Qwen-Image 2512 / Edit 2511 — NF4 for LoRA training
Pre-quantized weights used by the Qwen-Image 2512 trainer of AcademiaSD LoRAlab Trainer Studio. The trainer downloads only the folders it needs; you don't need to download anything by hand.
| Folder | Content |
|---|---|
transformer/ |
Qwen-Image-2512 transformer, NF4 (block 0 img_mod kept in BF16) |
transformer_edit/ |
Qwen-Image-Edit-2511 transformer, NF4 (same layout) |
text_encoder_NF4/, text_encoder_INT8/, text_encoder_BF16/ |
Qwen2.5-VL-7B text encoder in three precisions (used only by Pre-Cache) |
vae/, scheduler/, tokenizer/, processor/ |
Unchanged from the original repositories |
LoRAs/ |
Lightning 4-step LoRAs by lightx2v, used for fast training previews |
These files are meant for training with bitsandbytes; for inference use the original models.
Credits and license
- Qwen/Qwen-Image-2512 and Qwen/Qwen-Image-Edit-2511 by the Qwen team — Apache 2.0.
- lightx2v/Qwen-Image-2512-Lightning and lightx2v/Qwen-Image-Edit-2511-Lightning — Apache 2.0.
NF4 conversion by AcademiaSD. Same license as the original models (Apache 2.0).