AnimePro FLUX
A fine-tune of Flux.1 Shnell, AnimePRO FLUX produces DEV/PRO quality anime images and is the perfect model if you want to generate anime art with Flux, without the licensing restrictions of the DEV version.
Works well between 4-8 steps and thanks to quantisation will run on most enthusiast-level hardware. On my RTX 3090 GPU I get 1600x1200 images faster than I would using SDXL!
The model has been partially de-distilled in the training process. Using it past 10 steps will hit "refiner mode" which won't change composition but will add details to the images.
The model was fine-tuned using a special method which gets around the limitations of the schnell-series models and produces better details and colours, and personally I prefer it to DEV and PRO!
Workflows and prompts are embedded in the preview images for ComfyUI.
LoRA Training: use https://huggingface.co/jimmycarter/LibreFlux-SimpleTuner as the base model for your LoRA, and then apply it to this model.
The License is Apache 2.0 meaning you can do whatever you like with the model, including using it commercially.
Trained on powerful 4xA100-80G machines thanks to ShuttleAI
Description
Runs on 9.0GB VRAM with --lowvram in comfy
FAQ
Comments (8)
Looking for a LORA animepro.safetensors ? New to ComfyUI so I may be missing something basic
Workflow for Comfy please!
I know this is alot to ask, but could you point me in the right direction to using the simpletuner base that you listed? I'm having a real tough time finding tutorials on it. And I REALLY Want to create loras for your wonderful checkpoint!
how do you train lora on this? use base model dev or straight up using this model?
AssertionError: You do not have CLIP state dict! :( Use flux1-dev-bnb-nf4-v2. Help pls
Looking for a specific prompt that makes the eyebrows look " ^ ^ " or " \ / " instead of " / \ ". I really love this model's style, but I hope I can create energetic looking characters more constantly.
What parameters should be set for this model? Sampler, sheduler, cfg?
On positive note: One interesting thing noticed, FP8 model is faster than Q5 model and that was surprising (speed difference was 50 seconds on 6 steps)
On negative note regarding to all those Flux1 models (not only this one): Very underwhelming results, Flux-es are impossibly slow, yes, they can generate text but that is the only thing they do better compared to illustrious or SD1.5 even. To give you comparison, it takes same amount of time for Flux model to do 6 steps that it takes illustrious model to complete 40 steps (with same settings), Flux also refuses to listen when it comes to colors, you use every word you know for dim, toned colors and it will still generate you vibrant image. It capabilities at anatomy are also bad, very comparable to SD1.5, the way it generates brows, mouths and teeth is honestly often worse than SD1.5, but it certainly beats it when it comes to fingers or eyes (but only in terms of logic, not quality).
And lets not forget about elephant in the room. Setting up Flux model is absolutely TERRIBLE experience, nothing works out of the bat, workflows are worthless (they don't work) and you just must get lucky and read right comments and add 1+1+2 to get 5 then 3 then finally 4. For the effort it took to set it up, not worth it, so heads up if there would be anyone interested in those Flux1 models - don't do it, use illustrious = faster & better results
(i'm not hating on authors flux model, but on flux1 in general and one above was actually one of the best flux-es model, rest of them were even worse.)
What you need to run:
DualCLIPLoader - and in it t5xxl_fp8_e4m3fn.safetensors (or fp16 for better quality) + clip_I.safetensors
Load VAE - ae.sft (must be sft, not safetensors)
You use Load Checkpoint for FP8 model and Unet Loader (GGUF) for Q5 model
Use Euler or Euler_cfg_pp and with them Simple or Normal scheduler.
(supposedly all Eulers are good but i would not be so sure, and same goes to Beta scheduler, apparently is compatible but see yourself. Don't use karras)
Using FP8 with t5xxl_fp8 with clip_I and ae.sft will require nearly 32gb of RAM to load.