Clip specific materials.
■Material test images of spheres covered with fabric-like layers, like the sample, are quite common.
Sometimes you just can't help but wonder what's underneath. So I thought using a LoRA to remove the covering would resolve that curiosity. Now, there's no need to keep wondering about it.
■This LoRA was created for editing anime-style images.
■Details and prompts can be found in the shared workflow.
■If you're new to Kontext, try installing it using the instructions at the URL below.
It's well-documented and easy to set up.
https://docs.comfy.org/tutorials/flux/flux-1-kontext-dev
■This was trained 2D illustration images.
The scale is small, but the detection rate is surprisingly high.
■If you're not satisfied with the texture or detail after clipping, try i2i or inpainting with another model.
SD1.5 is lightweight and works well as a high-resolution refiner—recommended for i2i. It only uses about 2GB of VRAM, but it's very powerful.
It performs reliably up to 1024×1536px.
Try denoise values between 0.25–0.5
■If you're unsure which SD1.5 model to use, feel free to try my merged SD1.5 models first.
I provide both real and anime styles—use whichever fits your image better.
If the results don’t match your expectations, you can then look for a model closer to your ideal!
https://civarchive.com/models/1246353/sd15modellab
■If you have enough VRAM, SDXL could also be a good option.
■I used AI Toolkit to train this LoRA.
If you're interested in training, the developer has provided a tutorial at the URL below — give it a try!
I think you'll find it's easier than you might expect.
Description
The dataset has been increased to 2400 images.
As a result, this version tends to have better detection and stay truer to the original style than v01.But v1 is often more appealing.
Training prompts have also been updated—please check the workflow.
However, it's not a perfect upgrade over v01.
Due to increased variation, detection may fail or results may look worse in some cases.
You might get a better balance by using v01 and v02 together with 0.5 weight each.
FAQ
Comments (12)
Any suggestion on how to make the "spheres" clearer once lora has been applied? Seems to be all blurry or deformed, thanks to flux.
Thank you for trying it out.
In my case, I detect and convert first, then refine with another model using i2i for better texture.
If the object has unclear contours, changing the seed might help. Sometimes v01 works better, so try that too.
Usually, it keeps the shape well and only needs texture cleanup. But if your results are very different, it might be due to the Kontext model or workflow.
If you’re using my workflow and model from the docs and it’s still blurry, I’m not sure why...
https://docs.comfy.org/tutorials/flux/flux-1-kontext-dev
If nothing works, feel free to DM me on Civitai with the image URL you want to convert. I can check on my end too—sorry if the issue is due to my LoRA quality...
hjhf Just tried your v1 and it is way better for details underneath! Seems the v2 is very good at containing the shapes of the original input, but very potato design under. Almost feel like it's flat with no depth at all, almost drawned. V1 doesn't exactly keeps the perfect shape, but make it way more appealing.
And yeah, I've used the default Kontext workflow, can also be viewed inside Comfy under Workflow > Browse Templates > Flux > Flux Kontext Dev(Basic). Still learning how to use this witchcraft of model, it is very unique compare to any other as you can write specific modifications to I2I instead of a complete filter.
Love your other loras though! Still need to try them on some of my "Readings" as it'll definitely bring more life to the characters.
ItsThatTimeAgain
I'm glad to hear it worked better!
Your feedback was spot-on—I felt the same way.
v01 was made from 70 carefully selected and manually edited ideal difference images, so I agree it tends to give more pleasing results.
I created v02 by increasing the dataset to improve detection of objects v01 struggled with, but it lost some of its charm — sometimes too faithful in style or a bit rough.
If I find a better balance between style and detection through more careful curation, I might release a v3—but that’s still undecided.
Thank you for using the LoRA. I made it because I wanted something like this myself, and I'm really happy if it can help others too—for artwork or training resources.
Have you tried adding some flux LoRas with low strength?
The problem I have is that it's very 'off' or 'on' there is no middle-ground like what you can achieve with style transfer or general LoRa application.
zefy
Sorry, but I haven’t used the regular flux_lora, so I might not be the best person to answer.
This might not be exactly what you're looking for, but in my case with Kontext_lora, slightly lowering the weight helped reduce artifacts when it was a bit overtrained.
Also, applying v1 and v2 at 0.5 weight each sometimes gave a well-balanced result.
If you're refining an existing concept, blending might work, but for entirely new concepts, the transition might be more distinct.
* * * Workflow * * *
https://civitai.com/models/1831687/flux1-dev-krea-kontext-nunchaku-my-20sec-workflow?modelVersionId=2107703
Tested: version 3 / 2025.08
https://civitai.com/posts/20728753
--------------------------
This LORA get you awesome quality results
Workflow 2 faster than workflow 1
Absolutely great (tip: "use the training image"). If you only have one Flux Lora, this should be it.
It does exactly what you want it to do on everything it was tested on. Now I just need to work out a comparator mask, or something because sometimes I get two belly buttons.
Thanks for trying the LoRA!
I’m glad it gave good results.
As you said, detailed techniques will become more important—like preprocessing or post-fix after conversion to improve accuracy.
As for the LoRA itself, v2 detects better than v1 but lost some of the refined look. I plan to keep updating it with a more curated dataset until I find a better balance.
The issue might get resolved in the process.
hjhf For refining an image, this model does a great job (even without a prompt): https://huggingface.co/Anibaaal/Flux-Fusion-V2-4step-merge-gguf-nf4
Depending on the implementation, the 'denoise' has to be set to 0.32-0.4 for some programs, 0.2-0.3 for ComfyUI ("simple" scheduler)
Usually works best in 3-4 steps for this task.
Tends to clean up bad hands, fudgy faces, probably also extra belly buttons.
Thanks for letting me know!
It's great that it works with low steps even on Flux—makes it faster, which is nice. It might indeed be a good fit for fixing.
I'll consider using it as a correction model.