# RenderFormer — README Reference > A Cornell box turned into a cover page: a red, green and blue cube logo, a two-line ruled title, a black gallery of rendered scenes, then a nested contents list that maps the whole manual. **Theme:** light Read from `README.md` on the default branch `main` of github.com/microsoft/renderformer (last commit touching the file eec6a9b, 2026-10-13; captured 2026-10-28). Evidence is the raw Markdown, the HTML that GitHub's API returns for it, a 1440×900 first-screen capture and a 2700px tall capture of the light theme, with the README column 920px wide (856px content width). Image sizes and logo colors come from `medias/rf.png` and `medias/gallery-10.jpg` fetched from the repo. GitHub sets the type (its system stack, Segoe UI in the Windows captures), the 16px body size and the heading rules; nothing in the README overrides them. ## Overview The first screen is a centred stack that reads like a paper's title page set in GitHub's own type. A 320px square logo (a white "RF" cube on red, green and blue quadrants with a dark navy outline) sits on top, followed by the full paper title as a centred `

` that breaks onto two lines, a centred row of five bold linked author names separated by middle dots, and the venue as a centred `

` "SIGGRAPH 2025 Conference Papers". Both headings carry GitHub's grey bottom rule. Under the venue a full-width 5×2 gallery of rendered scenes on black backgrounds fills the bottom of the window. Below the fold come a centred plain-text caption, a centred row of five bold text links separated by `|`, one long left-aligned abstract paragraph, and then a `# Table of Content` with a four-level nested bullet list. The body uses `#` (h1) for every top-level section, so Installation, Usage, Bring Your Own Scene!, Acknowledgements, License and Citation all render at 32px with a rule. There are no badges, no emoji in headings and no news list. ## Header Block The whole header is wrapped in one outer `

`; GitHub's parser closes that paragraph as soon as it meets the inner `

`, so in the rendered HTML each element stands on its own and the outer `

` only contributes an empty paragraph. Rendered positions are y values in the 1440×900 capture (the README box starts at y≈25 with its "README.md" file bar; content starts at y≈80). 1. **Logo** (y≈110 to 430): `

`. The source is a 720×720 RGBA PNG; `width=320` renders it at 320×320, centred. GitHub wraps it in a link to the file. 2. **Title** (y≈482 and 522, rule at about 550): `

RenderFormer: Transformer-based Neural Rendering of Triangle Meshes with Global Illumination

`. GitHub h1, 2em (32px), weight 600, wrapping to two centred lines with a 1px grey rule under it across the 856px content width. The project name and the paper title are the same heading; there is no separate short name. 3. **Authors** (y≈579): one `

` with five `Name` entries separated by `·` on its own source line. All five names link to personal pages or Google Scholar. No affiliation marks, no affiliation line and no equal-contribution mark. 4. **Venue** (y≈630, rule at about 652): `

SIGGRAPH 2025 Conference Papers

`. GitHub h2, 1.5em (24px), with the same bottom rule; it also becomes an entry in GitHub's outline menu. 5. **Gallery teaser** (y≈670 to about 1013, crossing the fold): `
Examples of triangle-mesh based scenes ... and multiple light sources.
`. The JPG is 4345×1741 with no width attribute, so GitHub's `max-width: 100%` scales it to 856×343. It shows ten renders in a 5×2 grid, most on black. 6. **Caption** (y≈1030 to 1080): the plain sentence that follows the `` inside the same `
`, so it renders centred at body size directly under the image, three lines long, without italics or a "Figure" label. 7. **Link row** (y≈1126): `


Project Page | arXiv | Paper | Model | Official Code

`. Five bold links in GitHub link blue separated by plain pipes; the leading `
` adds one blank line above the row. 8. **Abstract paragraph** (y≈1160 to 1370): one plain Markdown paragraph of seven sentences, left-aligned, with no bold and no heading. It ends the header block. ## Badges The README uses no shields.io badges. The navigation is a row of bold text links that act as buttons. | Label | Message | Color (hex) | Style | Links to | |---|---|---|---|---| | (link row) | Project Page | #0969da (GitHub link blue) | bold text link, separated by ` \| ` | microsoft.github.io/renderformer/ | | (link row) | arXiv | #0969da | bold text link | arxiv.org/abs/2505.21925 | | (link row) | Paper | #0969da | bold text link | renderformer.github.io/pdfs/renderformer-paper.pdf | | (link row) | Model | #0969da | bold text link | huggingface.co/microsoft/renderformer-v1.1-swin-large | | (link row) | Official Code | #0969da | bold text link | github.com/microsoft/renderformer (the repo itself) | | (author row) | five author names | #0969da | bold text links separated by `·` | personal pages and Google Scholar | The link blue is the value in the github-markdown-css light theme that the capture uses; the README sets no color of its own. Other colors visible in the header come from the logo, not from badges: red #f23340 (top left quadrant), green #5e9d59 (top right), blue #3c91cb (bottom) and the navy outline and shadow #042a45 around the white "RF" cube. The gallery adds the red and green walls of a Cornell box on black, which echo the logo's red and green. ## Section Outline | Level | Text | What it holds | |---|---|---| | h1 (centred) | RenderFormer: Transformer-based Neural Rendering of Triangle Meshes with Global Illumination | title; authors follow without a heading | | h2 (centred) | SIGGRAPH 2025 Conference Papers | venue; gallery, caption, link row and abstract follow without a heading | | h1 | Table of Content | a nested bullet list of 28 anchor links on four levels (`-`, `*`, `+`, `-` markers by depth) | | h1 | Installation | h2 Prerequisites (two bold-led bullets: System, Hardware), h2 Environment Setup (one `bash` block), h2 Pretrained Models (four-column model table, a paragraph on safetensors, a `
` note) | | h1 | Usage | h2 Image Rendering with h3 Scene Conversion, h3 Rendering a Single Image Using Inference Script (h4 Available Arguments), h3 Inference with `RenderFormerRenderingPipeline` (a `python` block); h2 Video Rendering with h3 Download Example Data, h3 Rendering a Video (h4 Available Arguments) | | h1 | Bring Your Own Scene! | h2 Scene Definition JSON with seven h3 subsections (structure, objects, camera, example, HDF5 fields, remeshing, Blender reference), h2 Blender Extension, h2 Scene Setting Tips (a numbered list with nested ranges) | | h1 | Acknowledgements | a three-item code list and a long list of 3D model credits with bold model names | | h1 | License | one paragraph and two bullets for submodule licenses | | h1 | Citation | one sentence and a `bibtex` block | Absent from the source: badges, news or changelog, results table, demo GIF or embedded video, project logo in a `` dark variant, quick start separate from Installation, contact section, emoji in headings. Present but unusual: every top-level section is an h1 rather than an h2, so the body has seven 32px ruled headings; the venue is an `

` heading rather than bold text; the header is wrapped in an outer `

` that GitHub closes immediately; the table of contents is hand-written and three of its anchors do not match GitHub's generated ids (`#usage-1` where the heading id is `usage`, `#inference-with--renderformerrenderingpipeline-` where it is `inference-with-renderformerrenderingpipeline`, and `#bring-your-own-scene-` where it is `bring-your-own-scene`), so those three links do not jump; the "Available Arguments" blocks are argparse help text fenced as `bash`. ## Components **Centred stack in HTML.** Logo and gallery each sit in their own `

`; title, authors and venue are `

`, `

` and `

` with `align="center"`. Spacing comes only from GitHub's heading and paragraph margins plus one `
` before the link row. **Bold link rows.** Two rows use the same recipe: `Text` entries on separate source lines, joined by a separator on its own line (`·` between authors, `|` between destinations). Both render in #0969da at body size. **Table of contents.** `# Table of Content` followed by a nested list that uses `-` for the first level, `*` for the second, `+` for the third and `-` again for the fourth, each indented two more spaces. GitHub renders all of them as bullets (disc, circle, square) with 2em indentation per level. Inline code inside a heading name is kept in the link text ("Inference with `RenderFormerRenderingPipeline`"). **Bold-led bullets.** Prerequisites use `- **System**: ...` and `- **Hardware**: ...`, with key terms (Linux, MacOS, Windows, NVIDIA CUDA GPUs, Apple Metal GPUs) also in bold. **Model table.** `| Model | Params | Link | Model ID |` with `|-------|` delimiters and no alignment colons. Two rows (RenderFormer-V1-Base 205M, RenderFormer-V1.1-Large 483M); the Link cell is `[Hugging Face](...)` and the Model ID cell is inline code (`microsoft/renderformer-v1-base`). No bold best values. **Collapsible note.** `
Note on the released models
...
` hides a one-paragraph caveat about a shader bug and retrained models; collapsed by default. **Code blocks.** Fenced with language tags: `bash` for install, conversion, inference commands and argparse help listings, `python` for the minimal `RenderFormerRenderingPipeline` example (with a trailing comment showing the expected output shape), `bibtex` for the citation. GitHub renders them on its light grey code background with a copy button. **Field lists.** Scene and HDF5 formats are documented as bullets whose first token is an inline-code field name (`` `mesh_path` ``, `` `triangles` ``), with nested bullets for sub-fields and array shapes in brackets ("[N, 3, 3]"). **Credits list.** Acknowledgements list each 3D model as a bold name, linked when a source exists, followed by ": By Author" or ": from Repository". **Citation.** "If you find this work helpful, please cite our paper:" then `@inproceedings {zeng2025renderformer, ...}` with fields aligned in a column (`title = {...}`). ## Imagery | Image | Type | Source size | Markdown | Rendered at 920px | Placement | |---|---|---|---|---|---| | rf.png | PNG logo, square, RGBA | 720×720 | `` in `
` | 320×320 | top of the README | | gallery-10.jpg | JPG grid of ten renders, mostly black backgrounds | 4345×1741 | `` with no width, in `
` | 856×343 | after the venue heading; plain centred caption in the same `
` | The gallery is the only figure in the README. It is a 5×2 grid of RenderFormer outputs (Cornell box, glossy objects, an "RF" letter scene with colored walls, a tree, a fox, crystals and others), so the first screen shows results rather than a method diagram. No GIF, no video, no `` dark variant. The logo's flat quadrants (#f23340, #5e9d59, #3c91cb) are the only designed color on the page. ## Adapting to Other Content - **A library.** Keep the 320px logo, put the library name alone in the `

` (one line), replace the author row with a one-line description, keep the bold `|` link row (Docs, PyPI, Examples, Code) and keep the nested Table of Contents; drop the venue `

`. - **A dataset.** Keep the full header, swap the gallery for a sample grid at its natural width (scaled to 856px), and put a Download section with a table shaped like the model table (Split, Size, Link, ID) first in the contents list. - **A benchmark with scores.** Keep the header and add a Results section after Installation with a pipe table, numeric columns right-aligned (`| ---: |`) and the best value in bold; RenderFormer has no such table, so this is an extension. - **A repo without a teaser.** Let the bold link row follow the venue heading directly and move the abstract paragraph up; the logo alone carries the first screen at 320×320. - **A shorter manual.** Use `##` for top-level sections instead of `#` and drop the Table of Contents once there are fewer than about six sections; keep the order Installation, Usage, Citation. ## Do's and Don'ts | Do | Don't | |---|---| | Put the logo at `width=320` in its own centred `
` | Let a large square logo render at its 720px source size | | Join author names with `·` and destinations with `\|`, all as bold links | Mix badges and text links in the same row | | Let a results gallery fill the 856px width with a centred caption under it | Shrink the gallery into a thumbnail | | Map a long manual with a nested Table of Contents | Hand-write anchors without checking them against GitHub's ids (three of RenderFormer's do not resolve) | | Hide caveats about released weights in `
` | Put model-release caveats inline above the install commands | | Tag every code block (`bash`, `python`, `bibtex`) | Leave argparse help untagged in a long block | ## Agent Prompt Guide 1. "Write a README header in GitHub HTML: a centred `
` with a square logo at `width=320`, a centred `

` with the full paper title, a centred `

` of bold author links separated by `·`, a centred `

` with the venue, a centred `
` holding a full-width results gallery image (no width attribute, scales to 856px) followed by a plain caption sentence, and a centred `

` that starts with `
` and holds five bold links (Project Page | arXiv | Paper | Model | Official Code) in GitHub link blue #0969da. No badges." 2. "After a one-paragraph abstract, add `# Table of Content` as a nested bullet list of anchor links on up to four levels, then top-level `#` sections in this order: Installation (Prerequisites, Environment Setup, Pretrained Models), Usage, Bring Your Own Scene!, Acknowledgements, License, Citation. Check every anchor against GitHub's heading ids." 3. "Write the Pretrained Models table as `| Model | Params | Link | Model ID |` with a `[Hugging Face](...)` link and the model ID in inline code, follow it with a `

Note on the released models` caveat, and document input formats as bullets whose first token is an inline-code field name." ## Similar Pages - readme-vveicao-freeorbit4d also opens with a logo image above a centred title, venue line and authors, but uses `for-the-badge` badges and an HTML results table where RenderFormer uses bold text links and one gallery image. - readme-deepmancer-hairport shares the paper-like reading order (title, venue, authors, figure with caption) but puts a news list and numbered figure captions before installation and has no table of contents. - readme-eku127-satnav uses the same centred `

` and `

` pair with rules, but with a short project name in the `

` and a uniform badge row. ## Quick Start ````markdown

ProjectName: Full Paper Title That May Wrap Onto Two Lines

Author One · Author Two · Author Three

Venue Year Track

One plain sentence describing what the gallery shows.


Project Page | arXiv | Paper | Model | Official Code

One abstract paragraph, left-aligned. # Table of Content - [Installation](#installation) * [Prerequisites](#prerequisites) * [Environment Setup](#environment-setup) * [Pretrained Models](#pretrained-models) - [Usage](#usage) - [Bring Your Own Data](#bring-your-own-data) - [Acknowledgements](#acknowledgements) - [License](#license) - [Citation](#citation) # Installation ## Prerequisites - **System**: Tested on ... - **Hardware**: Tested on ... ## Environment Setup ```bash git clone https://github.com/org/name cd name pip install -r requirements.txt ``` ## Pretrained Models | Model | Params | Link | Model ID | |-------|--------|------|----------| | Name-Base | 000M | [Hugging Face](https://huggingface.co/org/name-base) | `org/name-base` |
Note on the released models
One paragraph of caveats.
# Usage ```python from name import Pipeline ``` # Bring Your Own Data - `field_name`: what it holds # Acknowledgements # License # Citation If you find this work helpful, please cite our paper: ```bibtex @inproceedings{key, title = {...}, } ``` ```` ## Provenance - Repo: https://github.com/microsoft/renderformer (`README.md`, branch `main`, last README commit eec6a9b on 2026-10-13). - Paper: "RenderFormer: Transformer-based Neural Rendering of Triangle Meshes with Global Illumination", SIGGRAPH 2025 Conference Papers, arXiv 2505.21925; first author Chong Zeng. Project page https://microsoft.github.io/renderformer/. - Selection: a model first screen, with a square logo, centred title, author line, venue line, a sharp captioned teaser grid, a single row of text links, then a full nested table of contents that maps the whole manual before installation starts. - Evidence: raw Markdown and GitHub-rendered HTML (saved 2026-10-28), 1440×900 and 2700px captures at a 920px README column, `medias/rf.png` and `medias/gallery-10.jpg` fetched from the repo for pixel sizes and logo colors, heading ids in the rendered HTML checked against the table of contents anchors. - Written 2026-10-28.