Kalaido-qwenedit-lora — Reinforcement learning enhanced Qwen-Image-Edit-2511
Introduction
Kalaido-qwenedit-lora is a LoRA adapter for Qwen-Image-Edit-2511, trained with reinforcement learning. It is intended for general editing and has better background preservation and edit instruction following than the base model.
Compared to the base Qwen-Image-Edit-2511, the adapter improves:
- Better background preservation: No modifications to the unedited regions in the image.
- Better instruction following: The model has better prompt following and understanding than the base model.
Example usage
Install a recent diffusers build (Qwen edit pipelines evolve quickly):
pip install git+https://github.com/huggingface/diffusers
Weights in this repo are shipped as a single Safetensors file: qwen-2511_general_edit_lora_v1.safetensors.
import torch
from PIL import Image
from diffusers import QwenImageEditPlusPipeline
model_id = "Qwen/Qwen-Image-Edit-2511"
lora_id = "FractalAIResearch/Kalaido-qwenedit-lora"
pipe = QwenImageEditPlusPipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
pipe.load_lora_weights(
lora_id,
weight_name="qwen-2511_general_edit_lora_v1.safetensors",
adapter_name="kalaido",
)
pipe.set_adapters(["kalaido"], [1.0])
pipe.to("cuda")
image = Image.open("input.png").convert("RGB")
prompt = "Your edit instruction here."
negative_prompt = " "
out = pipe(
image=image,
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=40,
true_cfg_scale=4.0,
guidance_scale=1.0,
generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
out.save("edited.png")
Evaluation
We report ImgEdit benchmark scores.
Kalaido-qwenedit-lora is the open release in this repository (FractalAIResearch/Kalaido-qwenedit-lora); Kalaido-edit is our stronger production editor. For our latest closed models, APIs, and demos, see kalaido.ai.
In the table below, the highest score in each row is bold (ties are bolded for both).
| Task | GPT-4o Image | Qwen Image Edit | Google Nano Banana Pro | GPT-5.1 | FLUX.1 Kontext | Kalaido-qwenedit-lora | Kalaido-edit |
|---|---|---|---|---|---|---|---|
| Addition | 4.65 | 4.61 | 4.70 | 4.70 | 4.08 | 4.65 | 4.71 |
| Remove | 3.81 | 4.39 | 4.69 | 4.89 | 3.23 | 4.50 | 4.72 |
| Replace | 4.49 | 4.70 | 4.55 | 4.83 | 4.42 | 4.70 | 4.77 |
| Attribute alter | 4.26 | 4.62 | 4.70 | 4.66 | 3.96 | 4.64 | 4.71 |
| Motion change | 4.76 | 4.82 | 4.83 | 4.92 | 4.54 | 4.80 | 4.92 |
| Style transfer | 4.75 | 4.77 | 4.87 | 4.91 | 4.59 | 4.73 | 4.84 |
| Background change | 4.62 | 4.54 | 4.50 | 4.90 | 4.15 | 4.53 | 4.68 |
| Object extraction | 2.96 | 4.19 | 3.57 | 2.91 | 2.27 | 4.21 | 4.40 |
| Hybrid edit | 4.54 | 4.03 | 3.77 | 4.71 | 3.42 | 4.37 | 4.64 |
| Average score | 4.32 | 4.52 | 4.46 | 4.60 | 3.85 | 4.57 | 4.71 |
Qualitative comparison
Side-by-side examples on ImgEdit-style general edits: input (left), Qwen-Image-Edit-2511 baseline (center), Ours (right).
License
Kalaido-qwenedit-lora is licensed under the Apache License 2.0.