Skip to content
Bud's Wiki

Qwen-Image-Edit-2511-Lightning — Step-Distilled 4-Step LoRA for Qwen-Image-Edit-2511

TL;DR

Section titled “TL;DR”

Qwen-Image-Edit-2511-Lightning is LightX2V’s day-0 step-distilled release on top of Qwen/Qwen-Image-Edit-2511, shipping a 4-step LoRA (BF16 and FP32) and a fused FP8 (e4m3fn scaled) checkpoint that bakes the distilled LoRA into the base weights. The distillation collapses the upstream 40-step recipe to 4 steps (~10× DiT speedup), and combined with LightX2V’s inference stack the team reports a 25× reduction in DiT NFEs and a ~42.55× end-to-end speedup with native support for NVIDIA, Hygon, Metax, Ascend, and Cambricon. FP8 quantization cuts memory roughly in half versus FP32 while preserving editing fidelity.

Key claims

Section titled “Key claims”

Method

Section titled “Method”

The release ships three artifact types over the same Qwen-Image-Edit-2511 base. Two are 4-step distilled LoRAs (BF16 and FP32 precision variants) that the user loads on top of the upstream pipeline via pipe.load_lora_weights(...). The third is a single-file FP8 (e4m3fn scaled) checkpoint with the distilled LoRA pre-fused into the base weights, intended for low-memory single-file loaders (ComfyUI, Diffusers single-file, Draw Things). The card frames the work as a combination of two orthogonal techniques: step distillation (training the LoRA to replicate the multi-step teacher trajectory in 4 steps) and FP8 quantization of the base. Usage is described against two host frameworks: the Qwen-Image-Lightning repo (the distillation pipeline itself) and LightX2V (the production inference engine reporting the headline 42.55× speedup).

Results

Section titled “Results”
  • 4 inference steps vs. 40 baseline → ~10× DiT-only speedup [model card §Key Optimizations].
  • FP8 fused variant: ~50% memory reduction vs. FP32 with “maintained editing fidelity” — no quantitative quality benchmark is provided on the card itself [model card §Key Optimizations].
  • The upstream repo claims 25× NFE reduction and 42.55× end-to-end wall-clock speedup with full LightX2V integration; no FID, CLIP, or EditReward numbers are given on the model card or in the repo announcement Qwen-Image — Image foundation model with text rendering and editing (QwenLM/Qwen-Image) News 2025.12.23.
  • 317,842 downloads in the month following release — a useful proxy for community adoption [model card §Downloads].

Why it’s interesting

Section titled “Why it’s interesting”

This is the canonical few-step-distill-of-a-frontier-edit-model recipe applied to Qwen-Image-Edit-2511, the same template that produced Eigen-Banana-Qwen-Image-Edit: Lightning-Fast Instruction-Based Image Editing with Pico-Banana-400K (Lightning distillation of Qwen-Image-Edit on Pico-Banana-400K) and that the prior Qwen-Image-Edit-2509 — multi-image editing and enhanced consistency (Qwen) release also attracted. The fact that the distill ships Day-0 alongside the base Qwen-Image Technical Report release, with cross-vendor accelerator support, signals that “frontier open image editor + few-step LoRA + FP8 fuse” is now the assumed release shape rather than a community follow-on — contrasts with Qwen-Edit-2509-Upscale-LoRA — community photo restoration adapter for Qwen-Image-Edit-2509 and Qwen-Edit-2509-Multiple-angles LoRA — community camera-control adapter for Qwen-Image-Edit-2509 which arrived weeks/months after the 2509 base. For distillation-method comparison see pi-Flow: Policy-Based Few-Step Generation via Imitation Distillation and FreeFlow: Flow Map Distillation Without Data which target the same 4-step regime with different training signals.

See also

Section titled “See also”