Depth Anything V4: what exists and what to run today

No official Depth Anything V4 exists as of 2026-10-28. The name surfaced in August 2026 on a preprint written by four researchers with no link to the Depth Anything team. That preprint was withdrawn two days later with the comment major errors in research. The official version line ends at Depth Anything 3. There is no v4 repository, no v4 weights, and no host serving a v4 anywhere. What does exist: a published version history, and a served model on this platform for depth work today, da3-small. The search demand for a fourth version is measurable too. This page is the watch. It holds the withdrawal record, the version line, the release signals worth checking, and a first call you can make now.

Last verified: 2026-10-28

What is Depth Anything V4

The short answer: nothing, officially. The Depth Anything family is built by a ByteDance Seed team led by Bingyi Kang. Their published line runs V1, V2, then DA3. Their official repository mentions no fourth generation anywhere in its news, read 2026-10-28. Their Hub organization lists 48 models, and not one id contains V4. What borrowed the name was a third-party preprint, and its story is short.

The withdrawal record

The arXiv record tells the whole story in its own fields. The paper was titled Depth Anything V4: Dynamic 4D Scene Reconstruction via Riemannian Flow Matching on 4D Gaussian Splatting. It was submitted on August 18, 2026 by Jiaming Fan, Jian Lu, Jinling Jia, and Chenbin Zhang. None of those four names appears in the author list of any official Depth Anything paper. Two days later, on August 20, 2026, the first author withdrew it. The version history shows the withdrawal stub, one kilobyte, replacing the full two-megabyte paper. The comments field reads major errors in research. The license field reads no license for this version due to withdrawn.

Before withdrawal, the abstract made three claims. It claimed a framework that reconstructs dynamic 4D scenes from ordinary monocular video. It claimed a flow matching method that treats rotation, scale, and opacity parameters on their natural curved spaces instead of flat Euclidean ones. And it claimed a small accuracy edge, 0.806 versus 0.762 on its internal F-score comparison, alongside a pre-training cost of 360 GPU-hours. These were claims of a paper that no longer stands. Treat every number in that abstract as retracted with the paper itself.

The spread explains why the name stuck. A paper-summary account posted the preprint on August 20, 2026. Two days later it posted a withdrawal alert: the central claim was found to be false, and the issues included basic inconsistencies inside the paper itself. A reader on that thread asked the obvious question the same week. Did these authors have any relation to the v1 through v3 teams. They did not. The code repository the summary page linked is now deleted. That was checked two ways on 2026-10-28. The repository address returns not-found from two separate clients. The personal account that owned it still returns a live page. The account survives. The code does not.

The real version line

The family's actual history is short and fully sourced:

• V1: Depth Anything, submitted January 19, 2024, accepted at CVPR 2024. Practical monocular depth from large-scale unlabeled data.
• V2: Depth Anything V2, submitted June 13, 2024, accepted at NeurIPS 2024. Synthetic training data and a larger teacher.
• DA3: Depth Anything 3, submitted November 13, 2025, an ICLR 2026 Oral. Geometry from any number of views, with or without known poses.

The gaps run five months, then seventeen months, then eleven months and counting. A cadence that irregular supports no release-date prediction, and this page offers none. What the record does support is a shipping habit. The team releases paper, code, and weights together, as they did on November 14, 2025 for DA3. A genuine v4 would arrive the same way, not as a lone preprint from unknown authors.

Why depth anything v4 github finds nothing

Try the search and the engines finish your typing. On 2026-10-28, Google's own suggestion box for depth anything v4 returned five learned completions. Github sat at position two. Withdrawn stood as its own suggestion. People keep typing the phrase. The artifact they hope for keeps not existing:

• A GitHub repository search for Depth-Anything-4, run 2026-10-28, returned fourteen fuzzy matches and zero actual v4 repositories. The nearest thing is a zero-star fork of Depth Anything 3 renamed Depth-Anything-4D. It is the streaming variant of DA3 under a new name, not a fourth generation.
• An issue search inside the official repository for v4 returned zero results. Not one person has opened a when-is-v4 issue there.
• A Hub model search for v4-shaped ids returned nothing real. The five fuzzy hits were V2 variants and camera-number ids.

The demand is real and dated. The supply is zero and dated. That gap is this page's reason to exist.

Use cases

What would a fourth generation even do. The withdrawn paper pointed at one target: dynamic 4D reconstruction, meaning geometry that moves, recovered from ordinary single-camera video. That target is real and unsolved in general, which is why third parties keep reaching for it. But the withdrawn paper is evidence of appetite, not of capability.

The family that exists today serves a different, working set of jobs. Per the official repo and the small tier's model card, DA3 reads one image or many, with or without known camera poses. It returns a depth map per view, a confidence map, and the camera parameters. Exports include GLB and PLY meshes for 3D pipelines. A streaming variant, released December 11, 2025, handles ultra-long video under 12 GB of GPU memory with a sliding window.

Where that lands in practice: scene capture for 3D content work, where photos become meshes. Depth also drives ComfyUI and Blender scene flows for rendering and compositing. And robot builds use depth plus camera pose from cheap cameras to feed navigation and manipulation. Community threads keep comparing versions on real footage. One running argument asks whether v2-large beats v3-giant on certain scenes. That argument is the demand side of this page in miniature. People care what the next version changes.

What would a genuine v4 have to change. The official record exposes three open items. One is the license split. The small and base tiers ship Apache 2.0. The large and giant tiers ship CC BY-NC 4.0. One is the streaming variant, which ships beside DA3 rather than inside it. And one is raw accuracy on the any-views task, where the project page already claims clear wins over the prior state of the art. Anything beyond those three items is speculation. This page does not do speculation.

API usage

There is no v4 call to document, on this platform or any other. The row that exists here is da3-small, on the public list since 2026-10-06, metered per image. The three tabs below are that call: check the catalog, make the request, read the raw contract. The same catalog check doubles as the watch tool. When a v4 id ever appears in it, the listing state will say so on the day you ask. No page's date sits in between.

curl

The catalog check, one request:

curl -sS "https://api.ironstratum.com/v1/models" \
  -H "Authorization: Bearer $KEY"

The answer lists every model id this platform serves, on the day you ask. As of 2026-10-28, da3-small is in it and no v4 id is. Repeat the call later and you learn whether that changed. No page's date sits in between, including this one's.

The depth call itself, against the served id:

curl -X POST "https://api.ironstratum.com/v1/vision" \
  -H "Authorization: Bearer $KEY" \
  -F "model=da3-small" \
  -F "file=@scene.jpg"

Metering is per image, from the count of images in the request. Rates for listed models sit on the pricing page, never in prose on pages like this one.

python

The same call from python, with the key and the host read from the environment:

import os
import requests

resp = requests.post(
    os.environ["API_BASE"] + "/v1/vision",
    headers={"Authorization": "Bearer " + os.environ["KEY"]},
    files={"file": open("scene.jpg", "rb")},
    data={"model": "da3-small"},
)
print(resp.status_code)
print(resp.json())   # depth result, confidence, usage for the metered call

No v4 id exists to swap into the model field. If you self-host instead, the publisher's repo documents the python class for every tier. The small tier needs no GPU at all. Your license choice and your hosting choice stay separate decisions.

http

The raw request the vision route serves under:

POST /v1/vision HTTP/1.1
Host: api.ironstratum.com
Authorization: Bearer $KEY
Content-Type: multipart/form-data; boundary=frame

--frame
Content-Disposition: form-data; name="model"

da3-small
--frame
Content-Disposition: form-data; name="file"; filename="scene.jpg"
Content-Type: image/jpeg

<the image bytes>
--frame--

Every field and limit lives in the vision route guide. A request naming an id that is not on the public list returns a named not-found error. That is the honest answer, and it is what the platform's own docs promise.

Benchmarks

There is no v4 benchmark to report, anywhere, because there is no v4. The only numbers that ever carried the name are the withdrawn paper's internal ones, the 0.806 versus 0.762 F-score comparison. They stand retracted with the paper. The record that does stand:

• DA3 was accepted as an ICLR 2026 Oral, per the OpenReview record, published January 26, 2026.
• The official repository recommends its retrained 1.1 checkpoints over the original large and giant releases, per its own model table read 2026-10-28.
• The official project page claims roughly 35.7 percent better camera pose accuracy than the prior state of the art. It claims roughly 23.6 percent better geometric accuracy. Both claims are the page's own, on its own benchmarks.
• The small tier carries 234,136 downloads in the last month on the Hub, per its model card read 2026-10-28.

When a fourth version appears with real measurements, this section is where they land, dated and sourced. Until then, every v4 benchmark claim you meet traces back to one withdrawn preprint. The honest answer to all of them is the record above.

Getting started

Depth work does not need to wait on a version number. The path from this page to a first metered call:

1. Sign up on the console and create an API key. The wallet is prepaid, and its balance is the hard stop on spend.
2. Make the catalog check from the curl tab. Confirm da3-small is listed on the day you start.
3. Send one image through the vision call. That single request is your first metered call, billed per image.
4. Read the depth result fields, then scale from one image to your real batch.

The served sibling has its own model page. The vision lane page holds the whole machine vision lane. The pricing page carries the current rate card. The models index is the live catalog.

If you are watching for the fourth version, watch the three signals from the FAQ. Read the official repo's news section. Check the team's author names on any new arXiv record. Look for new ids in the official Hub organization. A learned autocomplete suggestion is none of those. This page carries its Last verified date at the top. The record above is that date's. The catalog check answers the listing question live, any day, without trusting any page at all.