Any structure in. Any world out.

SolarRenderer

Few-Step Generative Reference Rendering

The agent builds it. We render it.

PaperCodeWeightsData

Overview

Game engines and agents can build the layout, motion and camera of a scene. SolarRenderer turns that structure into finished video with a control LoRA on MiniMax-H3, and a few-step LoRA post-trained with structure guidance renders it in four steps, 12.3 times faster than fifty. A block-causal version of both LoRAs renders the video as a stream, block by block.

One model, many structures

16

structure types, 11 of them never seen in training

Four steps

12.3×

faster sampling than 50 steps, with better image quality and structure

Open

208

benchmark cases released with the weights and training data

Structures

InputSolarRenderer
Point cloudunseenIndiana
InputSolarRenderer
2D pose skeletonunseenTech stage
InputSolarRenderer
WireframeunseenKaiju
InputSolarRenderer
Normal mapunseenJedi
InputSolarRenderer
Pencil storyboardunseenPixel art
InputSolarRenderer
VoxelsunseenWesteros
InputSolarRenderer
Child's crayon drawingunseenUkiyo-e
InputSolarRenderer
SilhouetteunseenBaghdad
InputSolarRenderer
Colour-mapped depthunseenSpace opera
InputSolarRenderer
3D bounding boxes (driving format)unseenMoscow snow
InputSolarRenderer
Mixamo characterunseenCairo
InputSolarRenderer
Object masks + skeletonsDaylight

Styles

Choose a structure type and one of its samples to watch a single input clip rendered in several styles, where the structure keeps the layout, the motion and the camera of every frame while the first frame and the text decide what the world looks like.

Input
Altair
Ezio
Connor
Edward
Arno
Bayek
Eivor
Naoe

Method

A control LoRA teaches the frozen backbone to follow a structure video, and a few-step LoRA post-trained with structure guidance renders it in four steps.

Results

A Gemini judge (gemini-3.8-flash) compares the methods in pairs on every case, and one Elo rating per method is fit for each of the four questions. A blind test with human judges agrees with the direction of the judge’s preferences.

MethodSteps Generalization benchmark208 casesGeneralization208 casesIn-distribution test set53 clipsIn-distribution53 clips
SolarRendererours 4 1 1441 1 1510
LynnReal-Omnisame backbone 4 1 1441 2 1410
Our control modelwithout acceleration 50 3 1389 3 1299
PAI ControlNet-Union 40 1221 1189
Official backboneMiniMax-H3 Ref2VA 50 1075 651
Streaming CWMsame backbone 4 1075 988
StreamRender-H3same backbone 2 580 651
Cosmos-Transfer2.5 35 389 651
Wan2.2-VACE-Fun 50 389 651
MethodSteps Generalization benchmark208 casesGeneralization208 casesIn-distribution test set53 clipsIn-distribution53 clips
SolarRendererours 4 1 1420 1 1424
LynnReal-Omnisame backbone 4 2 1412 2 1359
Our control modelwithout acceleration 50 3 1339 1196
PAI ControlNet-Union 40 1228 3 1304
Official backboneMiniMax-H3 Ref2VA 50 1086 586
Streaming CWMsame backbone 4 1100 968
StreamRender-H3same backbone 2 560 586
Cosmos-Transfer2.5 35 368 991
Wan2.2-VACE-Fun 50 489 586
MethodSteps Generalization benchmark208 casesGeneralization208 casesIn-distribution test set53 clipsIn-distribution53 clips
SolarRendererours 4 1 1338 1 1386
LynnReal-Omnisame backbone 4 3 1241 3 1295
Our control modelwithout acceleration 50 2 1276 2 1308
PAI ControlNet-Union 40 1083 1187
Official backboneMiniMax-H3 Ref2VA 50 627 554
Streaming CWMsame backbone 4 930 955
StreamRender-H3same backbone 2 737 750
Cosmos-Transfer2.5 35 1007 1012
Wan2.2-VACE-Fun 50 763 554
MethodSteps Generalization benchmark208 casesGeneralization208 casesIn-distribution test set53 clipsIn-distribution53 clips
SolarRendererours 4 1 1428 1 1522
LynnReal-Omnisame backbone 4 2 1423 2 1362
Our control modelwithout acceleration 50 3 1395 3 1346
PAI ControlNet-Union 40 1150 1029
Official backboneMiniMax-H3 Ref2VA 50 1153 660
Streaming CWMsame backbone 4 1199 1102
StreamRender-H3same backbone 2 498 660
Cosmos-Transfer2.5 35 377 660
Wan2.2-VACE-Fun 50 377 660

Fifteen published works on their own demos

86 examples taken from the works' own pages, rendered by SolarRenderer from the structure they released. * The 95% bootstrap interval excludes zero.

49 wins 13 ties 24 losses
+0.29*

Elo from the Gemini judge's verdicts on every pair of methods on a split, one rating per question (Bradley–Terry, 400 points per factor of ten in odds, mean 1000). Each case is judged twice with the two videos swapped. Medals mark the top three of each column; methods that lose every comparison share one rating. The four-step LoRAs we started from are compared in the paper's ablations.

Comparisons

We render 86 examples that fifteen published works show on their own pages, from the structure they released, with their reference image or the first frame of their result as our first frame. Each row shows their input, their result and SolarRenderer.

Input · grey box
Magpie
SolarRenderer · 4 steps
Input · grey box
Magpie
SolarRenderer · 4 steps
Input · grey box
Magpie
SolarRenderer · 4 steps
Input · grey box
Magpie
SolarRenderer · 4 steps
Input · grey box
Magpie
SolarRenderer · 4 steps
Input · grey box
Magpie
SolarRenderer · 4 steps

Benchmark and data

Everything below is released, except the training clips we may not redistribute.

Generalization benchmark

208

cases over 16 structure types, 11 unseen in training

In-distribution test set

53

clips with ground truth

Training data

136,289

clips in the training index, one sample each; 57,738 released with video

Weights

4

LoRAs: control and four-step, for full-clip and for causal streaming rendering

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}
}