AI tools, scored for your job
Get paid to train AI
🤝 You're in the Free & Open Source library — community-built tools you can use without paying. A different concept from our scored AI directory.
Home › Free & Open Source › Cybersecurity › HY-SOAR
✦ Free & open sourceCybersecurity

HY-SOAR

Train rectified-flow diffusion models using reward-free self-correction to fix denoising trajectory errors.

Self-Correction Alignment for Diffusion Models

Post-training for diffusion models that fixes exposure bias without reward models or preference labels.

Free forever · source-available · Licence not stated in the repository
✓ Needs IT✓ Browser-based✓ Real-time collaboration
5.9
ZEKAI
Advanced flow-model training
Free & open source · Cybersecurity
Ease of setup2
Usefulness8
Maintenance7
Community6.5
★ 752
GITHUB
Apr 2026
UPDATED
Licence not stated in the repository
LICENSE
🛠 Don't want to self-host? We can set it up →
⚡ Quick answer

HY-SOAR is an open-source post-training framework developed by Tencent-Hunyuan for rectified-flow diffusion models like Stable Diffusion 3.5. You should use it if you want to eliminate trajectory exposure bias during generation without relying on terminal reward models or labeled preference pairs. It serves machine learning engineers and generative AI researchers fine-tuning diffusion pipelines.

Runs on
Cross-platform
Hosting
Self-host
Languages
Python
License
Licence not stated in the repository

What it is

The tool samples on-trajectory states from a diffusion model rollout, re-noises off-trajectory states toward a noise endpoint, and calculates analytical correction velocities. This creates an on-policy, per-timestep correction signal that subsumes standard SFT. It provides training scripts, reward evaluation code, and DeepSpeed integration for multi-GPU hardware.

Who it's for

It suits AI researchers and ML engineers post-training flow-matching text-to-image models for task-specific refinement. It does not suit non-technical users or teams looking for a turnkey visual generator, as it requires multi-GPU infrastructure and deep PyTorch expertise.

Key features

Exposure-Bias Trajectory Correction

Supervises the denoiser at the exact timestep errors occur by re-noising off-trajectory states toward the original noise endpoint.

Reward-Free Dense Objectives

Provides per-timestep geometric correction targets anchored to clean latents without using reward models or negative samples.

On-Policy Training Signal

Generates auxiliary states from the active model's own rollouts so the training distribution evolves alongside model parameters.

Multi-Metric Evaluation Suite

Integrates local evaluation setups for GenEval, PickScore, CLIPScore, HPSv2.1, and PaddleOCR.

How to start

Up and running in 3 steps

1
Clone the repository and build the Python 3.10 Conda environment with PyTorch 2.6 and CUDA 12.6.
2
Download reward evaluation checkpoints and compile native CUDA dependencies for MMCV and MMDetection.
3
Authenticate via huggingface-cli to access Stable Diffusion 3.5 Medium and launch sora.train_soar_sd3_5m with torchrun.
Real talk

Our honest verdict

What's great

  • Fixes compounding denoising errors without needing preference datasets or reward model training.
  • Acts as a drop-in replacement or pre-stage for standard Supervised Fine-Tuning (SFT).
  • Includes pre-configured DeepSpeed ZeRO-2 launch scripts for distributed training.
  • Provides simplified local evaluation pipelines for benchmark metrics like GenEval and HPSv2.1.

Where it falls short

  • Requires significant GPU compute, with default training configured for an 8-GPU node.
  • Demands complex environment setups including custom C++ builds for MMCV and CUDA 12.6.
  • No official license is provided in the repository metadata, leaving legal usage terms unclear.
Cost

Pricing & license

HY-SOAR is free to download from GitHub, but the repository does not specify an open-source license. The only operational cost comes from provisioning the high-performance GPU hardware needed for distributed model training.

Paid work

Open remote AI-training roles

880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.

See all open roles →
Compare

Alternatives

🎁 Free & open-source alternatives
Get more

Use it, or have us set it up

📄 Free PDF: HY-SOAR for Cybersecurity
A plain-language guide to getting your first result. Sent to your inbox.
🔔 64 people want a full video tutorial — add your vote
🛠 Don't want to set it up yourself?
We'll install, host & configure HY-SOAR for you — and train your team.
Get it done for you ↗
Questions

FAQ

Does HY-SOAR require preference labels or human feedback?

No, HY-SOAR calculates analytical geometric correction targets directly from generated states and clean latents without external preference labels.

Can HY-SOAR be used alongside Reinforcement Learning stages?

Yes, it extends flow-matching objectives and can replace standard SFT prior to applying downstream RL alignment like Flow-GRPO.

What base model does HY-SOAR train out of the box?

The provided training scripts default to Stable Diffusion 3.5 Medium (stabilityai/stable-diffusion-3.5-medium).

Radar

Latest AI news

✉️ The weekly free toolkit
The best new free & open-source tools for Cybersecurity — one short email a week, in your language.
Free · 1 email/week · unsubscribe anytime
Get paid to train AI →

See Zekai first in Google