Please use with v3.3g workflow
https://civarchive.com/models/933018/
This lora trained on simple cloth and model pair. It is designed only on try on purpose.
Flux Template:
[Try On], The image pair highlights a transformation from a clothing sample photo to the effect of actually wearing it. ; [CLOTH_IMAGE] {Image1_desc}; [MODEL_IMAGE] {Image2_desc};
Enter the description of image1 in the following text box:
Input Image Description Refer to [IMAGE1]
Example:
The image depicts a white T-shirt with a graphic print on the front. The print features three iconic animated characters from left to right, Micky mouse, Goofy dog and Donald duck, crossing a city street. below them, the phrase "NEW YORK! NEW YORK!" is prominently displayed. The characters appear to be in motion, walking across the crosswalk, with buildings lining the street in the background. The overall design combines elements of popular culture and urban scenery.
Enter the description of output image in the following text box:
Output Image Description Refer to [IMAGE2]
Example:
In a serene dusk outdoor setting, fullbody shot of a whiteman wear the [CLOTH_IMAGE] t shirt in a street
Description
FAQ
Comments (21)
I'm not having much luck with this. Maybe 1/15 generations I'll actually get a result with a person and the product. Also the product isn't identical to the input image either. Out of the 15 generations, I'll get like 10 of just a copy of the product, 4 will be with a person and no product, and 1 will actually be person + product.
There are several reasons. First, your product might be very specified unique and the flux model very hard to match with your product. Second, you need to adjust your input image prompt to match with your product and following the instruction in the description. Third, the output prompt might contradict with the image, you need to carefully adjust the prompt and avoid the contradiction. Might you post your attempt in the gallery and let me try? Product image and the result.
@xiaozhijason Thanks for responding. I added it to the gallery: https://civitai.com/posts/9330237
The prompt is broken down into 3 sections with the template, image 1 description, and then the output image. I kept the last prompt simple as I'm just trying to test this out to see how it works. You'll see in the image I posted, I'm not getting the product at all.
@hexus0 I think I get the main point. The try on lora is specific trained on cloth images and your input image is a hat. It tried to convert your hat as a cloth. I suggest to use my another lora IC_CONSIS for this task. https://civitai.com/images/40957131 Example: https://civitai.com/posts/9331631?returnUrl=%2Fmodels%2F933026%3FmodelVersionId%3D1069974
@xiaozhijason does this workflow use only the image prompt aka text to get the lora to recreate the product?
@namee You have three input, image input, image description and generation prompt. The image description would also affect how the final image genarated.
Tip: Replacing the prompt part of the template workflow with Redux greatly improves the output. Would be cool to see the creator's official take on this
@GitarooMan You might use the version 2.4 workflow which only use redux and fill
I’m deeply impressed, thank you for sharing!
Is it possible to view the training images and settings for this LoRA?
The example is how training images labeled. lr 1e-4 adamw8bit 20 repeats 10 epoch
Did you use ai-toolkit or automatic1111 for training?
@alarraz151 No, this lora was trained on civitai lora training. Newer lora was trained on my personal training script.
@xiaozhijason Thank you for your answer!!! I'm trying to train an IC-LoRA using ai-toolkit from ostris and everything goes well but, in inference time the results are worse than expected, what script do you use to infer?? an inpainting pipeline?
@alarraz151 +1 on this . the ic lora repo suggests ostris trainer but results are average
Could you kindly let me know if it’s possible to access the training datasets used in your work? Thank you!
No, you might try this dataset. https://github.com/shadow2496/VITON-HD
Is there such a training code to train flux dev that generates another image from one conditional image?
It is just a lora which is no different than normal lora. generate conditional image is not doing in the train
Somehow this lora seems to help changing outfits in combination with flux kontext dev. Very interesting!
would you mind sharing your wokflow with flux kontext?
@lunaticfringe
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.