Tencent-Hunyuan

Tencent-Hunyuan / SAGE-GRPO

Public

Official Implementation of SAGE-GRPO:Manifold-Aware Exploration for Reinforcement Learning in Video Generation

dungeonmassster.github.ioSAGE-GRPO-Page
84
1
100% credibility
Found Mar 30, 2026 at 85 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

SAGE-GRPO is a research framework for fine-tuning video generation models like HunyuanVideo-1.5 to produce higher-quality, better-aligned outputs using specialized optimization techniques.

How It Works

1
🔍 Discover smarter video AI

You find a helpful tool that trains video generators to create exactly what you imagine, like making animations more lifelike.

2
📥 Get the starting pieces

Download ready-made video models and simple guides to begin improving them.

3
🛠️ Set up your training space

Connect computers and prepare sample videos with descriptions of what makes them great.

4
🚀 Start the magic training

Hit go and watch the tool learn from your examples, blending them into better video skills.

5
📹 Test new creations

Generate videos using your trained model and see smoother, more accurate results.

✨ Videos come alive perfectly

Enjoy videos that match your vision exactly, feeling professional and delightful every time.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 85 to 84 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is SAGE-GRPO?

SAGE-GRPO is the official GitHub repository delivering a Python-based post-training framework for aligning video generation models like HunyuanVideo-1.5 using GRPO reinforcement learning. It tackles instability in video RL by enabling manifold-aware exploration during generation, ensuring noise stays on the data manifold for reliable reward signals and higher-quality outputs. Developers run provided shell scripts for multi-node training or single-GPU inference, producing aligned models that excel in visual metrics like CLIPScore.

Why is it gaining traction?

Unlike DanceGRPO or FlowGRPO, it uses precise SDE exploration with logarithmic corrections and dual trust-region KL to prevent drift, yielding stabler training and superior videos—check the gallery comparisons beating baselines at 20/40 steps. Gradient norm equalization balances updates across timesteps, while FSDP and sequence parallelism scale seamlessly to 64 GPUs. Early adopters praise the plug-and-play setup on PyTorch and Diffusers for quick RLHF-style video alignment.

Who should use this?

ML engineers at AI labs fine-tuning diffusion models for text-to-video or image-to-video apps, especially those chasing better motion coherence and prompt adherence. Video generation researchers experimenting with reward models like VideoAlign on multi-GPU clusters. Teams extending HunyuanVideo who need production-ready post-training without rewriting distributed pipelines.

Verdict

Solid official implementation for manifold-aware RL exploration in video generation, with strong docs, visuals, and 64-GPU defaults—but 84 stars and 1.0% credibility score signal it's nascent; test on small setups first before scaling. Grab it if you're in video gen and have the hardware.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.