Due to the controversial updated terms of service on Civitai, in the future, I will only release the Base version of the model. The self-used Plus version will only be shared with members on Patreon. If you are interested in this, you are welcome to join my Patreon membership, and I have also prepared a small gift for you.
About LoRA
Converting photos into hand-drawn illustrations with a realistic touch has good compatibility with most 3D styles and hand-drawn styles. It supports speed LoRA, all schedulers are good except "simple". Most of the test images you see come from "sgm".
Differences between the Plus version and the Base version
1. The colors of the Plus version are more vivid
2. The Plus version makes fewer changes to the image.
3、The texture of the Plus version is more delicate, and the details are more abundant
Get the Plus version
Tips
While converting the image style, the image content can be edited. From the test results, the Plus version is also superior to the Base version in this aspect.
prompt：changed the image into realcomic style, Turn the woman holding a baby into a robot
🙈🙉🙊🐵OK! That's all
If you like my work, please give it a 👍.
Thank you for reading this far.
English is not my native language. The above is translated by AI. If there is anything inappropriate, welcome to correct me. Finally, I hope you have fun！
👉👉👉My CivitAI homepage👈👈👈
Description
FAQ
Comments (12)
so base is just a less trained checkpoint?
Why ask that?
why complain ?
@vjleoliu because it's a good question. Why are they different?
@ArtfulGenie69 That's indeed a good question, but I need to know the reason why he asked it this way. Because I often receive inquiries about model training, and they all explain their situations, so that I know how to help them. Of course, some people don't really want to discuss technology; then I can only answer based on my mood or ignore them. So, why do you ask this?
Idk but for me I love to have the option between full stylized and more raw anime styles, exactly as you've produced it
Can this model substitute your "ColorManga" in a 3D->Drawn->Real scenario like you explained here https://www.patreon.com/posts/how-to-convert-141187041, but this time using Qwen2509? Also would you recommend the speed lora over the base inference or it alters the output of your lora in a significant way? Is there a specific trigger sentence or words to use ( like "change the image into realcomic style")?
If you're looking to achieve a 3D-Realism effect, I don't recommend purchasing this LoRA because I'm working on a better solution, and it will be free. It might be released later today, but I need to do some testing first.
Just from looking at your examples, this looks super good. Awesome stuff!
thx，bro，If you think this model is good, please feel free to upload your test results and give it a like. This will help more people discover it.
The performance of your Lora is excellent. May I ask how you conduct your training?
Please allow me to first describe my issue.
I am training a artist's style on Qwen-Edit-2509. I used 16 sets of data, and their composition was as follows:
Target image: From this artist's work.
Control Image: I used Flux Kontext to convert the painting into a realistic photo.
I once successfully trained a very effective model on Kontext using these data.
Now I want to replicate it on Qwen-Edit-2509, but the preview images during the training process are very poor. Steps reach 6000/10000 now. However the preview images are either not different from the control image, or have no relation to the structure of the control image (like a new work by this artist that didn't exist before, just like the performance that a SD15 Lora of "style" would have had).
What is the reason for this? Could you tell me what I should do?
The following are some parameters I used for training on Qwen-Edit-2509.
type: "lora"
linear: 32
linear_alpha: 32
conv: 16
conv_alpha: 16
lokr_full_rank: true
lokr_factor: -1
dtype: "bf16"
batch_size: 1
steps: 10000
gradient_accumulation: 1
train_unet: true
train_text_encoder: false
gradient_checkpointing: true
noise_scheduler: "flowmatch"
optimizer: "adamw8bit"
timestep_type: "weighted"
content_or_style: "balanced"
lr: 0.0001
prompt: " Convert the picture to an illustration style."
you are the best ! I hope you could also make another styles for qwen, for it's the only model that can transfer the style and preserve the original content. I have tried so much methods and this lora the only that stands up. unfortunately the number of style lora for qwen edit are very very limited. so we hope that there is more coming in the near future