Sign In
homemodelsimagesvideosarticlescomicschallengesupdatesshop

mokuba_colab_anima

0

mokuba_attack's Avatar

mokuba_attack

Jun 13, 2026

tool guide
mokuba_colab_anima

I introduced hdae/diffusers-anima module to mokucola. But you need an abundance of time to make images in Google Colab (you need ~6 minutes to make a 1024-1024 image.), because T4 GPU doesn't support bfloat16.

requirements

Change the runtime type to T4 GPU.

Next, run next code on Notebook.

!pip install https://raw.githubusercontent.com/MokubaAttack/scripts_anima/refs/heads/main/mokuba_colab_anima/mokucola-42.11.128.tar.gz

import mokucola

explanations

mokucola.mokuani(

loras, lora_weights, prompt, n_prompt, pic_number, gs, step, sample, sgm, seed, out_folder, base_safe, url, p, dtype, dev, ser, del_pipe, si, mode, up, Interpolation, step2, ss

)

  • base_safe : str

It is the checkpoint file.

  • loras : str list

It is the name list of the lora file excluding extension. If there is not that file in the working folder, you must input the absolute path. LyCORIS is supported too.

  • lora_weights : float list

It is the lora's weight list.

  • prompt : str

It is the prompt.

  • n_prompt : str

It is the negative prompt.

  • step : int

It is num_inference_steps.

  • gs : float

It is guidance_scale.

  • sample : str

It is the scheduler type.
flowmatch_euler
euler
euler_a_rf
euler_ancestral_rf

  • sgm : str

It is the noise schedule and the schedule type. 
uniform
beta
simple
normal

  • pic_number : int

It is the number of the output images.

  • seed : int or int list

It is the seed or the seed list. If you input zero, the random seeds are made.

  • out_folder : str

It is the output folder path. If the folder doesn't exist, that is made.

  • si : bool

If you choice True, output images are shown in the output window.

  • url : str

If you input the webhook url of discord, images are sent to discord.

  • del_pipe : bool

If you choice True, the mokupipe object is deleted and None is returned.

  • x : int

It is width of output image.

  • y : int

It is height of output image.

  • ser : str

In google colab, please input "colab". In kaggle, please input "kaggle".

  • dev : str

It is the device that calculates. Choices are cuda and cpu.

  • p : mokuanipipe object

If you input the return of this module, you can use same pipeline without making the pipeline.

  • dtype : str

It is the calculation accuracy. Choices are f32 and bf16.

  • ss : float

It is denoising_strength ( a parameter of hires.fix ).

  • mode : int

It is the working mode.
0 : normal
1 : hires.fix

  • up : float

It is the upscale ( a parameter of hires.fix ).

  • Interpolation : int or str

It is the interpolation method of the upscaling. If you input pth file of ESRGAN, images are upscaled by ESRGAN.
1 : NEAREST
2 : BOX
3 : BILINEAR
4 : HAMMING
5 : BICUBIC
6 : LANCZOS

  • step2 : int

It is Hires steps ( a parameter of hires.fix ).

  • return : mokuanipipe object

Image files are output by naming (index)_(the seed).jpg in the output folder path. If safetensors files have CivitAi's Version ID in a item of "id" of metadata (In case of a lora file, lora's weight in a item of "weight" is needed too) , Generation metadata is baked in Output files.

(Example)

lora file

"id" : "111111", "weight" : "1"

merged lora file

"id" : "111111,222222", "weight" : "0.5,0.5"

ckpt file

"id" : "123456"

The metadata is read in CivitAi.

0