Read 2026-11-16 · RGB only on this site

Iris-Z text to RGB-D is not this form

Iris-Z is a Speridlabs image model that denoises an image and its depth in one sequence. The paper is arXiv:2611.16010, published 2026-11-11. Nano AI is not affiliated with Speridlabs or Black Forest Labs.

The button opens GPT Image 2.6, text to image, 1K, 12 credits. It returns one ordinary RGB still. It does not run Iris-Z, and it does not return a depth map, a colorized depth image, or a float32 array. Eligible new accounts get 16 credits, so one still leaves 4. It does not call Hugging Face, fal, or the Iris-Z SDK.

Original prompts for this page

These sentences were written here. Copy runs in your browser. It does not call Mera Prompt, ChatGPT, or a model.

  • Slot of stone

    New still

    A narrow sandstone slot, warm light on the near wall and a darker bend farther in. No people, no text, no color legend.

  • Ledge and headland

    New still

    A pale rock ledge in the foreground and a far headland across blue water. Clear day. No people, no text, no false-color overlay.

  • Empty road

    New still

    A straight empty road through low hills, near gravel and a far ridge. Morning light. No cars, no people, no text.

  • Dusk tower

    New still

    A white tower on a dark cliff at dusk, water below, one warm window. No people, no text, no logos.

Copy does not spend credits. Opening a sentence fills GPT Image 2.6 for a new 1K image at 12 credits. It does not upload a photo.

Made by
Speridlabs
Released
November 11, 2026
Run on Nano AI
GPT Image 2.6 · 12 credits per image
Page updated
November 16, 2026

Source: huggingface.co

Photographs, not depth inference

King of Hearts, Lower Antelope Canyon, CC BY-SA 4.0, 2018-11-20. Dietmar Rabich, a coastal cliff and a far headland, CC BY-SA 4.0, 2019-10-24. Reviewed 2026-11-16. Not Iris-Z output and not a colorized depth preview.

Wikimedia Commons: Lower Antelope Canyon, photographed by King of Hearts, CC BY-SA 4.0, 2018-11-20. No person and no logo. A photograph, not a depth map, not an Iris-Z inference, and not a GPT Image 2.6 output.
Wikimedia Commons · From the official showcaseLower Antelope Canyon, photographed by King of Hearts, CC BY-SA 4.0, 2018-11-20. No person and no logo. A photograph, not a depth map, not an Iris-Z inference, and not a GPT Image 2.6 output.Source
Wikimedia Commons: A coastal cliff with a far headland, photographed by Dietmar Rabich, CC BY-SA 4.0, 2019-10-24. No person and no logo. A photograph, not a depth map, not an Iris-Z inference, and not a GPT Image 2.6 output.
Wikimedia Commons · From the official showcaseA coastal cliff with a far headland, photographed by Dietmar Rabich, CC BY-SA 4.0, 2019-10-24. No person and no logo. A photograph, not a depth map, not an Iris-Z inference, and not a GPT Image 2.6 output.Source

What the authors published about Iris-Z

Read on 2026-11-16. The model card is https://huggingface.co/speridlabs/iris-z. The SDK README is https://github.com/speridlabs/iris-z-sdk. The project page is https://speridlabs.com/research/iris-z. The paper, https://arxiv.org/abs/2611.16010, is titled Iris-Z: Scaling One Generative Model Past Specialist Geometry Models, by Alejandro Cañada and Chema Garabito. The arXiv abstract page links http://creativecommons.org/licenses/by/4.0/ and shows the by-4.0 license icon. The paper was published 2026-11-11. The Hugging Face repo was created 2026-11-04 and last modified 2026-11-13. The card lists two base models, black-forest-labs/FLUX.2-dev and black-forest-labs/FLUX.2-klein-base-9B. The README says Iris-Z-big is 32.2B parameters with guidance 3.5, and Iris-Z-small is 9.1B with guidance 4.0. Both list resolution 512 and 1024. The paper says Iris-Z-big starts from FLUX.2 [dev] and Iris-Z-small from FLUX.2 [klein] Base 9B.

Depth is relative, not metric. The paper's limitations say the prediction is defined up to a scale and shift in log space, so Iris-Z gives neither metric distances nor the camera. The method stores depth as per-image normalised log ray distance. The README writes the npy file as relative log depth, and the colorized png as near = red, far = blue. Training frames were converted to metric ray distance, and the released prediction is still the normalised relative value. This page does not copy the benchmark tables. The paper reports its own runs. A hosted Inference Provider is not listed on the model card, which says the model is not deployed by any Inference Provider. The README links a Space at https://huggingface.co/spaces/speridlabs/iris-z. That page, opened 2026-11-16, showed a loading shell. No prompt was submitted, so it is not described as a checked demo.

License, as read that day. The GitHub repository license is Apache-2.0, and the README says the code is under the Apache License 2.0. The model card metadata says license other, license_name flux-non-commercial-license, and points license_link at the FLUX.2-dev LICENSE.txt. That txt file returned 401, so it was not read. LICENSE.md on the FLUX.2-dev repo was read. It is titled FLUX Non-Commercial License v2.1 and says the weights are made available for non-commercial and non-production use. This page does not treat the unread txt file as the same text. The install note says the FLUX.2 autoencoder is gated. Copy on this page stays in the browser. The open button fills GPT Image 2.6, text to image, 1K, 12 credits. It does not upload a photo and does not return depth. /gpt-image-2-6, /nano-banana-2-1, and /flux-2 are other stills on this site. None of them is Iris-Z. /iris-z and /text-to-rgbd stay 404. There is no alias.

Card, paper, and this button

Paper
arXiv:2611.16010, published 2026-11-11. Alejandro Cañada and Chema Garabito. Abs page links CC BY 4.0
Weights
Iris-Z-big 32.2B, guidance 3.5. Iris-Z-small 9.1B, guidance 4.0. Resolution 512 and 1024 on the card
Depth
Relative log depth. Not metric distances and not a camera, as the paper's limitations state
Code license
GitHub license Apache-2.0. README: the code is under the Apache License 2.0
Weight license
Card: flux-non-commercial-license. LICENSE.md read: FLUX Non-Commercial License v2.1. LICENSE.txt was 401
Hosted API
The card says it is not deployed by any Inference Provider. The Space page was a loading shell. No prompt was sent
This button
GPT Image 2.6, 1K, 12 credits, one RGB still. 16 credits at sign-up leave 4. No depth file
Still 404
/iris-z, /text-to-rgbd. No alias

What you can and cannot run here

01

One RGB still

GPT Image 2.6, text to image, 1K, 12 credits. One file. No depth alongside it.

02

Relative, not metric

The paper says Iris-Z does not give metric distances or the camera. This form does not give depth at all.

03

16 credits buy one still

A new account has 16 credits. One 12-credit still leaves 4. That is not an Iris-Z credit.

04

No fake depth

The pictures under the form are photographs. They are not colorized inference and not a float32 array.

Iris-Z and the still this site sells

The Iris-Z column is what the paper, the model card, and the README say. The other column is the button on this page. They are not the same model.

Iris-Z, as publishedThis button
TaskPrompt to image and depth, image to depth, or depth to imageText to one RGB still
DepthRelative log depth, not metric, no cameraNo depth map, no colorized depth, no float32 array
Sizes named512 and 1024 on the card1K on the GPT Image 2.6 request
Price hereNot sold on this site12 credits. 16 at sign-up leaves 4

Open one RGB still here

  1. 1

    Copy a line from this page

    The four sentences were written here. Copy stays in the browser and does not call the Iris-Z SDK.

  2. 2

    Open GPT Image 2.6

    Text to image, 1K, 12 credits. The button fills this site's generator. It does not open the Hugging Face Space.

  3. 3

    Expect one RGB file

    There is no second download for depth. A new account has 16 credits, so one still leaves 4.

Prompts for the GPT Image 2.6 still

Original lines for GPT Image 2.6, text to image, 1K, 12 credits. They ask for an ordinary photograph. They do not ask for a depth map.

  • Slot of stone

    A narrow sandstone slot, warm light on the near wall and a darker bend farther in. No people, no text, no color legend.

  • Ledge and headland

    A pale rock ledge in the foreground and a far headland across blue water. Clear day. No people, no text, no false-color overlay.

  • Empty road

    A straight empty road through low hills, near gravel and a far ridge. Morning light. No cars, no people, no text.

  • Dusk tower

    A white tower on a dark cliff at dusk, water below, one warm window. No people, no text, no logos.

FAQ

Iris-Z questions

Sources read 2026-11-16: the Hugging Face card, the SDK README, the project page, the arXiv abstract, and FLUX.2-dev LICENSE.md.

Does this button run Iris-Z?

No. It opens GPT Image 2.6 for one 1K RGB still at 12 credits. It does not call the Iris-Z weights, the SDK, or a depth file.

Is the depth metric?

The paper says no. Depth is relative, defined up to a scale and shift in log space, and the model does not give metric distances or the camera. This site does not output that depth either.

Who released it?

Alejandro Cañada and Chema Garabito, Speridlabs. The paper arXiv:2611.16010 was published 2026-11-11. The weight repo was created 2026-11-04.

What license applies?

The SDK repository is Apache-2.0 for the code. The weight card says flux-non-commercial-license. LICENSE.md, which was readable, is the FLUX Non-Commercial License v2.1 for non-commercial and non-production use. LICENSE.txt returned 401 and was not read.

Is there an official hosted API on the card?

The card says the model is not deployed by any Inference Provider. A Space URL is linked from the README. The Space page showed a loading shell, and no prompt was sent.

What does a new account buy here?

One 1K still is 12 credits. Eligible new accounts have 16 credits, so one still leaves 4. No depth is included.

Which spellings stay 404?

/iris-z and /text-to-rgbd. /iris-z-text-to-rgb-d is the only host. There is no alias.

Related pages

Nano AI is not affiliated with Speridlabs, Black Forest Labs, or Hugging Face. The button runs GPT Image 2.6 at 12 credits for one 1K RGB still and returns no depth. The photographs are Wikimedia files reviewed on 2026-11-16.