🌟 Qwen Image Edit Remix
Qwen Image Edit Remix is a high-performance Qwen-based model designed for Image Editing, Image-to-Image, and Text-to-Image tasks.
It focuses on stability, speed, and subject consistency, while still allowing flexible and creative remix-style generation.
The model runs in FP8 precision and includes acceleration LoRA, significantly improving inference speed and reducing VRAM usage without sacrificing output quality.
This model supports NSFW content. Please ensure responsible and lawful usage.
📦 Model Variants
🔹AIO v2.0 (All-In-One)
The AIO version comes with baked-in CLIP and VAE—ready to use right out of the box.
Installation: Download and place the model file into your
models\checkpointsfolder.Usage: Simply use the
Load Checkpointnode in ComfyUI to load the model.
🔹 Standard Version (without VAE / CLIP)
Contains only the core model weights
Requires users to load their own VAE and CLIP
Recommended for advanced users with existing pipelines or custom components
⚠️ Aside from the inclusion of VAE and CLIP, both versions are identical in structure, performance, and output quality
⚠️ Both versions run in FP8 precision and include the same acceleration LoRA
🎉 AIO v2.0 Update Notes
This v2.0 release brings several major upgrades to visual quality and control:
🧍 Enhanced Human Pose Accuracy: Significantly improves skeletal structure in complex dynamic poses. Limbs are generated much more naturally, bidding farewell to awkward anatomy.
🧑🤝🧑 Reduced Distortion in Multi-Person Scenes: Specially optimized for multi-subject interactions. Effectively minimizes limb blending, dislocations, and abnormal limb counts when generating multiple people.
🎯 Increased Prompt Sensitivity: The model now understands and responds to your prompts much more precisely, keenly capturing and reproducing the specific details and styles you ask for.
✨ Core Capabilities
Image Editing
Precise instruction-based editing of input images, including character, clothing, background, style, and detail adjustments.Image-to-Image (I2I)
Redraw, enhance, or stylize images while preserving the original composition and subject structure.Text-to-Image (T2I)
Generate images purely from text prompts without requiring any input image.Remix-oriented Generation
Designed for re-creation rather than full regeneration, maintaining key visual elements while introducing new creative variations.Efficient Inference
FP8 + acceleration LoRA provides a strong balance between speed, VRAM efficiency, and visual quality.
⚙️ Recommended Settings
Sampler
euler_ancestralScheduler
beta
This combination offers a good balance between stability, detail preservation, and overall visual coherence, especially for image editing and remix workflows.
🎯 Use Cases
AI image editing and retouching
Image-to-image redraw and style transfer
Text-to-image content creation
Outfit, pose, and scene modification
Character-consistent remix and iteration
Posters, covers, and visual concept design
ComfyUI / Diffusers image generation workflows
Description
FAQ
Comments (21)
等待standard版（搓手）
Fantastic release, this works better than the QWEN Rapid AIO v23 imho
V23 is the only qie2511 I ever use.. guess it's time to upgrade, hope the hype is real.
I tried out QIE a few months back and for whatever reason I didn't get the results I wanted. But I just got bored with F2K and tried this model. Wow! It does nearly anything I ask of it. Kind of crazy really. I can just give it 2 pictures of random people and prompt "These two people are wearing USPS uniforms and playing hacky-sack in front of a burning building" and I get EXACTLY that.
From me, I tried this one. it handles Chinese better than English.
can I use Qwen 2512 lora with this?
大佬，2.0的人脸一致性有问题。如果表情有变化，那么人脸的黑痣数量会有增加，原本两颗痣，结果里能多出来三四颗。
in my opinion very low image detail, plastic skin, worse than rapid aio v23, but slightly better face preservation. But unfortunately the image quality itself is very low, with few details.
can I take a look at the output?
@g1263495582 Example right in the topic. https://civitai.com/images/125724219 - low-quality photo, looks unrealistic, no skin details are visible, low hair quality and the image is blurry
@g1263495582 I believe that the problem is in lora. similar problems can be found in rapid aio, but in most cases the number of details is higher there
@cheyokey576 It's probably this part — 'The lighting is soft and ambient, enhancing her figure and the textures around her. The overall image exudes a powerful blend of vulnerability and self-assured allure.' — that might be triggering it.
@cheyokey576 In my experience, if an image generation model doesn't describe texture or style, the image actually turns out better than over-prompting it — being too specific can break the output. Also, I don't really use REMIX/Rapid-AIO in T2V mode that much.
@g1263495582 I agree that there's no need to describe this, and I also use the model in I2I, but the textures +- the same as in the example. Details are still scant, and I still think the problem is with LoRa.
There seems to be an issue with the Civitai download model. Could you please provide a link on HuggingFace? Thank you
In wan2.2 remix, there was a download link for a workflow using that model on the page, but I can't find one in qwen. The instructions say "load chatpoint," but that's not enough information. Can I just replace the model in the workflow installed from the official ComfyUI template and use it right away? Or is there a workflow that is best suited for this model? Any information would be greatly appreciated.
Addendum: As I am Japanese, I do not understand English or Chinese. I cannot use subtitles to translate the video, and I haven't watched it to the end yet, but if the video clearly explains things using mouse cursors or other visuals even if I can't read the text, I will try to use those as a reference.
I found the workflow. It was in the video description, my apologies.
it would takes more than 24g vram..