ANIMA Crossover Couple Generation using Regional Sampler

This guide applies to generating the following:

• Characters from the same title

• Characters designed by the same artist (similar drawing style), but from different titles

• Characters from entirely different works

• Generations with or without LoRAs

How DiT-Based ANIMA Improves Couple Generation

Overall, Anima makes it much easier to generate the "reference image (base image)" required for couple generation.

1. Less Concept Bleeding: It is significantly easier to generate couples from the same title without their traits mixing.

2. Better Small-Area Rendering: Wide, two-shot framing has a much higher success rate.

The Limitation: Style Bleeding

What can't Anima solve completely? Characters with drastically different artist styles.

For example, Lina Inverse has large, angular eyes. If you attempt a couple generation involving Lina and Nadia (The Secret of Blue Water), the "Lina eye style" will completely dominate Nadia, forcing Nadia to be drawn with Lina-style eyes. See the example below:

While typical concept bleeding (swapped hair color, eye color, or chest size) is heavily reduced in Anima, style bleeding still cannot be avoided natively.

What are your options

1. Accept the style bleeding. Most character crossovers will naturally blend into a relatively consistent middle-ground style that won't look overly weird for either character.

2. Solve it. That is exactly what this guide is for.

Methods to Solve Style Bleeding

Here are the primary methods for regional generation, ordered by how they interact with the model:

1. Attention-Level Regional Prompting

   • Note: A very recent method for DiT-based models is FreeFuse, though I haven't rigorously tested it with Anima yet.

2. Image-Level Regional Prompting

   • Inpainting, FaceDetailers.

3. Latent-Level Regional Prompting (The Focus of This Guide)

   • RegionalSampler from the ComfyUI-Impact-Pack.

Why Use Regional Sampler?

While latent-level methods like Regional Sampler are slower than standard sampling, they have major advantages:

• It is universally compatible with SDXL and DiT models like Anima.

• It isolates LoRAs cleanly. Because loading a LoRA typically "taints" the base model, RegionalSampler inherently solves this by applying the LoRA solely to the RegionalPrompt.

• It produces far fewer artifacts and harsh seams compared to traditional inpainting or FaceDetailers.

Couple Generation with Regional Sampler Overview

Prerequisite

If you are unfamiliar with the basic node wiring, please read the Impact Pack documentation.

You can build your ComfyUI workflow by referencing the "Regional LoRA" wiring graph at the very bottom of that page.

Please note that the documentation's graph is designed to compare the RegionalSampler against other regional strategies. You do not have to wire every node exactly as shown in their entire graph -- just focus on copying the RegionalSampler branch!

Overview

1. Generate a Reference Image (Base Image).

2. Feed the exact same prompts, sampler, scheduler, steps, CFG, and seed into the RegionalSampler for that reference image.

3. Download your generated reference image and load it into a LoadImage node.

4. Draw a mask in the LoadImage node over the area you want to change. This will be your Region 1.

5. Feed the mask, your LoRA-attached model, your "region prompts", and the same sampler/scheduler into the RegionalPrompt pipeline.

6. Adjust the parameters in the RegionalSampler.

7. Generate and done!

The entire structure functions like a 2-pass workflow, but crucially, it generates the final image entirely within the same latent space.

Note: Because of this 2-pass nature, the process is not trivial and is slower than a typical KSampler.

Crafting the Reference Image

This is where Anima truly shines. The reference image is the base image that feeds into the RegionalSampler. For a 2-character generation, you will encounter one of three scenarios:

Case 1: ANIMA Native (Both Characters Understood)

Anima natively understands both characters. For example, Amy and Suletta are both understood in Anima-Base v1.0. However, Suletta's art style might still dominate Amy's.

1. We simply prompt them in a reference image.

2. Later, we will apply RegionalPrompt to Amy's region in that latent space.

Case 2: Partial Native (Only 1 Character Understood)

Anima natively understands one character, but you need a LoRA for the second. For example, to make a couple image for Kofune Mio and Higa Kana (Okitsura).

1. We prompt a reference image for Kofune Mio and Kofune Ushio as they are natively understood in Anima Base v1.0.

2. Later, we will apply RegionalPrompt along with a (Higa Kana only) LoRA-attached model to Kofune Ushio`s region to swap it to Higa Kana.

Case 3: Non-Native (Neither Character Understood)

Anima doesn't know either character natively. To make a couple image for two entirely unsupported characters, you have two choices:

Option A (Double Proxy - Preferred Method):

1. We prompt a reference image using a structurally similar couple that is natively understood in Anima Base v1.0.

2. Later, we run two separate RegionalPrompt operations (each with their respective LoRA-attached model) to swap both proxies to your target characters.

Option B (Single Base LoRA):

1. We prompt a reference image using LoRA 1 to generate the first character, alongside a proxy for the second character. (Note: Because the base model is loaded with LoRA 1, this proxy will inherently be tainted by LoRA 1's concepts).

2. Later, we apply a RegionalPrompt (using only the LoRA 2 attached model) to that tainted proxy's region to swap them into the second character, effectively overwriting the tainted style.

Strategy for Choosing the Reference Couple

For all three crossover types, the strategy for creating the base reference image remains the same:

1. Prefer Native Generations:

   • Rely on what the model can generate natively before introducing a LoRA.

   • Native generations typically produce much more natural results that take full advantage of Anima's high-quality base aesthetic, without being stiffened or constrained by a LoRA's specific dataset limitations.

   • For example, instead of wiring a LoRA to force Higa Kana into the initial base image, we natively generate Kofune Mio & Ushio pair, and replace Kofune Ushio with Higa Kana later.

2. Minimize Necessary Modifications (The Proxy Method):

   • Because Mio and Ushio are from the same title, you prevent style-mixing issues right out of the gate, ensuring Mio's style is perfectly accurate.

   • Ushio has long blonde hair, which is a great structural match for Higa Kana.

   • I intentionally prompt Ushio to have Higa Kana's traits in the base image tanned skin, yellow eyes).

   • Takeaway: This is a crucial technique!

     • Prompt native proxy characters to have the target character's traits so the RegionalSampler has to do less heavy lifting later.

     • For example, if you leave the proxy character with fair skin in the base image, you will later have to loosen your constraints in the RegionalSampler (such as lowering your base_only_steps or masking a larger area) to force the transition to tanned skin. While that isn't necessarily wrong, aligning the physical traits in the base image early saves you time and makes the regional generation much smoother.

Prompting the Reference Couple Image in ANIMA

To be short, the following is the most effective method I experienced so far to reduce concept bleeding and get a good reference image.

(More real-world examples and prompts can be found on my main portfolio).

Example Base Prompt:

masterpiece, best quality, score_7, highres, absurdres, suletta mercury, gundam suisei no majo, nika nanaura, gundam suisei no majo, @OPTIONAL_ARTIST.

An anime screenshot featuring two sweaty girls standing in front of a futuristic sci-fi vanity mirror.
Both girls are fully clothed, ... [common traits].
The taller girl on the right has short black hair with colored inner hair, and expressive blue eyes.
She is ... [some actions].
The shorter girl on the left has rich dark brown skin, long red hair with low ponytail, ahoge, black hairband, short thick eyebrows, and expressive aqua eyes.
She is ... [some actions].

A loaded interior vanity room indoors, futuristic sci-fi, intricate details, natural light, subsurface scattering, light particles, light rays, heavy chromatic aberration, and prominent film grain.

Key Prompting Tips for DiT:

1. Structured Prompt Flow: A highly effective formula is to place Booru-style tags at the beginning, use natural language in the middle to describe the physical relationship and character details, and prompt the background/environment at the very end.

2. Character Tag Stability: In my experience, explicitly repeating the series tag for each character at the beginning of your prompt (e.g., kofune mio, summertime render, kofune ushio, summertime render) is much more stable than grouping the character names together (e.g., kofune mio, kofune ushio, summertime render).

   • When tags are grouped at the head of the prompt, the model is much more prone to trait bleeding (such as swapping hair lengths), even if you explicitly define their traits later in the natural language section.

   • Repeating the series tag seems to better anchor each character's core concepts, which is particularly helpful for characters with weaker native knowledge in the model (e.g.: Summertime Render).

3. Positional Tags: DiT models react incredibly well to spatial positioning. Using natural language phrases like "on the left/right", "in front of/behind", or "above/underneath" can dramatically reduce concept bleeding for body traits.

4. Avoid Complex Inline Relationships: Writing heavily nested relational prompts is highly prone to concept bleeding when the entire scene grows larger. Breaking the characters down into separate sentences using the positional style (as shown in the example prompt) yields a much cleaner separation of traits. Therefore, for larger scenes, try to avoid complex inline structures like the following (it's fine for simple scenes):

Suletta Mercury from Gundam Suisei no Majo with [traits...] is doing [actions...] to Nika Nanaura from Gundam Suisei no Majo with [traits...] 

1. Punctuation Matters: Use periods. In DiT models, periods act as explicit "hard stops" for the attention mechanism. Using periods to separate distinct concepts, actions, or character descriptions helps the model understand where one idea ends and another begins, which acts as another layer of defense against trait mixing.

Regional Sampler Parameters

Sampler Selection (Crucial)

MUST USE DETERMINISTIC SAMPLER! (For example: euler)

• Using euler a or er_sde won't break

  • But it requires a patch and will somehow change the composition and a lot of things in the final generated image.

  • You don't want to largely change the composition while you are fixing the style conflicts.

• Higher steps won't help, just use what you typically use.

  • I usually stays at 32-33 for euler along with sgm_uniform and cfg 5.0.

• For other parameters not mentioned below, you can safely leave them as is. (inactive)

Main Control Factors in Regional Sampler

These factors affect together to control how successful the "face swapping" operation will be, ordered from top to down:

1. The mask you draw in LoadImage for loading the base reference image.

2. base_only_steps.

3. The LoRA weights and prompts you feed into a RegionalPrompt branch.

4. overlap_factor.

The Masking Strategy

Key Takeaways for Masking:

1. As mentioned previously, choosing or prompting a proxy character that has similar traits (e.g., long blonde hair) in the base image can help you do the regional works.

2. In such case, you won't have to completely mask the entire head or face.

   • You can leave bangs, eyes, and ears unmasked.

   • This makes the constraint preserve the face angle, without having to use ControlNet.

3. The Rule: Mask less to get more constraint to the base image, mask more to get less constraint to the base image.

4. Targeting Extreme Style Bleeding:

   • For aggressive style bleeding issues (like the Lina-Nadia eye problem mentioned at the beginning of this guide), you can apply the rule above by drawing a very tight, constrained mask strictly over the inner top half of the face. This forces the RegionalSampler to overwrite only the conflicting eye style while letting the base image dictate the rest of the head structure.

The following images demonstrate these takeaways in action.

(Note: The masked image shown below was downloaded directly from the LoadImage node for clearer demonstration. While the transparent background looks a bit different from the standard in-editor black overlay, if you load this downloaded image back into the LoadImage node, it behaves identically).

The base_only_steps

This parameter controls how hard your reference image is going to bake into the final generation.

The core concept here is to estimate how many steps it takes for your masked region to reach a "converged state" where its underlying structure is firmly locked in.

• Higher steps: Lets the image reach a deeper converged state before applying the regional prompt, making the masked area look much closer to the base reference image.

• Lower steps: Interrupts the convergence earlier, giving the model more freedom to regenerate the masked area strictly based on what you prompt.

Recommended Values:

• A good start value could be 6 for a total generation steps of 32 for euler.

• If you find the final image for the masked character tends to have different angles, weird framing, or disconnected expressions, you should raise this value to 8 or 10 to force more structural convergence.

• Typically, when your masked "object" is relatively small compared to the entire image, it requires more steps to anchor its structure, meaning you will need a much higher base_only_steps, possibly 12-16, to swap a face in an extreme wide shot.

The Synergy with the Proxy Method:

This perfectly explains why prompting "similar traits" in your base image (mentioned earlier) is so powerful.

• If your base character already structurally matches your target, you can safely use a much higher base_only_steps because the model doesn't need to intervene early to fix major differences.

  • For example, if I leave Kofune Ushio with her native fair skin trait, I will have to lower the base_only_steps to give the model enough time to transition her into Higa Kana's tanned skin.

  • Lowering these steps means the overall composition, character interactions, and fine details could unintentionally shift (for example, generating entirely different hand postures).

  • This is exactly why generating a proxy character with similar traits in the reference image is so critical.

• By keeping the steps high, you preserve the high-quality aesthetics and composition of your native reference image.

• Crucially, if you are feeding a LoRA into the RegionalPrompt branch, a higher base_only_steps gives that LoRA far less time to process, which significantly reduces its ability to "taint" or bleed its specific style into the rest of the image.

Crafting the Regional Prompt

The prompt you feed to the Regional Prompt branch should be simple:

Example Regional Prompt:

masterpiece, best quality, score_7, highres, absurdres, anime screenshot, kana higa, long blonde hair, thick bangs, expressive yellow eyes, hair between eyes, yellowish-brown skin, soft expression, dim lighting, subsurface scattering, chromatic aberration, heavy film grain

Important Rules for Regional Prompting:

1. It does not need to contain framing or angle prompts.

   • For example, if your base image is generated with from above or foreshortening, you don't prompt it into this regional prompt.

   • In accumulation, your masked region will always have the entire image's knowledge like from above and foreshortening, even if you don't prompt it explicitly in this region. (the strength depends on base_only_steps).

   • Create a good reference image with similar traits, and let base_only_steps and your masks do the angle or framing control.

2. Prompt the traits and lighting.

   • Prompt the traits of the character (you can even omit style and quality tags), and prompt the lighting and finish effects you would like to see in that region.

3. If you are using a LoRA in this Regional Prompt, the weights also affect the result:

   • For larger masking regions, you should lower the weight to 0.8.

   • For smaller masking regions (e.g., the character in a super wide shot), you can increase the weight to 1.0.

Adjusting overlap_factor

1. Starts from 10. When you see artifacts, for example, color blinks at certain pixels around the masking region, you increase the region to 16 or 24.

2. If the artifacts still exist with a big overlap_factor, consider redrawing your mask completely.

How I Do Couple Generation

1. I have a ComfyUI workflow dedicated to generating the reference image.

   • A staged "regional prompt" for the target character is automatically written into the metadata of each generated image.

   • The key nodes involved in this workflow are:

     • comfyui-adaptiveprompts and comfyui-adaptiveprompts-extensions (used to load modular prompts saved on disk for a specific couple and scene).

     • comfyui-image-saver (writes the image metadata in an A1111-readable format).

2. I have a second ComfyUI workflow responsible for handling the regional operations.

   • It reads the reference image's metadata and automatically syncs all generation parameters for the RegionalSampler, including the base prompt and regional prompts.

   • The key nodes used here are:

     • comfyui-prompt-reader-node (reads all metadata from the base image)

     • comfyui-impact-pack (provides the RegionalSampler).

3. I typically run the first workflow to generate a batch of 50-100 base images.

   • (My settings: euler, sgm_uniform, CFG 5.0, 33 steps, 1216x832 or 832x1216, Anima-Base v1.0).

4. I curate the results and select the best reference image using IIB (Image Info Browser).

5. I download the chosen reference image and drag it into my second workflow.

6. I draw the mask over the target region.

7. I set up my LoRA weight to 0.8, base_only_steps to 8, and overlap_factor to 16.

8. I generate the first sample and adjust the parameters or mask based on the result.

   • Since we cannot alter the seed, if I get an unwanted expression or structural result, I troubleshoot by doing the following (in order):

     1. Adjust the base_only_steps.

     2. Adjust the LoRA weight.

     3. Adjust the regional prompt (adding new tags or reordering them).

     4. Redraw the mask (shrinking, enlarging, or shifting it).

Conclusion

This workflow gives you precise control over the artwork you truly care about. It is universally applicable with or without LoRAs, works for both SDXL and ANIMA, and produces fewer artifacts than traditional inpainting workflows. However, it is not suitable for batch generation. As you can see, the process is inherently cumbersome; even after automating the parameter synchronization, I still manually curate and match the best reference pairs for every image.

Ultimately, if you aren't overly bothered by minor style conflicts during crossover generations, simply allowing the model to blend the styles naturally is often a perfectly acceptable compromise that saves a lot of time.

(More examples of Regional Sampler anime works and 2-girl prompts can be found on my main portfolio).

🖼️ View the final outputs of this workflow in my latest Civitai gallery: https://civitai.com/posts/29794591

This guide is the result of months of testing and several days of writing to figure out these exact settings. If it saved you some headaches with your crossover generations, consider dropping a like - your support is highly appreciated!

Acknowledgement

• Thanks to the creators of ANIMA.

• Thanks to the ComfyUI-Impact-Pack team for the RegionalSampler.

• The only LoRA used in this guide's examples is Kana Higa by nochekaiser881.