News | Quick Start | Key Features | Documentation
- [2026/08] 🎉 Miles-Diffusion Release: RL post-training for diffusion models is here — Flow-GRPO, DiffusionNFT, and SFT under one trainer, with end-to-end validated recipes for SD3.5, Qwen-Image, Wan2.2-T2V-A14B, LTX-2.3, and the Cosmos3 MoT omni family. [Docs]
Miles-diffusion is currently a standalone repository built on Miles' design philosophy, focused on RL post-training for image and video diffusion models. SGLang-diffusion serves the rollout, and the DiT trains under FSDP2 on a backend that co-evolves with Miles' own. Models load from a diffusers pipeline, or from a native package when a family brings its own modeling. Shipped recipes carry explicit verification levels. Custom rewards, losses, and rollout functions plug in through flags.
- Verified recipes for the latest diffusion models. Launchers for Wan2.2-T2V-A14B, Qwen-Image, LTX-2.3, the Cosmos3 MoT omni family, and SD3.5.
TrainPipelineConfigallows for easy model support. - Quality control on three fronts. Deterministic mode supports bit-for-bit comparisons for recipes covered by committed E2E standards; SGLang-side monkey patches reduce train/rollout mismatches; and an FSDP2 param-dtype patch provides per-parameter fp32 control under the mixed-precision policy. See Deterministic Training and Dtype Control.
- SFT, DiffusionNFT, and Flow-GRPO under one trainer. The loss type, training-batch preparation, rollout function, and reward function are all replaceable components, so integrating a new algorithm — or swapping in your own customized component — is easy.
- SGLang native. Rollout runs on the inference engine itself — the SGLang-diffusion serving stack — with RL support and optimizations living engine-side. An optional curated set of monkey patches aligns selected engine operations with the training-side forward.
- Multiple parallelisms. The rollout engines scale with tensor and sequence parallelism to support large models and very long contexts; training scales with USP (Ulysses × Ring), built from each family's diffusers
_cp_plan— or a self-written one — for agile model integration. - LoRA training support. With
--lora-ipc-weight-sync, PEFT LoRA on the FSDP2 actor ships onlylora_A/lora_Bpairs to colocated rollout engines over CUDA IPC and merges them engine-side. See LoRA Training and Weight Sync.
Each model links to its recipe page; every documented recipe is labeled with a recipe verification level.
| Model | Task | Canonical Recipes |
|---|---|---|
| Stable Diffusion 3.5 | T2I | Flow-GRPO + OCR, DiffusionNFT + PickScore |
| Qwen-Image | T2I | Flow-GRPO + PickScore (flow_grpo-aligned) |
| Wan2.2-T2V-A14B | T2V | Flow-GRPO + PickScore, LoRA SFT |
| LTX-2.3 | T2V | Flow-GRPO + PickScore |
| Cosmos3 (Edge / Nano / Super) | T2I | Flow-GRPO + PickScore |
| MiniMax H3 | T2VA | Flow-GRPO + PickScore |
Miles-diffusion stands on the shoulders of giants and thanks the following repositories for their outstanding work: Miles · slime · SGLang · diffusers · VeOmni · Flow-GRPO · DiffusionNFT · Flow-Factory
- GitHub: https://github.com/radixark/miles_diffusion
- Miles (LLM RL): https://github.com/radixark/miles
- Documentation: https://miles.radixark.com/docs/diffusion
From noise, a world takes shape — one step at a time.