Enderfga/SolarRenderer-Data

SolarRenderer: Few-Step Generative Reference Rendering

🌐 Project Page • 📄 Paper • 🤗 Weights • 💻 Code • 🛠️ Data code

Training data and benchmark of SolarRenderer, which renders structure from game engines and agents (grey boxes, semantic maps, masks and skeletons, and structure types it never saw in training) into finished video.

Contents

Part	Size	What is in it
train/magpie	30,597 clips	render, scene grey box, first frame, prompt
train/carla	25,317 clips	render, semantic segmentation, first frame, prompt
train/unrealzoo	1,824 clips	render, object masks + skeletons, first frame, prompt
benchmark/hard	208 cases	structure, first frame, prompt, tags (16 structure types, 11 unseen in training)
benchmark/heldout	53 clips	structure, real first frame, prompt, ground truth
results/	522 renders, 44 files of verdicts	renders of both benchmark splits by SolarRenderer and its causal variant, and the judge verdicts behind the paper's Elo tables

📌 Of the training sources, the Magpie, CARLA and UnrealZoo clips are released; held-out test clips of the engine recordings with near-range depth are in benchmark/heldout.

Quick Start

import webdataset as wds
from huggingface_hub import HfFileSystem

base = "https://huggingface.co/datasets/Enderfga/SolarRenderer-Data/resolve/main/"
urls = [base + p.split("/", 3)[3] for p in HfFileSystem().glob("datasets/Enderfga/SolarRenderer-Data/train/magpie/*.tar")]
for clip in wds.WebDataset(urls).decode("pil"):   # first_frame.png -> PIL image, the videos stay mp4 bytes
    render, condition, first_frame, meta = clip["render.mp4"], clip["condition.mp4"], clip["first_frame.png"], clip["json"]
    break

Benchmark only: hf download Enderfga/SolarRenderer-Data --repo-type dataset --include "benchmark/*"; our renders and the judge verdicts: --include "results/*". Scoring code: eval/ in the main branch. To train on the shards, tools/hf_samples.py turns them into training samples (guide).

Format

• Training clips are exactly what the model saw: 24 fps, the first 158 frames, LANCZOS to 1344×768, the structure cut the same way. H.264, CRF 18 (render) / 15 (structure).
• Held-out clips (24 fps, 158 frames, 1344×768) are the very files the paper's held-out numbers were computed from, so eval/ on results/solarrenderer reproduces them.
• Training shards are WebDataset tars of about 2 GB; one clip = <key>.render.mp4, <key>.condition.mp4, <key>.first_frame.png, <key>.json. train/index.jsonl lists every clip with its shard.
• training_input in the index says which inputs training gave the clip; the prompt is written for that combination.

train/index.jsonl, train/<source>/<source>-00000.tar ...
benchmark/hard/cases.jsonl, benchmark/hard/<case>/{condition.mp4, first_frame.png, prompt.txt}
benchmark/heldout/cases.jsonl, benchmark/heldout/clips/<source>/<clip>/{render.mp4, condition.mp4, first_frame.png}
results/solarrenderer/{hard/<case>.mp4, heldout/<source>/<clip>.mp4}
results/solarrenderer_causal/...      same layout, the causal variant
results/judge/{hard,heldout}/<method>__vs__<method>.jsonl
page/                                 videos of the project page

Prompts

• Six parts: subject_definitions, summary, retention_analysis, detailed_description (350–500 words), overall_soundscape, non_diegetic_music.
• The parts that name the inputs are fixed sentences per structure type; the rest was written per clip by Kimi K2.6.
• Captioning pipeline: data branch. Prompt-writing skill: skills/solarrenderer-caption/ in the main branch.

Benchmark

• hard: 57 structure clips, each in several looks. 74 cases have a structure rendering not seen in training (condition_seen_in_training: false), including the DirtyMoCap SMPL renders with markers.
• Structures: mostly built by our agent in a game engine; 49 near-field shaded geometry cases from Depth Anything 3 depth of MiniMax-H3 clips; CARLA semantics, fitted SMPL, a Mixamo character, and the DirtyMoCap project video as is.
• First frames: Nano Banana redraws of the structure's frame 0 in each case's look.
• heldout: 53 clips from CARLA (14), engine recordings with near-range depth (9), Magpie (16) and UnrealZoo (14), held out by recording, scene or map (CARLA by clip, so its maps also appear in training). Each gives the structure video, the real first frame and the prompt; render.mp4 is the ground truth.
• results/solarrenderer: the SolarRenderer renders scored in the paper (four steps, seed 0).
• results/solarrenderer_causal: the renders of the causal LoRA pair, streamed block by block (four steps per block, seed 0), scored in the paper as Ours (causal).
• results/judge: pairwise verdicts in the format of eval/judge.py, among the main table's nine methods and the four-step LoRAs we started from (LongLive-Plug, DMAD, FastH3 with our control LoRA). eval/elo.py --methods with the nine methods gives the main table's Elo. The causal variant is judged only against SolarRenderer (solarrenderer__vs__solarrenderer_causal.jsonl); adding it as a tenth method and shifting the ratings so that the other nine average 1000 gives its row.

License

Part	License
Magpie clips	Apache-2.0, from Magpie
CARLA clips	CC BY 4.0; rendered by us in CARLA 0.9.16 (code MIT, assets CC-BY)
UnrealZoo clips	CC BY-NC 4.0; rendered by us in UnrealZoo (Apache-2.0)
Benchmark	ours, except inputs noted per case; held-out engine-recording clips for research use only
page/cmp	excerpts of the published demos of the 15 works compared on the project page, shown for comparison; each remains under its authors' terms
Other prompts, indexes, metadata	CC BY 4.0

Code: Apache-2.0. Weights: MiniMax H3 Community License (see the weights repository).

📮 Questions or removal requests: guianfang@u.nus.edu

Citation

@article{fang2026solarrenderer,
  title  = {SolarRenderer: Few-Step Generative Reference Rendering},
  author = {Fang, Guian and Qian, Shengju and Lu, Dongyue and
            Huang, Junchao and Jiang, Lutao and Wang, Xin and
            Song, Yiren and Jiang, Li and Shou, Mike Zheng},
  year   = {2026}
}