LightX2V releases 4-step Qwen-Image-2.1 Lightning LoRA, trained on 16 Huawei Ascend 910B NPUs
RisingSayak · x · 2026-11-10
LightX2V shipped a Lightning distilled LoRA for Qwen-Image-2.1 that cuts text-to-image generation and image editing down to just 4 inference steps — already at 435k downloads on Hugging Face.
The model card includes a full Diffusers snippet covering the scheduler config (dynamic shifting with exponential shift type) and 4-step sigmas settings. Training was done entirely on 16 Huawei Ascend 910B NPUs using DMD. Note: latents stay FP32 during training and LightX2V inference, so BF16 Diffusers runs may differ numerically.
The author also teased a Krea2 Turbo Distill 4-step LoRA, distilled further from Krea's 8-step version.
More from Multimodal
- Is ControlNet Obsolete? New DiT Image Models Ship Without It — Manicarus · 2026-11-10
- Full prompt revealed: how to recreate the Van Gogh convenience store in MiniMax H3 Image — mhdfaran · 2026-11-10
- 10 open-source GitHub repos for voice AI: from local ASR to zero-shot voice cloning — Shruti_0810 · 2026-11-10
- Fable 5.2 generates animated nature scene with ~70k trees from one prompt — Merzmensch · 2026-11-10
- MOSS-Image roundup: Control LoRAs, 461 new characters, and 20-minute LoRA training on a 5090 — optimisticalish · 2026-11-10
- StreamEval from Tsinghua, NVIDIA, CMU and Berkeley exposes gaps in streaming video generation across 38 models — jiqizhixin · 2026-11-10