This is the first big model I have worked on for several days. I suggest running it on comfyui. Below is a comparison of the original model and this model with the same parameters.
Description
FAQ
Comments (15)
I can't make it run in Forge, while all the other Flux 1.D models run well. Any tips? Really like what you created with this model. Thanks a lot!
Um, you can use comfy UI to generate images
😒 Hi, there,
I have an error message in Forge :
“AssertionError: You do not have CLIP state dict!”
can you integrate the missing encoders?
Thanks
Um, I'm also a novice, this is my first time trying a large model, I haven't used Forge before, you can use comfy UI,
@AIhuihuahua No, I don't want to use Comfy UI because it's very slow and it freezes if I use several loras at the same time.
Can you just make an “All in one” pack with all the appropriate encoders?
I don't know if it's complicated but many users do it for their models.
Thanks
This is fine-tuned based on flux-dev. I haven't used Forge yet, so I don't know much about it.
@AIhuihuahua ok, 😒
@Suzanne I tried to fuse a large model with lora. I need to learn. The worst result you can get is this lora
@AIhuihuahua for the moment it's not a lora problem but an encoder problem (vae, clip text, etc...)
If you've used specific models, you need to indicate them on this page so that we can install them too.
If you want to learn how to merge encoders or lora, you can ask those who have done it on their models, I think they'll be happy to help you. 😊
@Suzanne This fine-tuning model is originally an experiment. The training set only has 60 pictures. I just tried fine-tuning a large model. Normally, the training set of a large model fine-tuning should require hundreds, thousands, or even tens of thousands of pictures to achieve a good effect. Now I have re-trained Lora with the training set. You can try it.
Finally, You can download the (clip_l.safetensors ）model and try running it
@AIhuihuahua I've already tried clip_l.safetensors from the beginning, but it doesn't work.
What is the format of your checkpoint?
Can you convert it to NF4 FP8 format for example?
I read somewhere that it's easy to do....
@Suzanne The version is flux-dev, not quantized, you can quantize it, I tried it but it went wrong, you can try it with weibu
@Suzanne The quantitative version has been updated. The memory was not enough before, but it is now OK after adding two memory sticks.
@AIhuihuahua ok, I'll try it later, thanks to you
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.