady depth. A video gets one focal length for all its frames, the median over the first window. Memory grows with frame size and window length, not with duration.
None of that has a Marigold counterpart, because Marigold v1.1 is a single-image model. Run it on video and you run it frame by frame, and whatever flicker comes out is the flicker you ship.
Affine-invariant against metric
This is the section that decides most integrations, and it is not about quality. Marigold's output is affine-invariant: the predicted values sit between 0 and 1 and interpolate between the near and far planes the model chose. That is a perfectly good representation for a matte, a compositing depth pass, or anything that gets rescaled to fit a range — and it is the standard convention in the diffusion-depth literature, for the good reason that it is what these models are actually good at.
Depth Pro 2's output is a measurement in metres, with the focal length attached. For a robot deciding whether an obstacle is one metre or three metres away, for an AR overlay that has to sit at a physical distance, for a point cloud in a real coordinate frame, that difference is the whole job. Recovering absolute scale from an affine-invariant map is a separate step involving a known reference in the scene or a sensor, and it is work Depth Pro 2 has already done internally.
Neither is better. They are different products of the same word, and the one you want is decided by what consumes the map.
What Marigold V2 changed in 2026
The name "Marigold" now covers two generations, and the second one is not a small update. Marigold V2, huawei-bayerlab/marigold-v2-0, was created on 27 August 2026 with an initial release on 8 September 2026. It is a diffusion transformer rather than a UNet: a LoRA fine-tune of Qwen/Qwen-Image-Edit-2509, published with a SIGGRAPH Asia 2026 paper in ACM Transactions on Graphics 45(6), arXiv 2609.08084.
The capability surface widened considerably. Marigold V2 ships checkpoints for depth in log, linear and disparity parameterizations, layered variants that predict geometry behind glass, plus surface normals and albedo — the last of which is the kind of output that only makes sense from a generative model that already understands image formation. Inference needs about 17 GB of GPU memory at 1024² and 29 GB at 2048², so a 24 GB card covers the first and the second wants 32 GB.
Its depth output is still affine-invariant — depth up to an unknown scale and shift per image — so the scale argument above survives the generation change intact. What changed is accuracy and resolution: the reported zero-shot results under the Pixel-Perfect Depth protocol, AbsRel (lower is better) and δ1 in percent, are 3.6 / 98.0 on NYUv2, 5.4 / 97.4 on KITTI, 2.8 / 99.2 on ETH3D, 3.7 / 97.9 on ScanNet and 5.2 / 97.1 on DIODE. Surface normals come in at a mean angular error of 16.6 degrees and 61.2 percent within 11.25 degrees on NYUv2, and albedo on Hypersim at PSNR 20.78, SSIM 0.811 and LPIPS 0.195. Those are vendor-reported and unreproduced here.
What each side will let you cite
Marigold V2 publishes a table. Depth Pro 2 publishes a scatter plot with labelled axes and no per-model table. That asymmetry is the most under-reported fact about this pairing, and it cuts against the assumption that the newer model must be the better-documented one.

Depth Pro 2's card reports Boundary F1 and Jitter averaged over Sintel and Spring, claims to outperform prior work operating on a single frame or multiple frames on both, and plots roughly two dozen baselines — but Marigold is not among them. There is no Marigold marker on Apple's chart and no Depth Pro 2 row in Marigold V2's table, so neither lab has placed the two models on one axis. The citation block still reads arXiv:261X.XXXXX — a placeholder — and there is no research page for Depth Pro 2 on machinelearning.apple.com. The technical report that the card says compares the sizes on speed and memory does not exist yet.
So a procurement document can cite Marigold V2's AbsRel and δ1 today, with the caveat that they are vendor numbers on a protocol you did not run. It cannot cite a Depth Pro 2 number at all. That is not a knock on the model; it is a fact about the release, and it will change the day the report lands.
Licences: OpenRAIL, Apache, research-only
Three licences across three checkpoints, and they are not interchangeable.
• Marigold v1.1 — CreativeML OpenRAIL++-M, the Stable Diffusion 2 community licence, which carries use-based restrictions rather than a pure non-commercial bar.
• Marigold V2 — Apache 2.0 on the released weights, with the base Qwen-Image-Edit-2509 keeping its own licence, so your obligations depend on both.
• Depth Pro 2 — the Apple Machine Learning Research Model licence, which the repository defines as permitting use exclusively for research purposes, meaning non-commercial scientific research and academic development.
The practical consequence: Marigold V2 is the only one of these three that is plausibly shippable, subject to whatever the Qwen base model requires. Depth Pro 2 is not, at any size, and its code licence does not change that. Depth Pro 2's code separately acknowledges TIMM and DINOv3 as its foundations.
Which to reach for
Reach for Marigold V2 if you want a depth map as one output among several — normals, albedo, layered see-through depth — and you are working at roughly 1024², and the affine-invariant convention is fine because your consumer rescales anyway. It is the more flexible model, it publishes numbers, and its licence is the only permissive one here.
Reach for Depth Pro 2 if you need metres, if your input is a video or a live feed, or if you need sharp boundaries at full resolution and cannot afford to resample to 768. Start with S on a laptop, as Apple's documentation suggests, and expect to clip sky depth — the README's own advice, since sky pixels return infinite depth or a finite value of several kilometres on the small models.
What you should not do is read the two cards as a leaderboard. Marigold V2's AbsRel and Depth Pro 2's Boundary F1 are not the same experiment, and the only honest statement about their relative quality is that no one has run the comparison: Apple plotted two dozen baselines and Marigold was not one of them, while Marigold V2's paper evaluates against its own predecessors on the Pixel-Perfect Depth protocol.

