Multi-character Anima LoRA drew all 4 roles in one scene, 0 of 24 usable
Contents
Update (2026-08-13): Followed up with a 2×2 swap of standing positions and outfit assignments to pin down the foreground-right outfit mix-up → Anima 4-char LoRA put the outfit on whoever stood foreground-right, 12/12
In my previous run stripping the 4-character prompt down, the four trigger words on their own passed 0/3, while P3 with ID anchors, position lines and outfit lines passed 3/3. But lining four girls up against a white background and drawing one scene where each of them does something different are two different problems. So I put Kurara and Kei talking in the foreground, Koharu with her arms crossed on the right, and Kana waving from the back into a single classroom, and generated all 8 on/off combinations of those three actions at 3 seeds each. For the path where the prompt becomes conditioning through Qwen3-0.6B and T5 token sequences, see the run that split the prompt across Qwen and T5.
Four girls in frame is not a pass
On the exploratory runs I did first, I was counting an image as a success when four girls simply stood in front of and behind each other. Four people can be in frame while nobody is talking, no eyes meet, and hands or props are broken, and that is not a finished illustration.
I was also looking at each output and then adding or rephrasing prompt lines. Done that way, there is no telling which condition changed what.
I withdrew the conclusions from the exploratory runs for the time being, and settled on a fixed set of items to score against.
| Axis | Pass condition |
|---|---|
| Head count and separation | Exactly four girls, one each of Kurara, Kei, Kana and Koharu |
| Outfit and action assignment | Outfits, expressions, postures and gestures appear only on the girl they were assigned to |
| Bodies and contact | Hands, feet, crossed arms, gestures and contact with the floor or furniture are not broken |
| Space | Foreground and background, scale, occlusion, ground contact and the door position hold together as one continuous classroom |
| Readable without a caption | You can tell who is doing what without reading a caption |
| Usable as-is | Not just four girls present, but usable as a finished illustration |
A pass required all six axes at once. Separation alone being fine does not make an image a success.
The scene and the generation settings I fixed
The scene was fixed before the experiment started. Kurara at front left and Kei at front center look at each other and talk, Kurara speaking with one hand open, Kei turning her face and upper body toward her to listen. Koharu at front right crosses her arms, watches the two and looks exasperated. Several meters behind, in the open classroom doorway, Kana raises one hand above her shoulder and waves at the other three. Each of the four gets her own uniform specified, and no extra people or props are allowed.
| Item | Setting |
|---|---|
| Environment | ComfyUI on an M4 Mac mini |
| model | anima-base-v1.0 |
| text encoder | qwen_3_06b_base |
| 4-character LoRA | anima-4char-v1_epoch100 |
| Turbo | none |
| sampler | er_sde / simple |
| steps / cfg | 25 / 4.0 |
| Resolution | 1344×768 |
| seed | 42, 1234, 9999 |
I split the scene’s actions into A, B and C and generated all 8 combinations from 000 to 111.
| Factor | Role it requests |
|---|---|
| A | Kana at the back raises one hand above her shoulder and waves at the three in the foreground |
| B | Kurara and Kei face each other, and Kurara talks while gesturing with one hand |
| C | Koharu crosses her arms and looks at the two with exasperation |
The four regions, ID anchors, outfits, classroom, camera and negative prompt were fixed across every condition. Until all 24 images were out, I never edited the prompt, even after looking at intermediate output.
Re-running July’s hero image prompt at 3 seeds
The hero image of the article where I built the 4-character LoRA has Koharu looking back, Kurara gesturing with both hands, Kei with a hand at her mouth and Kana sitting on a desk, all drawn properly.
I pulled the positive and negative prompts out of the original PNG and regenerated them at 3 seeds without changing a character. All three came out with the four girls separated, in different postures, with depth, as one continuous classroom. Once hands, gaze and facial detail were counted, though, none of them was an image I could adopt as a finished illustration.
They were not exact matches to the original either. At seeds 42 and 1234 Koharu picked up Kurara-style stud earrings, and Kana was not as far away as the far background I had specified. This prompt works as a baseline for a complex four-girl picture being possible, but it was not a strict success case.
All 24 images laid out first
Laid out together, the 24 images did not simply get worse as the amount of information went up. Some conditions broke because the actions interfered with each other, and in others adding another action made the conversation pose come out more clearly. Images that satisfied all six axes came to 0/3 in every condition.
| Cell | Main result | Separation | Scene confirmed |
|---|---|---|---|
| 000 | Static, 3 in the foreground and 1 behind | 3/3 | diagnostic |
| 001 | C’s crossed arms and displeased face come out. Gaze is seed-dependent | 3/3 | 3/3 |
| 010 | B alone. The conversation partner gets mistaken, and the gesture shows up on someone else | 3/3 | 1/3 |
| 011 | B+C. The conversation and Koharu’s reaction come out together | 3/3 | 3/3 |
| 100 | A alone. Kana’s wave at the back is stable | 3/3 | 3/3 |
| 101 | A+C. Two roles far apart from each other coexist | 3/3 | 3/3 |
| 110 | A+B. Two images work, one drops to three girls with mixed features | 2/3 | 2/3 |
| 111 | A+B+C. The whole scene as designed | 3/3 | 3/3 |
Poses that separate as silhouettes come out
In 100, with A alone, Kana’s one-handed wave at the back came out 3/3. In 101, A+C, Kana’s wave and Koharu’s crossed arms coexisted 3/3. Nobody disappeared either.
These results do not say that the 4-character LoRA cannot hold a pose. Poses that separate as large silhouettes, and the foreground and background arrangement, did not break at any seed in 100 or 101.
The face-to-face pair comes out stably
With only B on, in 010, Kurara and Kei clearly faced each other in 1 of 3. At seed 1234 Kurara’s conversation partner was Koharu rather than Kei, and Kei was at the back together with Kana.
Add C, though, and 011 gave me the two talking and Koharu exasperated beside them in 3/3. With A added, 110 produced the conversation composition in 2/3.
Here are B alone, B+C and A+B at the same seed 42.
Building the whole scene, including the one watching them and the one waving from the back, pulled the composition together better than writing only the two talking, on some seeds. But 110 at seed 1234 produced three girls, and the one at the back was a single figure mixing Kana’s hair with Kei’s blue ribbon and white thighhighs.
111 got the roles out and did not reach a finished illustration
At all three seeds, 111 let me tell the roles apart without reading a caption. Kurara and Kei face each other in the foreground with Kurara talking with an open hand, Koharu crosses her arms and turns an exasperated face toward them, and Kana is small in the doorway at the back, waving.
These three still carried mistakes in hands, gaze, faces and outfit assignment, and none was adoptable as-is. What they gave me is rough composition candidates and comparable failure data generated from a fixed prompt.
Here is the positive prompt actually sent for 111.
masterpiece, best quality, safe, clean sharp lineart, cel shading, crisp anime style, 4girls, multiple girls, exactly four girls total, only kurara, keichan, kanachan and koharu, wide shot, inside one continuous after-school classroom with a wooden floor, rows of desks, a blackboard, an open classroom doorway and large windows, warm golden light, strong depth. Exactly three girls occupy the foreground. Exactly one girl occupies the background several meters behind them. In the foreground on the left and center, kurara and keichan form a face-to-face conversation pair. kurara, with long rose-brown hair, stud earrings and light makeup, turns her face and upper body right toward keichan, makes eye contact with her, and speaks while gesturing with one open hand. keichan, a blonde girl with a blue ribbon, turns her face and upper body left toward kurara, makes eye contact with her, and listens with a smile. In the foreground on the right, koharu, with short dark hair and red eyes, stands apart from the conversation pair with her arms crossed. She turns her eyes and face toward kurara and keichan, watching them with half-lidded eyes, lowered eyebrows and a clearly exasperated expression. In the background, kanachan, with a brown side ponytail and an ahoge, stands inside the open classroom doorway several meters behind the other three. She appears noticeably smaller because she is farther away and is partially seen past a row of desks. She looks toward the foreground group, raises one hand above her shoulder, and waves to the other three with a cheerful smile. kurara wears a white shirt, a red necktie and a navy pleated skirt; keichan wears a white shirt, a red ribbon bow, a navy pleated skirt and white thighhighs; kanachan wears a white shirt, a red necktie, a grey skirt and black tights; koharu wears a white shirt, a red ribbon bow and a navy pleated skirt. There are no other people in the room.Hands and lower-body outfits were broken
All three were scenes, not four girls just standing there. Whether an image can be adopted as a finished illustration, though, and whether the details came out as written in the prompt, were two different things. The open-hand talking gesture assigned to Kurara either showed up on Kei as well, or on Kei instead of Kurara. The grey skirt and black tights written for Kana were gone in 3/3. Koharu was wearing that same grey skirt and black tights in 3/3, rather than the navy skirt she had been given. At some seeds Koharu also picked up a Kana-style ahoge.
The four faces and the broad roles are drawn apart from each other, and yet the hands mid-conversation and the lower-body outfits came out on someone else.
Whether to score black tights as part of Kana’s identity is its own question. The original character LoRAs are trained without baking outfits into the trigger, so that outfits can be changed at inference time. Getting a different outfit every time when none is specified is by design.
This time I specified all four outfits, and Koharu wore the one written for Kana. Freedom to change outfits and being able to put an outfit on the girl I named are separate, so I scored them separately.
This 4-character LoRA got as far as the composition of a complex scene. Getting everyone’s clothes and fine-grained actions out exactly as specified needs more adjustment.