ModelTC/LightX2V-Qwen-Image-Lightning

📌 What is Qwen‑Image‑Lightning?

Qwen‑Image‑Lightning is a set of distilled (i.e., faster, lighter) checkpoint files for the Qwen‑Image text‑to‑image model released by the QwenLM team. The “Lightning” models keep the original’s ability to render complex Chinese/English text inside images, but they require far fewer diffusion steps (4 or 8 steps instead of the 100‑step baseline), giving 12–25× speed‑up with only modest quality loss.

---

🔧 Main components

Component	What it is	Typical use
Lightning checkpoints	.safetensors files (fp32, bf16, fp8) that contain the distilled weights for Qwen‑Image (base) and Qwen‑Image‑Edit (editing).	Load with 🤗 Diffusers, ComfyUI, or any pipeline that supports Qwen‑Image.
LoRA weights	Low‑rank adapters (“Lightning LoRAs”) that can be applied on top of the base model to improve quality or adapt to FP8 bases.	Fine‑tune or combine with existing LoRAs for better visual fidelity.
ComfyUI workflows	JSON files that describe a ready‑to‑run graph in the ComfyUI visual editor.	One‑click generation or editing inside ComfyUI.
Evaluation scripts	Example scripts (e.g., in the Nunchaku repo) showing how to run the model with cache‑acceleration or 4‑bit quantisation.	Benchmark speed/quality or reduce GPU memory.

---

🚀 Why use the Lightning version?

• Speed – 4‑step models run in a fraction of the time of the original 100‑step diffusion.
• Memory‑friendly – FP8 / 4‑bit variants let you run on a single consumer‑grade GPU.
• Compatibility – Works with the official 🤗 Diffusers QwenImagePipeline, ComfyUI, and third‑party tools like Nunchaku and Cache‑dit.
• Quality‑focused updates – V2.0 improves over‑saturation and skin texture compared to V1.x.

---

📂 How to get started

1. Pick a checkpoint – Choose the step count (4 or 8) and precision you need (fp32, bf16, fp8). Links are in the Latest News section.
2. Load with Diffusers

   from diffusers import QwenImagePipeline
   pipe = QwenImagePipeline.from_pretrained("lightx2v/Qwen-Image-Lightning-4steps-V2.0")
   image = pipe(prompt="...", num_inference_steps=4).images[0]
   

3. Optional LoRA – If you are using the FP8 base, download the matching Lightning LoRA (see the “Using Lightning LoRAs with FP8 Models” table) and apply it with pipe.unet.add_adapter().
4. ComfyUI – Open the provided JSON workflow (workflows/...json) to run the model visually.

---

📊 Reported performance (from the README)

Model	Steps	Speed‑up vs. base	Typical quality notes
Qwen‑Image‑Lightning‑8steps‑V1.1	8	~12× faster	Good overall, but hair‑like details may be slightly blurred.
Qwen‑Image‑Lightning‑4steps‑V1.0	4	~25× faster	Very fast; may lose fine text rendering in dense scenes.
V2.0 (both 4‑ and 8‑step)	4 / 8	similar speed	Reduced over‑saturation, better skin tones.

The README also notes that for very dense or small text the original 100‑step model still performs best, and that extremely complex scenes can challenge all versions.

---

🛠️ Known issues & fixes

• Grid artifacts when using the original FP8 base with LoRAs trained on BF16. The repo supplies two remedies:
  1. Use the newly‑distilled Lightning LoRA that was trained on the FP8 base.
  2. Switch to the scaled FP8 base weight (qwen_image_fp8_e4m3fn_scaled.safetensors) which aligns better with existing LoRAs.
• Compatibility with the official Qwen‑Image FP8 model was fixed in Oct 2025 (see Issue #32).

---

📚 Where to learn more

• Diffusers docs – https://huggingface.co/docs/diffusers/main/api/pipelines/qwenimage
• ComfyUI tutorials – https://docs.comfy.org/tutorials/image/qwen/qwen-image
• Nunchaku example script – https://github.com/nunchaku-tech/nunchaku/blob/main/examples/v1/qwen-image-lightning.py
• Cache‑dit 3.5‑step inference – https://github.com/vipshop/cache-dit/blob/main/examples/pipeline/run_qwen_image_lightning.py

---

🎯 TL;DR

Qwen‑Image‑Lightning provides fast, low‑step distilled checkpoints (4 / 8 steps) for the Qwen‑Image text‑to‑image model, with multiple precision formats and ready‑to‑use LoRA adapters. It integrates with the major diffusion toolkits (🤗 Diffusers, ComfyUI) and offers a clear upgrade path from the original model while keeping the standout ability to render complex Chinese/English text.