YAAE | Yet Another Anima Edit

As the title implies, this is not the first Anima Edit model and mostly a proof-of-concept test of a diffusion-pipe update to Support Anima Edit Training as of the current version.

Version 0.2

769 source/target pairs or 1538 total images.

Used exclusively non-synthetic data for this version, which was a bit of work for each edit task to prepare source and target pairs.

Trained at 512^2, 1024^2, 1536^2 resolutions on a combination of several edit task datasets for:

• Text control

• Clothing control

• Time of day control / relighting

• Sketch, Line-art and Detail Level control

• Personification

Notes:

• Text control is mostly best at removing, with adding or changing not having as high of a success rate

Version 0.1

Used 984 source-target image pairs subset of the NHR-Edit synthetic dataset.

Trained at 512, 768, 1024 resolutions, input images should not exceed 1024 for best results.

Can use a resize node to ensure this on input images in ComfyUI as I have in the preview workflows.

Target edit tasks, 90 image pairs each:

• Add Background

• Add Object

• Change Background

• Change Color

• Change Object

• Remove Background

• Remove Object

And mixed 354 images of all categories in the "anime" style.

The "augmented_instructions" column from the dataset is used for the target captions.

The dataset for this proof-of-concept version consists of synthetic Flux outputs, which isn't very compatible with the anime focus of Anima.

Usage

This LoRA is only usable in local generation with ComfyUI or Forge Neo.

The basic idea is to take an input image and run img2img generation with a prompt that contains your edit instruction.

Ideally the edit instruction should fall under one of the categories that were explicitly trained for this LoRA (see version details), but the model is capable of some level of generalization.

Prompting

Use natural language with the clearly worded edit instructions, for example including the words Add details.
Use tags to guide the generation as you would for a regular t2i generation.

The preview images mostly use the ideal prompt structure, which is something like:

Add details and soft textures to the image.
Change the girls expression to a slight smile.
Change her outfit to black shorts and a bikini under a wet white shirt with the text "HOSHIMACHI".
masterpiece, best quality, hoshimachi suisei, star in eye, bikini under clothes, garter straps, thigh strap, black shorts, white shirt, translucent, smile

Explanation of the workflow

The source image is a real image that is loaded and resized to a 1024^2 pixel SDXL resolution (832x1216), optionally scaled by 1.5 to a 1536^2 resolution using the custom JPS-Nodes, and the ComfyUI KJ-Nodes - Resize Image v2 node.

*The resizing is important to keep the source image in a normal range that matches both Anima's training and this LoRA's pixel resolutions with sides divisible by 16, version 2 is trained at 512^2, 1024^2 and 1536^2 image resolutions.

The Anima Model and this LoRA are all loaded normally, the PC Schedule Prompt custom node is optionally used mix artists in the prompt with the [artist1|artist2|artistN...] format.

The ComfyUI-Cosmos-Reference custom node is used to apply the Cosmos latent reference of the source image to the generation, which is required for this LoRA to function.

In the "Save" group, the Draw Text, Resize and Image Composite Masked nodes are finally used to write the "Reference Image" text above the source image, and compose it over the generated image in the top left corner.

ComfyUI

The ComfyUI-Cosmos-Reference custom node is required to use the workflow.

The workflow is included in the preview images, you can download the images and drag them into ComfyUI.

Forge Neo

Forge Neo also supports cosmos reference natively, so you can use this LoRA in the img2img tab. (Make sure to enable it in Settings->Stable Diffusion)

Citations

@article{Layer2025NoHumansRequired,
    arxivId = {2507.14119},
    author = {Maksim Kuprashevich and Grigorii Alekseenko and Irina Tolstykh and Georgii Fedorov and Bulat Suleimanov and Vladimir Dokholyan and Aleksandr Gordeev},
    title = {{NoHumansRequired: Autonomous High-Quality Image Editing Triplet Mining}},
    year = {2025},
    eprint = {2507.14119},
    archivePrefix = {arXiv},
    primaryClass = {cs.CV},
    url = {https://arxiv.org/abs/2507.14119},
    journal={arXiv preprint arXiv:2507.14119}
}