You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$25,000 left to reach the funding goal.
A generated world can look convincing but fall apart when you move, change an object or ask a guide about what you see. I am asking for USD 25,000 to build and test an open component that keeps these actions tied to the same scene. A USD 5,000 minimum would fund a smaller pilot, not the full study. Watch the research film: https://kumarnavish.github.io/sciescape-research/#film
Give rooms, openings and objects stable identities. Let rendering, movement, edits and the guide use one versioned scene state. Check changes before showing them; a failed change should leave the previous state intact. Release a small viewer, test scenes and replay tools that work without the private Sciescape application. Saved-scene replay should need no paid API. The aim is a useful, reproducible step toward coherent worlds, not a promise to finish a photorealistic avatar or the whole experience.
Over 12 weeks: first reproduce failures and pilot the baselines; then build shared state, checked edits and recovery; finally run and publish a fixed comparison. After development, freeze 21 new scene requests, two seeds, movement/edit scripts, scoring and software versions. Compare four conditions: the current depth-panorama pipeline, shared state, separate interaction state, and a runnable public baseline with compatible licensing. That gives 168 planned journeys. Record passable routes, object identity on return, persistent edits and unsupported guide claims, alongside visual quality, latency, memory and cost. Independent renderer evidence checks what the guide can see. Count timeouts and refusals as failures. Any necessary scope reduction happens before freezing the evaluation.
Success is a standalone tool other researchers can run, complete results and a measured improvement in whole journeys without hiding worse visual quality or excessive cost. Compare outcomes by scene and report uncertainty. An independent engineer will reproduce the runner and review scoring; nobody has yet been appointed. If shared state does not help, release the negative result and explain why. Text2Room, DreamScene360 and explicit scene representations already exist; the research question is whether this combination provides a useful, measurable improvement.
The full USD 25,000 budget is USD 16,200 for 180 hours of research/implementation, USD 4,800 for 60 hours of independent engineering/reproduction, USD 3,500 for GPU and host compute, and USD 500 for storage/transfer. I propose 15 hours per week for 12 weeks alongside my doctoral work. These are budget allowances, not supplier quotes or existing commitments. Checkpoints, CPU replay, bounded retries and automatic shutdown will conserve compute. Spending starts only after funding and terms are agreed and stays within available project funds.
At the USD 5,000 minimum: USD 3,600 buys 40 applicant hours; USD 800 buys 10 hours of independent review; USD 500 covers compute; USD 100 covers storage/transfer. This funds a small shared-state prototype, failure replay and a six-scene pilot comparing two conditions with one seed: 12 planned pilot journeys. It does not fund the 168-journey evaluation, a new avatar or a complete voice experience. Results would be labeled exploratory. Intermediate funding would support an explicitly revised scope and budget before work begins.
I am Navish Kumar, an employed doctoral researcher at the University of Basel, Switzerland. Sciescape is my independent personal project. Basel is my affiliation, not a project sponsor; no advisor, faculty PI or institutional signatory is involved. I am the sole applicant and use coding agents as engineering tools. The public release contains five original panoramas with provenance and a checkpoint/resume utility whose 23 CPU tests passed on 15 November 2026. Its evidence record also reports only 2 of 21 geometry development cases completed. Those cases informed development and are not independent validation. The full free-moving world, lifelike guide and natural voice journey remains unsolved. No external adoption is claimed.
The main risk is that the scene representation preserves state but cannot repair poor geometry, or that better consistency costs too much latency or visual quality. GPU capacity, baseline licensing and third-party model behavior may also limit comparisons. Pilot compatibility and cost first; keep the evaluation small enough to finish; preserve failures and release useful replay tools even if the hypothesis fails. Synthetic scenes and licensed voices avoid private recordings and real-person likenesses. This study cannot establish learning gains or human acceptance.
Sciescape has used promotional Google Cloud credits and Kaggle/Colab allocations for prototyping. Their total monetary value has not been reconciled. Applications to Lambda, CloudRift, fal, OpenAI, EGI, Yotta, inference.net and the Sequoia OSS Fellowship are pending. The same USD 25,000 work package has been sent to Sentient for preliminary review. These are alternative or coordinated funding routes, not separate claims for the same costs. No award is assumed. Before accepting funding, I will disclose decisions, remove overlapping costs or agree distinct work packages, and never charge the same cost twice.
New funded code and documentation will be Apache-2.0. Evaluation materials and examples will have appropriate open permissions, subject to upstream rights. The tools must run independently of the private Sciescape product. Private application code, credentials, restricted assets and third-party model weights are excluded. Noncommercial film/media terms do not become Apache-2.0 through this proposal.
AI tools helped research this opportunity and draft and edit this proposal from my project records and stated goals. This assistance is disclosed; it is not independent validation of the research or a claim that the proposed results already exist.