Please take a look at the following article for a full guide on how to use this workflow effectively:

https://civarchive.com/articles/29870

Changelog:

v1.0

Initial release

v1.1

Adding Regional Prompting!
Please Note: Regional Prompting uses a node that is not available in the standard ComfyUI manager. You must manually install it if you want to use it, which requires some basic git knowledge. https://github.com/Sen-sou/Comfyui-Anima-Regional-Conditioning/tree/main
I highly suggest you give it a try. This node set works great for Anima.
Changed the Face Detailer model. I found that the model I used in v1.0 tended to get into infinite loops for Anima. See the bottom of this guide for the download link.

V1.2

Added quick turbo mode switch • Added “Negate Prompts” to Turbo mode by using CLIPNegPip.
Moved 1st refinement step to legacy (The process can mess up images when using it with CLIPNegPip in Turbo mode so it is disabled by default. You can still use it, just be aware that it does not play nice with negative values in the positive prompt)
Implemented new NSFW detailer https://civarchive.com/models/1313556/anime-nsfw-detectionadetailer-all-in-one?modelVersionId=2266294

V1.3

Added ControlNet support
Added Prompt Crafter section (styles, backgrounds, subjects, actions, and more)

V1.4

Added Inpaint support (Page 10)
Reworked Image2Image to use standard KSampler instead of SDSU
Rearranged nodes to fit better on smaller monitors
Swapped base model for LoRA patched model for all detailers to maintain art style consistency

V1.4 (hotfix1)

Moved ClipNegPip outside of the Turbo section and into additional options. It was discovered that there is incompatibility with ClipNegPip and Anima Regional Conditioning. It otherwise works fine with everything else in the workflow

V1.5

Added new mode! Edit mode using the Anima Edit Lora.
Moved Inpaint section outside of Img2Img to allow for use in other modes in the future.
Regional Prompting patched model now routes to Img2Img and Txt2Img to allow for regional prompting in Img2Img mode. Also routes to detailers for higher adherence to regional prompting
Added Set Latent Noise Mask to Inpaint section to isolate masked area in generation

V1.6

Added LLM assisted Prompt Generator using Gemma 4
Added SeedVR2 Upscaling support
Larger output images (67 mega pixels) by using SDSU + SeedVR2 + RTX super resolution. (No idea why you would need an image that big…. But you can do it!)

V1.6(hotfix1)

Replaced nodes for all ControlNet sections. Upstream changes caused conflicting names on nodes. This also includes Inpainting.

V1.7

Complete rework of Regional Prompting using Anima Regional Canvas V2.0

V1.8

Added multiple small features
- SAMLoader to better isolate detailed areas with pixel exact masks
- CFG Zero Star to use higher CFG values
- Node to remove backgrounds from outputs
Added comfyui-anima-tools nodes for browsing artist, character, pose, clothing, and backgrounds.
Moved eye detailer before final SDSU run to hopefully get higher detailed eyes

V1.9

Implemented Dual Ksamplers

Removed some node dependencies
- Miosp/ComfyUI-FBCNN
- sipherxyz/comfyui-art-venture
Added additional post processing effects

V1.10

Fixed Prompt Generator and LLM system prompt support
Added Inpaint support to Edit Mode.
Fixed an issue that causes LoRA metadata not to be carried over to the generated image
Replaced problematic rgthree nodes with BCNodes

V1.11

Added Ref2Img mode!
Removed ComfyUI-basic_data_handling as a dependency by reworking SeedVR2 math.
Fixed issue with LoRA’s not being included in image metadata

QUICK NOTE
💖 Thank you everyone for trying this workflow out! 💖
It's been fun to put together and I have learned a lot doing it. Please feel free to suggest new feature for me to add!

Below you will find a quick overview, but I encourage you to take a look at the full guide.

So what makes this workflow special? It is specifically how your detailers work.

A common annoyance I have is if you are generating large batches of images with a generic prompt for your detailers, they will tend to “normalize” the area they are detailing. Example: An image of a character with an angry expression gets sent to a face detailer with “face, beautiful face,” it will slowly morph the face into a neutral of smiling expression with multiple passes.

To combat this problem, this workflow will feed your face and eye detailer prompts into your body prompt, and all three will feed into your base prompt. Here is a quick visualization of what is happening.

Essentially, you input your detailer text and wildcards on one central location. Eye detailer text will feed into your face detailer, your face detailer will feed into your body detailer. And all detailers will feed into your global prompt.
Here is an example base prompt. This is Anima, so you can use natural language in your prompt.

And now I input prompts for the detailers

You will notice I included emotions of the characters in the face detailer. This will get fed into the base prompt but also ensure that the face detailer knows about the expressions.

I also included the names of the characters in the eye detailer. This may seem odd at first but bear with me. I did this because the eye detailer is fed into both the face and body detailer. This ensures every detailer knows what character it is drawling, all the way down to their eyes. (Refer to the diagram at the top)

Note: At the bottom, there is a preview box so you can see what your prompt looks like when it gets sent for initial image generation. This is helpful when you are first starting to use this workflow.

Note about LoRA's

This workflow has an switch that will enable the Turbo LoRA, and configure it properly with CFG 1, and 12 steps. You can modify these settings if desired in the Turbo section.

https://civarchive.com/models/2560840/anima-turbo-lora

Normally, negative prompts are not evaluated at CFG 1, but this workflow uses CLIP NegPip to get around that limitation. You can then add tags to your positive prompt and give them a negative value, allowing you to achieve results much like a negative prompt. Below is an example of this.

Original image Adding (coffee cup:-1) to positive prompt

Clip NegPip can mess up your generation in the first Refinement step in the workflow. This is why I have disabled it by default. You can still enable if needed. (Works fine when not using CLIP NegPip)

If you want to generate at resolutions higher than 1024x1024, you can use this LoRA
https://civarchive.com/models/2540444/anima-highresaesthetic-boost?modelVersionId=2855073

Set your aspect_ratio to custom, and manually enter the size. Setting aspect_ratio to anything other than custom will override your entered values for width and height.

IMG2IMG

Enable the IMG2IMG mode

Load an image and specify a Denoise level. The closer the Denoise is to 1, the more the image will change.

The image will follow the same detailer pipeline and logic as the TXT2IMG one. Detailer prompts will be sent to your initial image prompt, and face/eye detailers will be sent to your body prompt. The results are very consistent and give high prompt adherence.

For this image, I included things like “slit pupils, glowing eyes” in the eye detailer, “evil smile, mischievous, fangs” in the face detailer, and “large breasts” in the body detailer.

Outside of that, I just described the scene through natural language:
night, darkness, candlelight, shadows, moon, dim lighting,

succubus, devil wings, devil tail, horns, succubus outfit, revealing clothes,

Three adult women are sitting on a couch in a living room with a fireplace together. They all have different eye and hair colors. The girl in the center is taller than the other two and her mouth is closed. The two smaller girls are on either side of the taller girl. The two girls have their mouths open revealing fangs.

The denoise was set pretty high since I wanted the image to change quite a bit. You should generally start lower and work your way up (Around 0.5 or 0.6 perhaps)

It’s important to realize that Anima is not an “edit” model. You shouldn’t try to edit an image like Qwen edit. This is simply IMG2IMG.

Edit Mode

The Anima edit LoRA gives you the ability to prompt for specific edits to an image, much like Qwen Edit. Although the prompt adherence may not be as good as Qwen, the LoRA is small and works very quickly! This works without Inpainting as well. https://civarchive.com/models/2650553/anima-edit

1. Load your image in the Edit Mode sections

2. Ensure the path is correct for the Anima Edit LoRA

3. Prompt in natural language and only include what you want to change

Since you should only include what you want to change in the image, I recommend that you DISABLE the Send Detailers to Main Pompt option in the Base Configuration section.

This will ensure only the text you enter into the primary positive prompt will make it to the CLIP text encoder node.

Note that this workflow does not support Anima Edit with inpainting yet. I will plan this for a future release.

Ref2Img Mode **NEW**

Create an image with one or two characters from reference images! No character LoRA needed! This functionality is made possible by ElectronGirl/anima-ref-adapter . This mode also works with dual Ksamplers.
Attach one or two images. If you are using two images, enable 2nd Ref

Select a size for your image

Prompt for the scene!

---------------------------------
For one character images:

- Attach an image of a character reference

- Adjust the size of the desired image

- Prompt for the scene. Example: cowboy shot, sitting, cafe, holding cup, coffee, smile, looking at viewer, ;)

The resulting image should be the character in the character reference performing the scene you described.

Source:

Result:

For two character images:

- Attach two character references. the Image Stitch node will combine them; one on the left, and one on the right.

- Describe in your prompt who is on the left and who is on the right with a few words each.

(Example: the boy on the right has long light blonde hair narrowed light purple eyes, broad shoulders, with a annoyed expression and has his hands at his side, wearing black armor with gold trim, he is large compared to the girl, the girl on the right has teal hair, closed eyes, a big smile, and is hugging the boy)

The prompt seems a bit long but a lot of this is just to modify the pose. The important part here is to just include a very short description of the character on the left and the character on the right. (I like to use hair and eye color.)

Source:

Result:

Upscalers

You have three possible upscalers in this workflow: SDSU, SeedVR2, and RTX Super resolution. Each upscaler increases the resolution by a factor of 2.

The recommended setup for maximum quality is SDSU + SeedVR2

The recommended setup for faster generation is SDSU + RTX Super resolution

Both of these will output a final image of 4096x4096 (for 1:1 aspect ratio). SeedVR2 cleans up the image a bit whereas RTX Super resolution simply upscale pixel by pixel. Please note that the first time you use SeedVR2, the node will download the model and this will take some time (almost 16GB). Subsequent runs will be much quicker.

The default configuration uses the larger model and SDPA.

I personally recommend using sageattn_2 or 3. Note that you will need the sageattention package to use this. That setup is outside of the scope of this guide. (Google is your friend)

If you have a RTX 5000 series card, you can use sageattn_3, which will upscale faster.

If you don’t want to deal with sageattention, you can keep the mode set to SDPA for max compatibility and ease of use.

If desired, you can run all three for a final image of 8192x8192, but I think this is a bit unnecessarily big.

Inpainting

Inpainting is located in the Img2Img section. Be sure to download the Inpainting model on the last page of this guide.

1. Attach your image to inpaint

2. Adjust Denoise (higher values change the image more, lower values remain more consistent with the reference image)

3. Ensure the inpaint model is selected

4. Right click anywhere on the image and select “Open in MeshEditor.”

5. Mask the section of the image you want to be inpainted.

Here’s an example using inpainting to change the outfit of a character after masking the entire character as shown above.

Prompt: masterpiece,best quality,amazing quality, very aesthetic, high resolution, ultra-detailed, absurdres, newest, explicit, score_9, score_8, score_7, 1girl, standing, from side, night, fireworks, full body, honoka \(doa\), street, buildings, festival, hand on hip, smile, looking at viewer, 1girl, red eyes, pink eyes, dango, pink hair, long hair, hair ornament, large breasts, cleavage, pink kimono, fully clothed, one side up, A girl is standing and gazing at the fireworks. The fireworks can be seen in the corner of the image and the girl is to the side of the fireworks. Fireworks illuminate her. She is wearing a pink and white full length kimono.

Regional Prompting

This functionality has been completely rebuilt using Anima Regional Canvas .

https://github.com/ukr8b3g-cmyk/Anima_Regional_Canvas

Preparation: Download the nodes via git or ComfyUI Manager. Download the control net as well and put it in your model_patches/ folder: Sen-sou/Anima-LLLite-Regional-Controlnet

How it works: Draw regions for your characters or objects. Modify the size to match your image generation

Next, input your prompts for each colored region

This allows for up to 5 regions. One more than in previous releases.

Your base prompt for T2I will be pipped to the scene box. Use this to guide the overall composition.

Ensure the ControlNet is selected and adjust strength if needed

A strength of 1 is usually good enough, but you can increase if it if needed. If you find that regional prompting is affecting the composition of your image, you can also set the start precent higher. This will cause regional prompting only to start after a certain part of the image has generated. I usually will set it to 0.100 – 0.200 if I am having composition issues.

There are many ways to use this node . You can simply mask the areas you want the prompt to affect. You can draw a basic pose for it to follow. Here are two examples.

Control Net
You probably already have an idea of what is possible with ControlNets, but you can read up on the one used in this workflow here: https://huggingface.co/kohya-ss/Anima-LLLite

Check out my article if you want more info on using this section
https://civarchive.com/articles/29870

Prompt Crafter

Need a little inspiration? Use the Prompt Crafter section to help you select styles and general concepts for your image.

All text from this section will be integrated into your base prompt

Prompt Generator
Use an LLM to assist with your prompting! I recommend using a jail broken version of Gemma4 or Qwen. By default, the workflow calls for Gemma: https://huggingface.co/Bedovyy/gemma-4-E4B-it-ultra-uncensored-heretic-Comfy/resolve/main/text_encoders/gemma-4-E4B-it-ultra-uncensored-heretic-fp8.safetensors

Place the model in ComfyUI/models/text_encoders/

The system prompt is tuned for the hybrid prompts (tag + natural language) that Anima tends to work best with. There are four specific “modes” for this.

Txt2Txt – Takes or base prompt and enhances it with additional detail
Img2Txt – Interrogate an image and output the text into the prompt for your generation
Txt&Img2Txt – Interrogate an image and use your base prompt to craft a “hybrid prompt” for the generation
(Just attach an image and put something in your base prompt)
None – Input nothing and have the LLM make a prompt up

Feel free to change the system prompt to your liking. This will change the way the LLM modifies your prompt.

Dual KSamplers NEW
This allows you to use two separate models for a single generation to blend the style into the image. This feature can be used in every mode (T2I I2I and Edit

1. Select the model to be used on your second KSampler

2. Adjust the % of Checkpoint 1 value.

- Use a value between 0.1 and 0.99. The value is multiplied by your configured number of steps, and that is how many steps will be taken on the first KSampler. The remaining steps will be processed on the second KSampler

- Example: Steps = 40 | % of Checkpoint 1 = 0.2. (KSampler1 will process 8 steps, KSampler2 will process 32 steps)

Workflow ships at 50% for each KSampler.

Here’s an Example depicting different values.

For Image2Image, your denoise value is calculated first, and then the Dual KSampler calculation runs

Example: Denoise = 0.5 Steps = 40 %of Checkpoint 1 = 0.5
After denoise calculation (steps * Denoise value), I2I will run 20 steps on the generation. The % of Checkpoint 1 will then split the remaining 20 steps between the two models. 10 Steps are ran on model 1 and 10 steps are ran on model 2

Custom Nodes

You should be able to install everything that is needed through the ComfyUI manager. But just in case, I will leave a list of nodes:

Checkpoints: Use any Anima model you want!

Text Encoder (required): https://huggingface.co/circlestone-labs/Anima/blob/main/split_files/text_encoders/qwen_3_06b_base.safetensors

Text Encoder (optional): https://huggingface.co/Bedovyy/gemma-4-E4B-it-ultra-uncensored-heretic-Comfy/resolve/main/text_encoders/gemma-4-E4B-it-ultra-uncensored-heretic-fp8.safetensors (Only needed if you want to use the Prompt Generator)

CLIP: https://huggingface.co/circlestone-labs/Anima/blob/main/split_files/text_encoders/qwen_3_06b_base.safetensors

VAE: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/blob/main/split_files/vae/qwen_image_vae.safetensors

ControlNets: https://huggingface.co/kohya-ss/Anima-LLLite/tree/main

Model_Patches: https://huggingface.co/Sen-sou/Anima-LLLite-Regional-Controlnet/blob/main/anima-lllite-regional-exp-v3.safetensors

Inpainting: https://huggingface.co/kohya-ss/Anima-LLLite/blob/main/anima-lllite-inpainting-v2.safetensors

LoRAs
Anima Edit: https://civarchive.com/models/2650553/anima-edit

Turbo LoRA: https://civarchive.com/models/2560840/anima-turbo-lora

Detailer models

Upscale: https://huggingface.co/Kim2091/2x-AnimeSharpV4/blob/1a9339b5c308ab3990f6233be2c1169a75772878/2x-AnimeSharpV4_Fast_RCAN_PU.safetensors

Hand: https://huggingface.co/Bingsu/adetailer/blob/main/hand_yolov8s.pt

Face: https://huggingface.co/Anzhc/Anzhcs_YOLOs/blob/5029ed6e79052e5c62608161c914a68502ed055e/Anzhc%20Face%20seg%201024%20v2%20y8n.pt

Eyes: https://huggingface.co/Tenofas/ComfyUI/blob/d79945fb5c16e8aef8a1eb3ba1788d72152c6d96/ultralytics/bbox/Eyeful_v2-Paired.pt

Body: https://civarchive.com/models/2201851/personfemaledetection

NSFW Detailer: https://civarchive.com/models/1313556/anime-nsfw-detection-adetailer-all-in-one?modelVersionId=2266294

Extra notes for Anima: Take a look at their Huggingface repo. They have plenty of guidance for using their model. https://huggingface.co/circlestone-labs/Anima

Also, If you want to increase the generation speed, feel free to disable detailers. Anima tends to produce better base images than SDXL, Pony, and Illustrious. There are many times where detailing an image is simply not needed. Perhaps try only leaving the Refine and SDSU Upscale steps enabled, with all other detailers disabled.

If you want SUPER fast generations, Enable Turbo mode, leave your hand and eye detailers on, disable SDSU, and enable RTX Super resolution for your upscale.

This gives you a 2048x2048 image that has some quick detailing.

For all models used in this workflow, as well as in-depth examples on regional prompting, check out the guide article:
https://civarchive.com/articles/29870

Files with SHA256: 997c50f08eb38bb3e2e73e59fa8432a8814a87a2fce306ffa00f5a14150eb377