AlayaRenderer — Generative World Renderer

Skill by ara.so — Daily 2026 Skills collection.

AlayaRenderer is a two-stage framework for high-quality video rendering:

1. Inverse Renderer (RGB → G-buffers): Extracts albedo, normal, depth, roughness, and metallic maps from RGB video using a fine-tuned Cosmos-Transfer1-DiffusionRenderer 7B model.
2. Game Editing (G-buffers + Text → Stylized RGB): Synthesizes photorealistic, stylized RGB video from G-buffer inputs using a fine-tuned Wan2.1 1.3B model via DiffSynth-Studio.

Installation

Clone the Repository

git clone --recurse-submodules https://github.com/ShandaAI/AlayaRenderer.git
cd AlayaRenderer

Important: Use --recurse-submodules — DiffSynth-Studio is a git submodule required for Game Editing.

Two Separate Conda Environments (Recommended)

The two models have conflicting dependencies. Use separate environments:

# Environment 1: Inverse Renderer
conda create -n inverse_renderer python=3.10 -y
conda activate inverse_renderer
cd inverse_renderer
# Follow inverse_renderer/ instructions for Cosmos-Transfer1 setup

# Environment 2: Game Editing
conda create -n game_editing python=3.10 -y
conda activate game_editing
cd game_editing
# Follow DiffSynth-Studio setup instructions

Model Weights

Download and Place Weights

# Inverse Renderer — replace the base checkpoint
huggingface-cli download Brian9999/world_inverse_renderer \
  --local-dir inverse_renderer/checkpoints/Diffusion_Renderer_Inverse_Cosmos_7B

# Game Editing — place in game_editing models directory
mkdir -p game_editing/models/train/Wan2.1-T2V-1.3B_gbuffer
huggingface-cli download Brian9999/stylerenderer \
  --local-dir game_editing/models/train/Wan2.1-T2V-1.3B_gbuffer

Inverse Renderer Usage

The inverse renderer decomposes an RGB video into 5 G-buffer channels: albedo, normal, depth, roughness, metallic.

Setup

cd inverse_renderer
# Follow Cosmos-Transfer1-DiffusionRenderer environment setup
# Ensure checkpoint is at:
# inverse_renderer/checkpoints/Diffusion_Renderer_Inverse_Cosmos_7B/

Inference

Refer to the inverse_renderer/ subdirectory for the full inference script. The general pattern follows Cosmos-Transfer1-DiffusionRenderer conventions:

# inverse_renderer/run_inverse.py (typical pattern)
import torch
from pathlib import Path

# Input: path to RGB video
input_video = "path/to/rgb_video.mp4"
output_dir = "outputs/gbuffers/"

# The model outputs 5 synchronized channels:
# - albedo (diffuse color)
# - normal (surface orientation)
# - depth (scene geometry)
# - roughness (surface roughness)
# - metallic (metallic property)

Game Editing Usage

Quick Start — CLI Inference

cd game_editing

CUDA_VISIBLE_DEVICES=0 python \
    examples/wanvideo/model_inference/inference_gbuffer_caption.py \
    --checkpoint models/train/Wan2.1-T2V-1.3B_gbuffer/model.safetensors \
    --gpu 0 \
    --style snowy_winter \
    --prompt "the scene is set in a frozen, snow-covered environment under cold, pale winter light with falling snowflakes, creating a silent and ethereal winter wonderland atmosphere." \
    --gbuffer_dir test_dataset \
    --save_dir outputs/ \
    --num_frames 81 \
    --height 480 \
    --width 832

CLI Parameters

G-buffer Directory Structure

test_dataset/
├── albedo/
│   ├── frame_0000.png
│   ├── frame_0001.png
│   └── ...
├── normal/
│   ├── frame_0000.png
│   └── ...
├── depth/
│   ├── frame_0000.png
│   └── ...
├── roughness/
│   ├── frame_0000.png
│   └── ...
└── metallic/
    ├── frame_0000.png
    └── ...

Style Prompt Examples

# Cyberpunk night scene
--style night \
--prompt "neon-lit urban environment at night with rain-slicked streets reflecting colorful neon signs, creating a cyberpunk noir atmosphere"

# Golden hour / sunset
--style sunset \
--prompt "warm golden hour lighting with long shadows and a glowing amber sky, soft cinematic atmosphere"

# Rainy urban
--style rainy \
--prompt "overcast rainy day with wet surfaces, soft diffuse lighting, and atmospheric fog creating a moody cinematic look"

# Fantasy / stylized
--style fantasy \
--prompt "magical forest environment with bioluminescent plants, ethereal blue-green lighting, and mystical particle effects"

# Foggy morning
--style foggy \
--prompt "early morning dense fog with soft diffused light creating a mysterious and quiet atmosphere"

Multi-GPU Inference

# Run on specific GPU
CUDA_VISIBLE_DEVICES=1 python \
    examples/wanvideo/model_inference/inference_gbuffer_caption.py \
    --checkpoint models/train/Wan2.1-T2V-1.3B_gbuffer/model.safetensors \
    --gpu 1 \
    --style rainy \
    --prompt "heavy rainfall with dark storm clouds and dramatic lightning in the distance" \
    --gbuffer_dir my_gbuffers \
    --save_dir outputs/rainy_scene \
    --num_frames 81 --height 480 --width 832

Full Pipeline: RGB Video → Stylized Output

# Step 1: Extract G-buffers from RGB video (Inverse Renderer env)
conda activate inverse_renderer
cd inverse_renderer
python run_inverse.py \
    --input path/to/gameplay_video.mp4 \
    --output_dir ../game_editing/test_dataset/

# Step 2: Apply game editing style (Game Editing env)
conda activate game_editing
cd ../game_editing
CUDA_VISIBLE_DEVICES=0 python \
    examples/wanvideo/model_inference/inference_gbuffer_caption.py \
    --checkpoint models/train/Wan2.1-T2V-1.3B_gbuffer/model.safetensors \
    --gpu 0 \
    --style snowy_winter \
    --prompt "frozen tundra with blizzard conditions, pale blue-white lighting and drifting snow" \
    --gbuffer_dir test_dataset \
    --save_dir outputs/final_render \
    --num_frames 81 --height 480 --width 832

Online Demos

Dataset Overview

The AlayaRenderer dataset (release pending) features:

• 4M+ frames at 720p / 30 FPS
• 6 synchronized channels: RGB + albedo, normal, depth, metallic, roughness
• 40 hours from Cyberpunk 2077 and Black Myth: Wukong
• Average clip length: 8 minutes, up to 53 minutes continuous
• Weather variants: sunny, rainy, foggy, night, sunset
• Motion blur variant via sub-frame interpolation

Architecture Summary

RGB Video Input
      │
      ▼
┌─────────────────────────────────────┐
│  Inverse Renderer                   │
│  (Cosmos-Transfer1 7B fine-tuned)   │
│  RGB → [albedo, normal, depth,      │
│          roughness, metallic]       │
└─────────────────┬───────────────────┘
                  │  G-buffers
                  ▼
┌─────────────────────────────────────┐
│  Game Editing                       │
│  (Wan2.1 1.3B fine-tuned)           │
│  G-buffers + Text Prompt            │
│  → Stylized RGB Video               │
└─────────────────────────────────────┘

Troubleshooting

Submodule not found / DiffSynth-Studio missing

# If cloned without --recurse-submodules:
git submodule update --init --recursive

CUDA Out of Memory

• Reduce --num_frames (try 41 instead of 81)
• Reduce resolution: --height 320 --width 576
• Ensure no other processes are using the GPU: CUDA_VISIBLE_DEVICES=0

num_frames must follow 8n+1 pattern

Valid values: 9, 17, 25, 33, 41, 49, 57, 65, 73, 81

# Valid
--num_frames 81   # 8*10 + 1 ✓
--num_frames 41   # 8*5 + 1  ✓

# Invalid
--num_frames 80   # ✗
--num_frames 60   # ✗

Checkpoint not found

# Verify checkpoint placement
ls game_editing/models/train/Wan2.1-T2V-1.3B_gbuffer/model.safetensors
ls inverse_renderer/checkpoints/Diffusion_Renderer_Inverse_Cosmos_7B/

Version conflicts between models

Always use the two separate conda environments (inverse_renderer and game_editing). Do not install both models' dependencies in one environment.

Citation

@article{huang2026generativeworldrenderer,
    title={Generative World Renderer},
    author={Zheng-Hui Huang and Zhixiang Wang and Jiaming Tan and Ruihan Yu and Yidan Zhang and Bo Zheng and Yu-Lun Liu and Yung-Yu Chuang and Kaipeng Zhang},
    journal={arXiv preprint arXiv:2604.02329},
    year={2026}
}