Lightx2v releases RL-refined Qwen-Image-2.1 Lightning v2: 4-step generation and editing
linoy_tsaban · x · 2026-11-02
Lightx2v dropped v2 of its Qwen-Image-2.1 Lightning checkpoint on Hugging Face, an RL-refined (LightX2V RL) distillation supporting text-to-image, single-reference editing, and multi-reference generation in just 4 denoising steps at CFG=1.0. Weights ship as a single safetensors file (FP32, runnable in BF16); it's a full transformer replacement (not a LoRA), loaded with the original Qwen-Image-2.1 pipeline via a provided Diffusers quick-start and inferenceconfig.json.
More from Multimodal
- Midnight Run: an underground racing short film made with Seedance 2.5 — Hot-Decision9790 · 2026-11-02
- Tencent ships three speech models: ASR, TTS and real-time voice-preserving translation — solyarisoftware · 2026-11-02
- Sonic 2 VFX animator shows hybrid Maya + AI rendering pipeline — KarimHann · 2026-11-02
- Fibo 2.0 image editing ships a 2-step distilled version on Hugging Face — Illustrious_Row_9971 · 2026-11-02
- 4B Qwen3.5 fine-tune generates game-ready 3D assets with one-click GLB export — linoy_tsaban · 2026-11-02
- Krea2 character training: swapping LoRA for LoKR tanks results, author seeks fix — AkaToraX · 2026-11-02