Collov Labs
Collov Labs is the visual agent platform of Collov, Inc.: multimodal agents that perceive a scene, plan, generate and iterate. Three verticals sit on the same core, serving real estate, mobile capture and consumer design, with published research behind them.
What is Collov Labs?
Collov Labs presents itself as The Visual Agent Platform of Collov, Inc., and states its position bluntly: “Not a Product. Not a Studio. A System.” Its stated mission is “unlocking the latent potential of visual intelligence through autonomous optimization”, and its argument is that frontier models win isolated demos but fail on complicated visual work. The proposed answer is to coordinate perception, reasoning and generation into a single execution system, moving from one-step generation to coordinated task execution — summed up on the site as “Language agents explain. Visual agents execute.”
The architecture is described in three named layers. A Perception Layer performs open-vocabulary detection and segmentation and reads depth, lines and planes, turning an image into a structured, queryable scene state. An Agent Orchestrator interprets intent, reasons under constraints, decomposes tasks and plans multi-step visual actions. A Post-Trained Diffusion Engine then carries out production-quality edits at low latency, tunable per vertical and per deployment constraint. Around them runs a five-step loop — Perceive, Plan, Act, Evaluate, Improve — plus a learning loop in which every deployed agent interaction produces execution traces, corrections and outcomes that feed back into the model, the reasoning policy and the validators.
Three products sit on that core, described as domain-specific interfaces into the same intelligence core: NewEyes for intelligent capture, context understanding, visual creation and price comparison; Collov AI for real estate, with high-fidelity staging and listing image enhancement; and CozyAI, a gentler experience aimed at prosumers. Generation itself is split into human, scene and product tracks, with RL alignment on shadows, lighting and pose consistency, and claimed zero-shot spatial reasoning that respects physics.
The research claim is specific: Meissonic (ICLR 2025), DPaI (ICLR 2025), D-edit (AAAI 2025), FlexControl (ICML 2025), a paper on view conditions for image synthesis (IJCAI 2024), “No More Adam”, and one paper under submission. Code sits on GitHub, models on Hugging Face. Partnerships cover Intel OpenVINO and Core Ultra, Qualcomm Snapdragon on-device inference, an Unreal Engine “Epic Connector” and Cloudflare’s edge network. The flagship WorldFM demo, turning one static 2D image into a real-time 3D environment, is still marked Coming Soon.
What it does
- Detect and segment scene content with an open vocabulary — objects, materials, geometry and spatial relations, “not just labels”.
- Read depth, planes, boundaries, occlusions and real-world constraints, then turn the visual scene into a structured state the agent can query, track and update.
- Interpret an intent, reason over constraints, break the task down and plan multi-step visual actions.
- Execute production-quality image edits at low latency through a post-trained DiT diffusion engine.
- Evaluate the result against the constraints and the original intent, then correct it and run the loop again.
- Generate humans, scenes and products with consistent shadows, lighting, materials and perspective.
- Run on a local NPU for fast, private experiences (Intel OpenVINO and Core Ultra, Qualcomm Snapdragon).
When to use Collov Labs / When not to
A quick filter to help you decide if Collov Labs is the right fit.
When to use Collov Labs
- Computer vision and generative AI researchers, who can follow work published at ICLR 2025, AAAI 2025, ICML 2025 and IJCAI 2024 and run the models released on Hugging Face.
- ML and AI platform engineers evaluating a documented perceive-plan-act-evaluate-improve loop rather than a single-shot image generator.
- Real estate agents and brokers, through the Collov AI vertical: high-fidelity staging and listing image enhancement.
- Interior designers and architects who need room redesigns where materials, geometry, lighting and perspective stay consistent.
- On-device and mobile developers, thanks to models optimized for local NPUs (Intel OpenVINO and Core Ultra, Qualcomm Snapdragon).
When not to use Collov Labs
- Professionals who need licensed deliverables: the terms state plainly that Collov is not an architecture firm, an engineering practice, a licensed interior design studio, a real estate brokerage or a construction consultant, and that all outputs are for visualization only.
- Buyers who want to compare plans and subscribe unaided — no price is published, and the only routes in are a booked call or an early-access form.
- Developers looking to integrate the Labs platform through a documented self-service API: collov.com publishes none.
- Anyone who needs renders that are listing-ready by default: outputs carry no MLS, advertising or regulatory compliance guarantee, and presenting them as real photographs without proper disclosure is forbidden.
- Users under 18, and teams intending to build a competing product on top of the Services — both are excluded by the terms.
How to use Collov Labs
A typical end-to-end flow, from setup to results.
- Open collov.com — a single page whose navigation is made of internal anchors (#product, #research, #ecosystem) rather than sub-pages.
- Read the three-layer architecture and the Perceive-Plan-Act-Evaluate-Improve loop to judge whether the platform matches the workflow you have in mind.
- Review the research section: seven papers with their venues, plus code on GitHub (collovlabs) and models on Hugging Face (Collov-Labs), both open without an account.
- Use the two entry buttons at the top of the page, “Get Started” and “Contact Us” — bearing in mind that the “Get Started” buttons carry no destination in the static HTML, so what happens after the click is not observable.
- For an actual conversation, follow “Contact us” in the footer, which opens a Google Calendar booking page.
- Alternatively, enter your address in the footer “SIGN UP FOR FREE” block and submit “Get Early Access”.
- Explore the verticals from the ecosystem section: “Go to Site” for Collov AI, “Download Now” and “Learn More” for NewEyes, App Store and Google Play badges for CozyAI — all displayed without a clickable link in the static HTML.
- Expect account creation and a credit system on the Collov AI side, where one generated image consumes one credit.
- Wait for the WorldFM demo: uploading a static 2D image to obtain a real-time 3D environment is announced as “Coming Soon” and is not usable yet.
Pros & Cons
Pros
- Verifiable research footing: seven pieces of work cited with their venues, including two at ICLR 2025, one at AAAI 2025, one at ICML 2025 and one at IJCAI 2024.
- Public code and models: repositories on GitHub (collovlabs) and open models on Hugging Face (Collov-Labs).
- Architecture spelled out layer by layer, with a five-step execution loop documented on the home page instead of being left implicit.
- On-device NPU execution demonstrated with two silicon vendors, Intel and Qualcomm — a concrete latency and privacy argument.
- Recent, documented funding: a $23M Series A announced in April 2026, with Axios Pro and Fortune Term Sheet coverage relayed on the site.
- Three verticals already in service — real estate, mobile capture and consumer design — showing the platform is past the prototype stage.
- Users keep ownership of their inputs and outputs under section 6.1 of the terms, and generated environments can be carried into a real-time engine through the Unreal Engine connector.
Cons
- No public pricing whatsoever: no price page, no plan, no amount and no currency on collov.com.
- Single-page site: no detailed product page, no documentation, no about page, and neither robots.txt nor sitemap.xml.
- No API documentation published for the Labs platform.
- The action buttons — “Get Started”, “Download Now”, “App Store”, “Google Play” — carry no destination in the HTML, so where they lead cannot be verified, and the flagship WorldFM demo is announced as “Coming Soon”.
- Production interactions feed model training and no opt-out is documented.
- No published security certification (neither SOC 2 nor ISO 27001), no subprocessor list, no DPA, no GDPR compliance claim, no Article 27 representative and no stated legal basis for processing.
- Legal documents are hosted on the neighboring collov.ai domain and written in the name of Collov.ai; the privacy policy is dated October 21, 2024 and still carries leftover mentions of another vendor (“Styldod”), and no postal address is published anywhere.
Pricing & Plans
No pricing is published. Collov Labs shows no price page, no plan, no amount and no currency on collov.com; the pricing model is therefore a contact-sales one. Access runs through the “Contact Us” link, which opens a Google Calendar booking, or through the footer “Get Early Access” form. The footer wording “SIGN UP FOR FREE” belongs to that early-access form and not to a free product plan, and no free trial is announced for the platform. The Standard, Advanced and Premium tiers named in the terms belong to the Collov AI product and carry no amount either. Prospective buyers should expect a quotation obtained by contacting the company.
Data, GDPR & hosting
A consolidated view of how Collov Labs handles your data.
GDPR overview
The privacy policy, effective October 21, 2024, names www.collov.com and Collov Inc., and devotes section 10 to the GDPR: residents of the European Economic Area “may be entitled to rights under the GDPR”, and eight are listed — withdrawal of consent, access, a copy, rectification, erasure, portability, restriction of processing and objection. Exercising them is free, requests go to cs@collov.com, identity may be verified first, and complaints can be filed with a supervisory authority through the EDPB members page. The announced response window is 45 days, extendable to 90. What is missing matters as much: the document claims compliance with the California Consumer Privacy Act only, never with the GDPR. No Article 27 representative and no DPO are named, no legal basis for processing is stated, and no transfer mechanism (standard contractual clauses, adequacy decision) is mentioned, while section 9 confirms data goes to the United States.
Who owns the data?
Section 6.1 of the terms leaves you the owner of both what you upload (“Input”) and what the system generates (“Output”) — together “Content” — to the extent the law allows, and solely responsible for it. Section 6.2 grants Collov a worldwide, royalty-free, sublicensable, limited license to process, modify, store, reproduce and display your Input for the sole purpose of providing and improving the Services, and section 6.3 permits the reuse of anonymized, aggregated data. The privacy policy adds who else may see it: social networks, analytics providers, payment processors, third-party marketing providers and affiliates, each bound by a confidentiality agreement. In a merger or restructuring, information may be transferred without notice or consent. Collov states it does not rent or sell personal contact information for third-party direct marketing.
Reuse rights
Because ownership of the outputs stays with the user (section 6.1), the outputs can be reused without asking permission, within the limits of the law — but the terms attach conditions. “You are solely responsible for verifying all outputs before commercial use”, with no warranty of accuracy, uniqueness or fitness. Outputs are for visualization only and must not be relied on for construction, structural modifications or professional decision-making, and nothing guarantees they meet MLS, advertising or regulatory standards. Presenting AI-generated content as a real photograph without proper disclosure is prohibited. Watermarking depends on the subscription tier described in the terms — the Standard plan trial watermarks Photo Editing, Design Agent and Chat Editing, and watermarks are removed from the Advanced tier upward — a rule that applies to the Collov AI product tiers, since collov.com itself publishes no plan. Collov also disclaims liability for losses tied to a real estate transaction and for misleading listings.
Data retention & training
Hosting summary
Jurisdiction is the United States. The privacy policy states “we are headquartered in the United States” and that the Services are governed by United States law, then adds — in the conditional — that information “may be transferred to, stored, and processed in the United States where our servers may be located”. No data center, region or hosting provider is named, and the policy warns that the United States “might not offer the same level of privacy protection as the country where you reside”. Storage is described as handled by Collov or by third-party hosting providers, in environments protected against public or unauthorized access, with encryption in transit through SSL “or similar technologies”. Two network facts prove nothing about storage: the domain’s first A record, 216.150.1.1, geolocates to the United States on Amazon’s AS16509 but is flagged anycast — a point of presence, not a storage location — and the site is served by Vercel, which concerns page delivery. No hosting or security certification is published: neither SOC 2 nor ISO 27001, and no subprocessor list.
Things to keep in mind
Risks and trade-offs to weigh before adopting Collov Labs.
- Your work trains the system: execution traces and corrections from production feed the model, and no opt-out is documented, so every upload is a contribution you cannot take back.
- Convincing renders invite deception. The terms forbid passing AI output off as a real photograph without disclosure and explicitly contemplate the risk of a misleading property listing — a temptation that grows with the quality of the output.
- Outputs are not decision-grade: the terms warn they “may be inaccurate, distorted, incomplete, or unrealistic” and are for visualization only, never for structural, engineering or safety decisions, and they carry no MLS, advertising or regulatory compliance guarantee.
- The input license is broad — worldwide, sublicensable, covering processing, modification, storage, reproduction and display — and information may be transferred without notice or consent in a merger or restructuring, so what you upload deserves as much thought as what comes out.
- Data crosses to the United States with an explicit warning that protection there may be weaker than in your country of residence; tracking through Google Analytics, Facebook Pixel and Rakuten is declared, and the site states it does not respond to “Do Not Track” signals.
- Handing spatial and aesthetic judgment to a system that optimizes plausibility rather than fidelity to the actual place erodes the habit of checking a room against its render.
- Commercial opacity makes dependency hard to price: with no published pricing and no public documentation, the cost of relying on the platform only becomes visible after you have engaged.
Setup & Integrations
Technical difficulty
Two very different levels. Evaluating the platform requires no setup and allows none: collov.com is a presentation page with a booking link and an early-access form, without documentation, getting-started guide or API reference, and its buttons carry no destination in the HTML. The genuinely open technical route — the GitHub repositories and the Hugging Face models — assumes an ML profile and the usual prerequisites. Using the verticals is straightforward: Collov AI on the web, NewEyes and CozyAI as mobile apps, with account creation and credits, and API integration only under a separate agreement for enterprise accounts.
Deployment
Integrations
Behind Collov Labs
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What is Collov Labs?
How does Collov Labs relate to Collov AI, NewEyes and CozyAI?
How much does Collov Labs cost?
Is there a free plan or a free trial?
Does Collov Labs publish an API?
Are my interactions used to train the models?
Who owns the generated images?
Where is my data hosted, and is the GDPR covered?
Can the outputs be used in a property listing?
Who can sign up, and how do I reach the team?
Should you pick Collov Labs?
Collov Labs is the platform and research layer of Collov, Inc., not a consumer tool: it is judged on the architecture it describes and on the verticals derived from it. The technical credibility is documented rather than asserted — seven papers cited with their venues, code on GitHub, models on Hugging Face, and silicon partnerships with Intel and Qualcomm that put inference on a local NPU. A $23M Series A announced in April 2026 and three verticals already in service (Collov AI for real estate, NewEyes for mobile capture, CozyAI for prosumers) point to a company past the prototype stage.
The commercial opacity is just as real. There is no price, no plan and no API documentation for the Labs platform, the site is a single page whose action buttons carry no verifiable destination, and the flagship WorldFM demo is still marked “Coming Soon”. Two points deserve attention before any engagement. On data: the site states that production interactions become structured training signals, and no opt-out is documented. On legal footing: the terms and the privacy policy are hosted on the neighboring collov.ai domain and written in the name of Collov.ai, the policy is dated October 2024 and still carries leftover mentions of another vendor, it claims CCPA but never GDPR compliance, it names no Article 27 representative, and no postal address appears anywhere.
The result is a credible engineering story behind a deliberately closed commercial front door. The evaluation is worth the effort if you are assessing a visual agent platform, following research in the field, or considering one of the verticals for real estate, capture or design work. If you need published pricing, a self-service API or contractual guarantees such as a DPA, expect to obtain them by contacting the company, and to verify them yourself.
- Choosing a selection results in a full page refresh.
- Opens in a new window.