RealFusion¶

Development line: project:realfusion · thread realfusion-development
Last event: 2023-03-03 · 1 dated since 2023-03-03 · Researched: 2026-09-04 · confidence: high

What it is¶

RealFusion reconstructs a plausible 360-degree object from one masked image by combining single-image textual inversion, Stable Diffusion score distillation, and an Instant-NGP-style NeRF.

• token inversion: learns an image-specific token
• radiance field: optimizes against the input image and diffusion prior
• mesh export: exports a textured mesh optionally

Unseen geometry is inferred, not recovered; failed convergence and Janus-style duplicated faces remain documented failure modes.

Use it as a reproducibility baseline or a tunable research workflow, not as a current turnkey image-to-3D product.

Development line¶

• 2023-03-03 — RealFusion code and project resources linked. On 2023-03-03, the RealFusion development line recorded a code-release entry with links to the project's public page and GitHub repository. This establishes a dated public reference point, but no version, commit, or specific technical change is identified.

What changed¶

2023-03-03 — RealFusion’s public project page and code release made the single-image textual-inversion-plus-NeRF workflow runnable outside the paper implementation. 2023-04-08 — The later reference pointed back to the 2023-03-03 release; no separate RealFusion code, model, or method update is evidenced by that link.

How to use this¶

From 2023-03-03, practitioners could use the dated RealFusion project page and repository links as the starting point for locating its publicly referenced code-release resources.

1. Install the Python requirements, PyTorch separately, and the CUDA extensions; install nvdiffrast only when exporting a textured mesh. — https://github.com/lukemelas/realfusion/blob/main/README.md
2. Prepare a square RGBA image containing the object and mask; use the supplied mask-extraction script when starting from an unmasked image. — https://github.com/lukemelas/realfusion/blob/main/README.md
3. Run single-image textual inversion with Stable Diffusion v1.5 to create the learned object embedding. — https://github.com/lukemelas/realfusion/blob/main/README.md
4. Run reconstruction with python main.py --O, the RGBA image, learned embedding, and matching Stable Diffusion checkpoint. — https://github.com/lukemelas/realfusion/blob/main/README.md
5. Inspect novel views and tune camera pose, camera radii, losses, and random seeds for the specific image before accepting an output. — https://github.com/lukemelas/realfusion/blob/main/README.md

Best practices¶

• Start with a clean masked object image; the documented input contract is a square RGBA image. — https://github.com/lukemelas/realfusion/blob/main/README.md
• Use --O: the repository states that its CUDA raymarching path was developed and tested with that optimization bundle. — https://github.com/lukemelas/realfusion/blob/main/README.md
• Match pose_angle to the camera viewpoint and set radius_rot slightly above the training radius maximum. — https://github.com/lukemelas/realfusion/blob/main/README.md
• Run multiple seeds and select outputs after inspection; default parameters are explicitly not equally good for every example. — https://github.com/lukemelas/realfusion/blob/main/README.md
• Treat backsides as plausible extrapolations, and reject transparent fields, floaters, incorrect geometry, or duplicated-face outputs rather than presenting them as recovered geometry. — https://arxiv.org/abs/2302.10663

Superseded by this¶

• 2023-03-03 — The public repository’s refactored release superseded reliance on the paper’s original implementation details; its maintainer states that parts differ from the paper code.
• 2023-04-08 — Treating the later reference as a separate release is obsolete: it links to the earlier 2023-03-03 item rather than documenting a new release.

Still unknown¶

• No dated first-party release, checkpoint, compatibility test, or successor-method announcement was found for the 2023-04-08 reference; it appears to be a retrospective pointer to the earlier item, not a distinct development step.
• The current repository documents a CUDA-dependent research workflow but does not provide evidence of compatibility with current PyTorch, CUDA, Diffusers, or Stable Diffusion ecosystem versions.

Sources¶

Agent brief¶

• Subject: project:realfusion, thread realfusion-development, 1 dated events 2023-03-03 → 2023-03-03.
• Practical note: From 2023-03-03, practitioners could use the dated RealFusion project page and repository links as the starting point for locating its publicly referenced code-release resources.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.