Raemu Flux
A finetuned model built on top of Flux.1-dev, specifically focused on anime. It has been finetuned with dataset nearly 1,500 carefully curated, high-quality aesthetic anime images, resulting in noticeable improvements in anime scene rendering and visual quality.
Enjoying my work? Consider supporting the development of my models by buying me a coffee on Ko-fi
Also check out my other models on Mage: Raepix Focus XL / Raepix Focus XL Lightning / Raevitamix Realistic / Raevista XL
Recommended Setting
The model does not include VAE and CLIP, which are essential components. You must have them.
*Baked version already included all clips and vae.
Download both clip_l and t5 from here and place them in the respective folders: text_encoder for ForgeUI and clip for ComfyUI.
steps: 20 - 30
Distilled CFG/Guidance: 3.5 - 4.0
CFG: 1.0
Sampler: euler/dpm++ 2m simple
Vae: flux ae
clip: clip_l and t5xxl
Usage
Be sure to include "anime" in your prompt; otherwise, the output may appear realistic. For example: "A woman anime hold sign that says Hello World."
License
The model falls under flux-1-dev-non-commercial-license
Description
Baked version included all clips and vae
FAQ
Comments (18)
Is this a model learned through fine tuning? Or is it a model that merges the learned lora?
merged with a lora model that I trained
@Raelina Please just put this information in the tittle or information, so we don't waste time downloading a merge
@anonlinuxmint962 From what I understand, a checkpoint merge refers to combining models or LoRAs that are already available online from other creators. However, this model is one that I trained myself, so it hasn’t been merged with any other LoRA from other authors. Therefore, this is not misinformation.
@Raelina I definitely agree with you, I don't think you're misrepresenting anything, but I think you should consider releasing the LoRA weights by themselves. Flux models are gigantic as it is, you're asking people to have a whole 22gb network file around just to use your awesome LoRA weights!
@Raelina Merged means that you haven't trained all the layers in the model (complete fine tuning), and just merge different loras, yours or the others, with the original weight therein. Right now at least 90 percent of Flux checkpoints use this method, but don't make an advise about it, I know is too early for complete fine tuning models, and since is a 12B, is really hard to train with even a 4090, However is hard to follow the progress of people really fine tuning the model. Sorry for being rude at the start.
@anonlinuxmint962 Flux is a big model with 12B parameters. I think we won't see many complete fine-tuned Flux models in the community soon, even might be very rare in the foreseeable future. Considering fine-tune a Flux could cost much more than SDXL - maybe several times more, most people would find this too expensive to do.
is there a way to extract a usable lora out of a model like this? I dont mind downloading the 22GB and do the above processing, but, i dont want to consume that much storage long term when 200MB to store it as a LoRA or similar would suffice. Thanks
I'm not sure what I'm doing wrong. I tried to use similar prompts as you but I'm just not getting an anime style. It's blending realism with anime and the eyes are far too small.
This model is too big for my gpu to run, even my friend's 4090 is struggling ---- Other models can run and play the game at the same time. I see a lot of FLUX models are around 11G, is it possible to shrink this model a bit? So that a 16G gpu can also run it?
Thats why fp8 exist. Use fp8 instead. Even 8gb vram is enough for fp8 on forge.
Thank you. But this author's PONY and IL models are great, so I'd like to try his FLUX ones too. So far his FLUX models are too big.
@dcdcforever I could barely run this on an 8GB VRAM card (with Forge on the backend), but on a 16GB card it works fine. I'm using the "baked" version.
can i use this with webui?
It seems censored, I can't get it to do naked bodies :<
Flux doesn't support nsfw. If you're looking for nsfw, it's better to stick with sdxl