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 0.1
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
As far as I know, generation providers do not support Anima Edit workflows for online generation, so 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.
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)
Prompting
Use Natural Language for the edit instructions, some examples:
"Add a blue background."
"Add a glowing white rabbit on the right side."
"Switch the background from a snowy pine forest to an icy mountain vista."
"Change the blue hair color to white."
"Replace the flowing sundress with a black blazer."
"Remove the crowd of people."
Tags can be included with the Natural Language instruction and help avoid the Flux style influence somewhat, recommend including at least the quality modifier tags:
"masterpiece, best quality"
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}
}Description
Version 0.1
Used 984 source-target image pairs subset of the NHR-Edit synthetic dataset.
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.
FAQ
Comments (10)
đź‘Ť
this is going to sound stupid but how is this used
Hey, not at all - I've updated the details of the Usage section of this model page
If anything is still unclear please let me know
This LoRA is practically unusable. In my workflow, no matter how I adjust the settings or nodes, the output barely changes or doesn't change at all—there is hardly any editing taking place.
So I tested running the exact workflow from your image locally in ComfyUI without changing anything (settings, nodes, or prompts. also model clip and vae). The result was that the image didn't change at all.
Hey, sorry about this
> So I tested running the exact workflow from your image locally in ComfyUI without changing anything (settings, nodes, or prompts). The result was that the image didn't change at all.
Can you confirm a different Anima Edit LoRA works as expected in your local ComfyUI with the custom ComfyUI-Cosmos-Reference node?
Alternatively, you could try it in the latest Forge Neo.
@motimalu I am currently using that LoRA (v2). Although there are times (RNG) when it seems to fail or give incomplete results for unknown reasons (and I am not the only one experiencing this), it is actually functional. When it works, it performs exceptionally well despite having some limitations.
That LoRA works with both Mirumo0u0/ComfyUI-Cosmos-Reference and Anzhc/Anzhc-ComfyUI-Cosmos-Reference, and it functions properly even across a variety of node combinations.
My workflow also works seamlessly with the Colorizer ANIMA Edit LoRA as well. Based on various tests, I have concluded that the issue is unlikely to be on my end.
@motimalu My apologies—I checked again and realized I forgot to check the prompt properly when testing your workflow. After re-testing, your LoRA seems to work normally in your workflow (I think). I'll test it again when I'm free.
@motimalu I re-tested it. The factors that clearly affect the quality/performance of your LoRA are CFG and resolution while passing through KSampler.
I found that CFG behaves like denoise/strength: the closer it gets to 1, the lower the chance or effect of image editing occurring. Turbo LoRA cannot fix this issue (I haven't tested it with distilled models specifically designed for low CFG generation yet). Meanwhile, resolution acts like the strength parameter of Cosmos Reference: the lower the resolution during generation, the higher the chance of the image deviating from the reference latent due to the editing effects. Using a mask to limit the edit area works, but it doesn't improve the LoRA's overall quality or performance. This clearly impacts crop & stitch workflows. From its behavior, it seems the LoRA fails to understand what it's seeing if the latent has low context—meaning that even at high resolutions, editing is hindered (such as cropping just the head to specifically change hair color).
In short: CFG needs to be kept high (Turbo LoRA doesn't help), and image resolution must be high enough, otherwise the output will distort from the original. Both points directly affect users without high-end hardware specs or those looking to optimize generation as much as possible (like myself).
My testing was done using your workflow with only minor tweaks, not a complex setup. I believe my testing isn't completely exhaustive yet, but hardware limitations prevented me from experimenting further.
@motimalu A small addition—I haven't tested this enough yet, so this is just an inference based on a few results. It's possible that your LoRA's overall performance (specifically, editing with precise reference to the original) drops drastically when dealing with images that play with perspective or images requiring minor edits to a few overlapping details—such as unusual POV angles or removing censor bars.
Hey @Nanai771Â thanks for testing it further.
Though I do appreciate that it would be desirable to optimize generation, it is not expected that this LoRA would work with lower CFG or combined with a Turbo LoRA, since the Turbo distillations target fewer step generation of text-to-image outputs, not image-to-image outputs with the edit task.
As for what kind of tasks it can perform well with, I also don't expect too much generalization outside of the narrow tasks it has been trained with, since it just used a small test dataset of synthetic data so far.
Idea is to iterate from here and build a large dataset though, so I appreciate the feedback.