Deskup Anima Edit General
deskup-anima-edit-general.safetensors is a general-purpose image-editing LoRA designed for the Anima model. It was trained on approximately 80,000 image pairs and subsequently merged with Anima Edit.
The LoRA was trained for:
- Pose changes
- Background changes
- Character changes
- General scene modification
This LoRA is experimental and is unstable. The same prompt may produce inconsistent results, and some edits may require several retries before producing the intended outcome.
When changing background, original one might bleed into new image.
For background changes I suggest to use single color background as reference.
Usage
Use the LoRA through either of these interfaces:
- ComfyUI: Install the ComfyUI-Cosmos-Reference custom node, then begin with the provided sample workflow.
- Stable Diffusion WebUI: Use the neo branch of sd-webui-forge-classic, which supports Anima image-editing LoRAs.
Prompt primarily in natural language. Supplement the prompt with tags when natural-language instructions alone do not produce the desired result.
Prompting guide
The LoRA accepts both descriptive captions and direct edit instructions. For more stable results, describe the desired final image clearly:
A character wearing YYY, in pose ZZZ, against background XYZShort instructions can also work:
make character xxx a xyzchange yyy to xyzreplace zzz to xyzWhen replacing something, focus on the desired result or name the broader category instead of emphasizing the old item's exact appearance:
Prefer: Replace the outfit with XXXAvoid: Replace the white shirt with XXXIf an unwanted element remains in the result, add that element to the negative prompt. If the model fails to preserve an important character, object, style, or scene detail, explicitly mention what must remain in the positive prompt.
For difficult edits, keep the instruction concise and generate several variations rather than making the prompt increasingly complicated.
Description
FAQ
Comments (13)
You said it was merged with normal Anima Edit. Does that mean that if we have that Lora, we don't need this one?
Yes, you can use either https://civitai.red/models/2650553/anima-edit
or mine.
You should not likely use them together.
These loras are similar. Anima Edit lora is arguably more stable, but often fails to apply more drastic edits.
Thank you very much for your work! Based on my testing, compared with other editing LoRAs, this one seems much better suited for making larger changes, such as replacing characters or changing art styles, and the results are particularly impressive.
However, sometimes certain elements seem to be quite difficult to remove, such as a character’s hairstyle. For example, I’d like to change long hair or twin tails into medium hair. In most cases, the original hairstyle can be removed successfully, but sometimes only the hair color changes while the hairstyle itself remains unchanged. According to the instructions, I also tried putting “long hair” / “twin tails” in the negative prompt, as well as using remove/replace to modify the hair, but neither seemed to help much.
Generating variations with different random seeds can work around this issue, but it seems that some images have a higher failure rate for removing the hairstyle than others. I hope this can be improved in the future. Or perhaps there’s something I’m doing wrong?
No, you are not doing anything wrong. It is limitation of the lora itself unfortunately.
I would likely need to increase lora size and retrain it, as at the end of the training it was showing symptoms matching too small model size.
Using the anima turbo int8 model with the workflow I posted in the gallery, I was able to edit effectively(CFG4.0 is strange on anima turbo,but it works).
It doesn’t offer rewriting capabilities as powerful as qwen image edit,but it can generate results at a practical speed even on low-spec machines.
Thank you.
Very inconstant but it works.
Did you make your own dataset, or i just can't find the anime one
Yes, I made it myself.
At least part of it.
I used some high quality general datasets from huggingface to start training on.
And my own custom anime dataset of 20k image pairs. Captioned with mixture of gemini 3.6 flash and grok.
About 1k pairs was generated using nano banana, grok and qwen models, rest was scraped by my custom pipeline for picking most suitable images for edit instructions.
Though, when preparing it again, I would only use gemini models.
Captioning cost me a little though, and it's not like I'm getting any money from this, so it's not like I will post improved model any soon. (Unless somebody would paid me).
@Deskup that's the hard part
Can you share a screenshoot or something how do we use on webui Forge ?
Just use this lora as other any other type of loras,
and then use img2img tab
For some reason, it cannot change the character's pose. Modifications other than the character's pose all work correctly. This issue also occurs when using Anima Edit.
unfortunatelly, this lora is not as strong as dedicated image edit model