One LoRA on 6 Anima derivatives (M1 Max): trigger-only unstable, RDBT runs wild

In my post sorting 20-plus Anima derivative checkpoints, I wrote that since every derivative shares the same base, a single Anima LoRA loads on all of them. This is that, tested for real. I run my own character LoRA — baked for WAI-Anima — through several Anima derivatives at the same prompt and the same seed, and look only at the output differences. ComfyUI on an M1 Max 64GB, with the Turbo LoRA.

The derivatives are six: WAI-Anima (the training base), AnimaYume, RDBT, Kirazuri, Hikari, and AnimaIka. JANIMA was in paid early access at the time of writing, with a free release scheduled about ten days later, so I left it out this time.

The findings up front:

• Even with the same LoRA, derivatives diverge a lot in what they output. The biggest gap shows up with “trigger only,” specifying no outfit or anything else
• With just the trigger word, none of the models keep the character stable. The only full kanachan — side ponytail and all — came from AnimaYume, one shot of three (s200; ponytail on the viewer’s right, matching the training material). WAI-Anima never produced the ponytail across three shots, and RDBT morphs into beast-ears and military gear
• WAI-Anima is the steady type: it reliably puts out brown hair + ahoge every time. But the color runs dark and dim, and the outfit drifts to plain casual wear
• AnimaYume, when it hits, is cleaner and brighter with tidier art, but it only hits once in three shots — not stable
• RDBT has the muddiest color. With trigger only it bolts to beast-ears and military gear
• A heavy prompt that spells out the outfit converges every model on uniform kanachan. But it’s just a simple uniform, so it’s not much of a comparison
• keichan barely shows up with the trigger word alone (only Hikari held the shape). But add structure tags + a uniform and all six converge on uniform keichan

The LoRAs and the derivatives

Both LoRAs are my own characters, trained on top of WAI-Anima. The contrast is in where the training material comes from — that’s the interesting part this time.

Character	LoRA	Training-material origin	Identifying keys
kanachan	kanachan-waianima-rework-v4 ep180	Images generated with Gemini (outside the Anima family)	side ponytail + ahoge + blue scrunchie
keichan	keichan-waianima-v1 ep20	Images generated with WAI-Anima (inside the Anima family)	air intakes + half braid + blue ribbon

kanachan is a LoRA trained on Gemini-generated material fed into Anima, baked from a different distribution than the Anima family. keichan is a LoRA trained only on WAI-Anima output, baked from the same distribution as the base.

Label	Checkpoint	Lineage	Traits
WAI-Anima	waiANIMA_v10	WAI line (reference)	Training base for both LoRAs. Aesthetic already tuned
AnimaYume	animayume_v05	Mainstream fine-tune	One of the most-downloaded, NSFW-capable
RDBT	rdbtAnima_b1V0371	Distilled	Guidance-distilled, runs at CFG 1–4. Requires explicit style prompting
Kirazuri	kirazuriAnima_v20	Large-scale fine-tune	Trained on 35,537 images from preview3-base. Strong signature style
Hikari	hikariAnima_100	Contrarian	Drops quality-tag dependence, film grain
AnimaIka	animaika_v40	Merge	A merge that includes WAI-Anima as an ingredient

Generation settings

Every model uses the same settings. The text encoder (qwen_3_06b_base) and VAE (qwen_image_vae) are shared across derivatives, so they’re fixed.

Setting	Value
Resolution	832×1216
Steps	8
CFG	1.0
Sampler	er_sde / simple
Turbo LoRA	anima-turbo-lora-v0.1 strength 1.0
Seed	42 / 100 / 200 (fixed)
kanachan LoRA strength	model 1.0 / clip 0.8
keichan LoRA strength	model 1.0 / clip 1.0
Generation time	~40 s/image (M1 Max 64GB)

Baseline output (no LoRA): each model’s style

Before applying the character LoRA, let’s see what the six models put out on a generic prompt. Same “blonde hair + white robe + gold embroidery” standing-pose prompt as past posts, no LoRA, Turbo only. Seed 42.

Even on the same prompt, the style and color differ a lot. WAI-Anima and AnimaIka are nearly the same lineage and hard to tell apart (which tracks, since AnimaIka is a merge that includes WAI-Anima). Hikari is full-on moe 2D, the most anime-looking of the six. AnimaYume has a recent-anime kind of shading, soft and pale. RDBT leans realistic and off-anime, with a gothic Western-game standing-portrait texture and muddy color. Kirazuri is realistic-leaning 2D, and its strong signature style shows — it even adds props not in the prompt, like a lantern. AnimaIka sometimes carries a watermark (probably from its training data). This style spread is what drives the color and the direction of the runaway once you apply the character LoRA.

From here I apply the LoRA. Prompts come in two stages: light (trigger word only) and heavy (piling on outfit, hairstyle, color).

# Light prompt (trigger only)
{trigger}, 1girl, solo, standing, looking at viewer, full body, white background

# Heavy prompt (kanachan)
masterpiece, best quality, score_7, safe, 1girl, solo, kanachan,
side ponytail, ahoge, double parted bangs, medium hair, blue scrunchie,
white collared shirt, red necktie, looking at viewer, front view,
white background, simple background, full body, standing

How far trigger-only gets kanachan

This is where the gap is biggest. Give nothing but the single word kanachan — no outfit, no hairstyle — and see what each model produces. kanachan’s identifying keys are a side ponytail (on the viewer’s right in the training material) + ahoge + blue scrunchie.

WAI-Anima (steady but flat)

Brown hair + ahoge come out reliably across all three seeds, but the side ponytail never appears. The outfit is always plain casual — sweater, apron, cardigan. Being the training base, the face and features don’t wobble, but the result is flat and the color sinks a little dark.

AnimaYume (full hit on 1 of 3)

s42 drifted to a kimono, s100 to a T-shirt, and the full kanachan — side ponytail + ahoge + uniform — appeared only in s200 (on the viewer’s right, matching the training material). One in three isn’t stable, but it was the only shot across all models where the trigger alone picked up the ponytail. That hit had clean bright color and crisp lines — high completeness.

RDBT (morphs into beast-ears and military gear)

Across all three seeds it morphs into beast-ears + tail + ragged military gear, unrecognizable as kanachan. RDBT’s CivitAI page says explicit style prompting is required and that without it the style goes random. With only the trigger, this is exactly that unspecified case, and the character gets swallowed by RDBT’s survival-leaning default look. For the record, kanachan’s training material has no beast-ears, no tail, and no RPG-style outfit whatsoever.

Kirazuri

Something ponytail-ish faintly shows in the first shot (s42), but it’s closer to a small bunch tied at the side than a ponytail, and it’s on the viewer’s left — so the orientation doesn’t match the training material. The hair color is also off from the training material’s brown, and while ahoge shows in two shots, it’s far smaller than in the training material. On top of that, s100 and s200 drift into cat-ears + tail beast-girls, with muddy, dark color. Two of three being something else puts it close to RDBT; as-is, getting kanachan out is rough. The rendering also feels a bit over-the-top — even with a fair amount of spec, it sometimes throws in poses I didn’t ask for. It’s an impression from only a few shots so I can’t be sure, but it looks like a Kirazuri quirk.

Hikari

Brown hair + ahoge + casual wear, no side ponytail. There’s no beast-girl runaway, but the face shape turns rounder than the other models. As intended for film grain, the whole thing is slightly muted and soft.

AnimaIka

Brown hair + ahoge + casual wear — the closest output to WAI-Anima of the six. That makes sense, since the merge includes WAI-Anima v1. No side ponytail, and at s200 it drifts to cat-ears + tail (probably from the Cat Tower ingredient, also in the merge).

Lining up all six, the only trigger-only shot that produced the side ponytail was AnimaYume’s s200. The rest are either the WAI-Anima pattern (“brown hair + ahoge + casual, no ponytail”) or, like RDBT / Kirazuri / AnimaIka, drift toward beast-ears — nobody reliably produces kanachan.

Spell out the outfit and they converge (but the color gap remains)

Add a heavy prompt piling on uniform, hairstyle, and color, and every model converges on uniform kanachan. Even RDBT, which had morphed into beast-ears, comes back properly.

Kirazuri and AnimaIka, which had drifted to beast-girls, also come back to uniform kanachan once the outfit is spelled out.

The outfit it produces is a simple uniform — white shirt + red necktie + skirt — and at that level any model can do it once you specify it. But the side-ponytail orientation isn’t stable. Even in these six, Hikari and Kirazuri put it on the viewer’s left, not the training material’s right. Orientation wobbling with the seed showed up in past WAI-Anima testing too; even a heavy prompt doesn’t snap to the correct side in one shot. Hikari and Kirazuri are also the only two whose face shape looks a little different from the other four, so maybe they’re a different lineage somehow, but I can’t tell why.

Convergence itself is a given, so what actually differs here is the color. At the same seed and the same uniform, WAI-Anima runs darker with heavier blues and shadows, sinking a bit dim, while AnimaYume’s shading is lighter and brighter. RDBT has the muddiest, most sunken color of the three. A big part of why AnimaYume “looks pretty” is that light, clean brightness.

From here it’s speculation. kanachan’s LoRA is trained on Gemini-generated images, but that material was deliberately pulled away from existing anime shading toward a slightly watercolor-ish palette (not a point about Gemini’s default style — I mean I steered the training material that way). Because that material’s color is offset from the anime shading Anima is good at, WAI-Anima tries to pick up that color and sinks dark, while AnimaYume re-aligns toward Anima’s own anime style and goes brighter and cleaner — that’s my read. I can’t be sure how much of Anima’s style is baked in at LoRA-training time, so there’s no proof, but the trend in the output fits the hypothesis.

keichan (add structure tags + a uniform and all six produce her)

keichan’s identifying keys are air intakes (forward-facing scoops on both sides of the bangs) + blunt bangs + half braid + blue ribbon.

Hikari is the only one that keeps the shape from the trigger alone

With just keichan, the only one of the six that produces a recognizable keichan is Hikari. And Hikari holds nearly the whole shape — blunt bangs, the braid, and more — from a single trigger, which honestly surprised me. The other five fall apart from the blonde-hair-and-face stage; with the trigger alone they don’t become keichan. In the keichan LoRA-training post and the post on compositing two characters into one image I wrote that the trigger bakes in hair color, face, and body type but not the hairstyle structure — but moving to these derivatives breaks even that premise, and outside Hikari, even the hair color and face don’t stabilize.

The five models other than Hikari drift to the base’s default hairstyle with the trigger alone, and the air intakes don’t appear. Since you can’t leave the identifying keys to the trigger, these need the keys spelled out as separate tags. Compared with kanachan, where the trigger alone at least picked up ahoge + brown hair, keichan — baked lightly (ep20) on self-generated material — doesn’t ride fully on the trigger for the hairstyle, and Hikari pulling the whole thing is more the exception.

Add structure tags + a uniform and all six produce her

Add the structure tags established in past posts (blonde hair, long hair, blunt bangs, long sidelocks, half updo, braid, blue ribbon + a natural-language sentence describing the intakes) plus a uniform (white shirt + red ribbon tie + dark navy skirt + white knee-high socks), and all six converge on keichan in blonde hair + blunt bangs + blue ribbon + uniform.

The training base WAI-Anima of course, but also AnimaYume, RDBT, Kirazuri, Hikari, and AnimaIka all produce keichan. Spell out the uniform and even RDBT comes out normally instead of morphing into military gear. The differences between models ride only on the base style (Kirazuri’s hair is a bit wavy with softer lines, Hikari leans moe, and so on); the character reproduction itself holds on any model. The strength of the air intakes varies a little by model, but keichan’s skeleton — blunt bangs + long sidelocks + blue-ribbon half-up — holds on every model.

Where kanachan’s side ponytail wobbled in orientation even on a heavy prompt, keichan’s hairstyle is left-right symmetric, so there’s no orientation problem, and spelling out the uniform lands her cleanly. That said, the air intakes are generally on the small side except on Kirazuri, and I can’t tell whether that’s just variance without generating more samples. But the original training post didn’t include many big-intake diagonal-angle shots either, so this much is probably about what you’d expect.

Which one produces the LoRA’s character most obediently

At first I thought the training base WAI-Anima was the most obedient, but on a second look, the meaning of “obedient” splits.

• Stable and predictable (same face and features every time): WAI-Anima. Being the training base, it doesn’t wobble. But the color is dark and the art is flat
• Densest, prettiest hits: AnimaYume. It had the one shot that even picked up the side ponytail, and the color is light and bright. But it’s one hit in three — not stable
• RDBT presumes explicit style prompting; with the trigger alone it carries the character off into its own default look. It also has the muddiest color
• The only model that held even the full shape from a single trigger was, surprisingly, Hikari. The whole image comes out paler and rounder than the original character, so it’s not a full reproduction, but even keichan — weakly baked — came out from one trigger. It might suit a LoRA that turned out so-so when you just want the character in one shot. kanachan, baked from non-Anima Gemini material, comes out the same way, so it doesn’t look like ease-of-output is decided by whether the training material is Anima-distribution (though on kanachan the face skews round and the side ponytail doesn’t appear)
• Kirazuri leans RDBT-like: with kanachan’s trigger alone, two of three shots drift to beast-girls, and the hair color doesn’t match the training material. A heavy prompt brings the uniform back, but it’s hard to call obedient
• AnimaIka has the closest output to WAI-Anima and leans stable. With the trigger alone it doesn’t produce the side ponytail and sometimes drifts to cat-ears. No particular complaints, but this close, you start to wonder whether you might as well just use WAI-Anima — its use case honestly isn’t clear. If anything, maybe for when you also want beast-ears

Subjective, but the best-looking output came from AnimaYume. The density and brightness of that one hit beat the rest. That said, if you want it stable, you end up needing to pile on the prompt anyway.

RDBT doesn’t fix even with Turbo removed

RDBT is a distilled model, so stacking the Turbo LoRA doubles up the distillation. To isolate whether the breakage is Turbo’s fault, I fired the light prompt at bare RDBT with Turbo removed too, but the result was the same — still broken. Where the Turbo version drifts toward military gear, the no-Turbo version keeps the beast-ears and leans toward a swimsuit (it gets revealing, so I’ll skip the image). The direction differs, but the beast-ears not disappearing and the character getting swallowed is the same — the main cause of the breakage is RDBT itself, not Turbo.

Related posts

• Anima derivative checkpoints: 20+ sorted by type and aim
• Running Anima-Base v1.0 on M1 Max vs WAI-Anima: LoRA compatibility and 5.7× Turbo LoRA speedup
• Training an Anima original-character LoRA on Anima-only images: converged at ep20
• Pushing a WAI-Anima character LoRA to the official 12,000 steps hurt control; half that at ep150 hit 100%