[Contribution]: Investigate why LLM-assisted generation produces incomplete models vs. the official Pikachu demo

What are you proposing?

A bounded new fix

Target issue URL or number

new bounded fix

Proposed scope

The official img2threejs site showcases the "One photo in. A procedural model out." demo (Pikachu 10K Star Celebration): a complete, polished procedural model generated from a single reference photo.

However, when I try to reproduce this kind of result using mainstream AI assistants (Claude, Kimi, GPT) against the img2threejs pipeline, every attempt produces an incomplete / broken model — missing or malformed parts compared to the official render (screenshots available; official demo render vs. AI-generated result). Each retry consumes a lot of compute, and I have not been able to get a finished model in one shot.

I would like to propose a bounded fix: investigate why one-shot, LLM-assisted generation falls short of the official demo quality, and contribute either (a) a documented, reproducible prompt/workflow that reliably produces a complete model from a single reference photo, or (b) a small pipeline fix, if the root cause turns out to be in the pipeline itself.

Intentionally out of scope: redesigning the generation pipeline, new input modalities, or quality work beyond making a single complete character model reproducible.

Questions for maintainers:

1. Is the official Pikachu result actually achievable in one shot, or was it produced through many internal iterations?
2. Is iterative refinement the expected workflow for users, i.e. is "one photo in, finished model out" not literally one pass?

Acceptance criteria and validation plan

Maintainers can consider the fix complete when:

1. A documented workflow/prompt exists that produces a complete (non-broken) model from a single reference photo, verified with before/after renders; or
2. A pipeline patch lands that measurably reduces incomplete-geometry failures, with test renders attached to the PR.

I will validate by running the documented workflow / patched pipeline on the same reference photo used in the official demo and attaching the resulting renders as visual evidence.

Availability

I can dedicate a few hours per week over the next couple of weeks to testing prompts, collecting failure cases, and validating any fix.

Acknowledgement

• I understand this form does not assign or reserve the work, and I will reference the issue number in any implementation PR.