Train rectified-flow diffusion models using reward-free self-correction to fix denoising trajectory errors.
Post-training for diffusion models that fixes exposure bias without reward models or preference labels.
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.
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.
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.
Supervises the denoiser at the exact timestep errors occur by re-noising off-trajectory states toward the original noise endpoint.
Provides per-timestep geometric correction targets anchored to clean latents without using reward models or negative samples.
Generates auxiliary states from the active model's own rollouts so the training distribution evolves alongside model parameters.
Integrates local evaluation setups for GenEval, PickScore, CLIPScore, HPSv2.1, and PaddleOCR.
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.
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.
No, HY-SOAR calculates analytical geometric correction targets directly from generated states and clean latents without external preference labels.
Yes, it extends flow-matching objectives and can replace standard SFT prior to applying downstream RL alignment like Flow-GRPO.
The provided training scripts default to Stable Diffusion 3.5 Medium (stabilityai/stable-diffusion-3.5-medium).
See Zekai first in Google