Most comparisons of civitai alternatives score the wrong variable. They count models, or compare library sizes, or rank features. None of that tells you how long it takes to get from an empty prompt box to a usable anime image, which is the thing most people switching are trying to shorten.
So this is written up as an experiment rather than a review. One hypothesis, stated controls, one confound I chose not to remove, and a section at the end on what the results can and cannot support.
Hypothesis
A natural-language DiT model with no configuration will produce usable anime output, follow multi-part instructions, and support in-place editing more reliably than a general-purpose model with no configuration.
The system under test is Tsubaki.3, the newest anime model on PixAI, which takes plain sentences rather than tag lists. The comparison system is CivitAI.
Setup and controls
Identical prompts. Every test used the same prompt text on both platforms, word for word.
No configuration on either side. No LoRAs, no negative prompts, no tuned settings.
The confound, stated explicitly: CivitAI ran on DreamShaper, its default model. DreamShaper is a general model, not an anime checkpoint. A tuned anime checkpoint with appropriate LoRAs would perform very differently. This test measures the out-of-box experience only, and every result below should be read with that in mind.
Cost, recorded but not normalised. A single Tsubaki.3 generation cost 5,500 credits and a batch of 4 cost 10,000. A CivitAI generation cost between 3 and 10 blue buzz, and an edit cost around 4. Those are separate currencies with no stable exchange rate, so I've kept them as separate budgets.
For prompt-writing specifics on the model under test, the Tsubaki.3 prompt guide covers the syntax.
Test subject
One character, Rei, a conductor on a night train that carries spirits. Six anchors:
- Waist-length silver-white side braid with a red cord
- Midnight-blue conductor's coat with brass buttons
- One fingerless black glove, left hand only
- Brass ticket punch at the right hip
- Pale scar through the right eyebrow
- Amber eyes
Anchors 3 and 5 are asymmetric on purpose. Side-specific details are where drift shows first, so they act as sensitive indicators.
Evaluation criteria
Five properties, one or more tests each:
- First-generation anime quality with nothing added
- Prompt comprehension of sentences rather than tag syntax
- Multi-subject composition with several characters doing distinct things
- Character continuation across new scenes
- In-place editing that changes one element and preserves the rest
T1: baseline character generation
Method: new generation on both platforms.
A cinematic modern anime illustration. A young woman stands alone on a night
railway platform, waiting. Waist-length silver-white hair in a low side braid
with a red cord woven through it. A midnight-blue conductor's coat with brass
buttons and a high collar, worn open over a grey turtleneck. One fingerless
black glove on her left hand. A brass ticket punch hanging on a chain at her
right hip. A pale scar through her right eyebrow. Amber eyes. She holds a
signal lamp low in her right hand and looks down the empty track away from
the camera. Paper tickets drift past her ankles in the wind. Cold blue
platform lights above, warm lamp glow below, fog rolling along the rails.
Full body, low angle.
PixAI dashboard (Image generation screen
CivitAI dashboard (Image creation screen)
Tsubaki.3: 4 of 6 anchors. Braid, red cord, open coat over the turtleneck, glove on the correct hand, amber eyes, pose and both light sources all rendered. Cold blue fell on the roof structure and her coat, warm amber from the lamp on her boots and the platform tiles. Both hands resolved cleanly, including the one gripping the lamp handle. Misses: the scar was absent and the ticket punch came back as an indistinct shape.
DreamShaper: 0 of 6 anchors. Short braid with no red cord, a coat buttoned tight with a belt, no glove, no ticket punch, no scar, dark eyes rather than amber. It substituted a handbag for the signal lamp, turned her to face the camera, and dropped the tickets, the fog and both light sources. Style output was semi-realistic rather than anime.
Left: Tsubaki.3. Right: CivitAI on DreamShaper.
T2: multi-subject composition
Four passengers, four distinct named actions. This is where instructions typically start dropping.
Method: new generation on both platforms.
A cinematic modern anime illustration, interior of a night train carriage
carrying spirits. The conductor stands in the centre aisle with her back
three-quarters to the camera: waist-length silver-white braid with a red
cord, midnight-blue coat with brass buttons, brass ticket punch at her hip,
one fingerless black glove on her left hand. She is punching a ticket held up
by a passenger. Four translucent passengers fill the carriage, each doing
something different. An old man in a farmer's jacket sleeps against the
window with his hat over his face. A schoolgirl kneels backwards on her seat
facing away, drawing a shape in the fogged glass with one finger. A thin man
in a suit stands holding the overhead strap, reading a folded newspaper. A
small boy crouches on the floor, chasing a paper ticket that is sliding along
the aisle. Warm ceiling lamps inside, cold blue night rushing past the
windows. Wide shot down the length of the carriage.
Tsubaki.3: 4 of 4 actions. The schoolgirl kneels backwards drawing a spiral on the fogged glass. The suited man grips the overhead rail with one hand and his newspaper with the other. The boy is down in the aisle, grabbing at a ticket as it slides away. The conductor retains braid, coat, punch and glove at a three-quarter turn. Partial miss: 3 of 4 passengers translucent. The old man rendered solid, with his hat on his head rather than over his face.
Test 2:Tsubaki.3 result.
DreamShaper: 1 of 4 passengers present. The sleeping old man, the schoolgirl and the crouching boy were all absent. Every figure rendered solid, the conductor's anchors were missing, and the carriage came out as a photorealistic subway rather than an anime spirit train.
Test 2: CivitAI result.
T3: compound instruction following
Eight discrete requirements in one prompt, including two exact counts, one exclusion, and a physical property that distinguishes the living character from the dead ones.
Method: new generation on both platforms.
A cinematic modern anime illustration inside a night train carriage. The
conductor sits alone on a bench seat at the far end, facing the camera
directly: silver-white side braid with a red cord, midnight-blue conductor's
coat, one fingerless black glove on her left hand, pale scar through her
right eyebrow, amber eyes. She is the only living person in the carriage and
her breath fogs visibly in the cold air. Three translucent spirit passengers
sit along the same bench, and none of them breathe out any visible air.
Exactly three paper lanterns hang along the ceiling above the aisle, no more
and no less. Every spirit passenger is looking out of the window, and only
the conductor looks at the camera. Her signal lamp rests switched off on the
seat beside her. Low warm light from the three lanterns, cold moonlight
through the windows. Symmetrical one-point perspective straight down the
carriage.
Tsubaki.3: 6 of 8. Passes included the semantically hard requirements: her breath fogs while the spirits' breath does not, which requires parsing the instruction as a living/dead distinction rather than a lighting effect. Each spirit is turned towards the window, with the conductor alone facing the lens. The signal lamp rests unlit beside her. One-point perspective runs straight down the aisle with clean vanishing lines.
Test 3: Tsubaki.3 result 6 of 8.
Both count requirements failed. Three lanterns requested, four rendered. Three spirits requested, four rendered. Exact counts are the consistent failure mode and should be treated as unreliable in any prompt that depends on them.
DreamShaper: 3 of 8, with one pass by accident. No spirits were generated at all, so "no spirit breath" passed by default. The ceiling filled with dozens of lanterns, the conductor stood in the aisle rather than seated, her breath was absent, and the signal lamp was dropped. Perspective and the single figure facing camera were real passes.
Test 3: CivitAI result, 3 of 8.
T4: continuation under a lighting change
Night to daylight, T1 image supplied as reference.
Method: reference-based, T1 image attached.
The same conductor from the reference image, now off duty in the early
morning. She sits on the steps outside a shuttered station cafe with her coat
unbuttoned and folded over one arm, holding a paper cup in both hands,
laughing at something off frame. Keep her face, the pale scar through her
right eyebrow, the amber eyes, the silver-white side braid with the red cord,
the fingerless black glove on her left hand and the brass ticket punch at her
hip. Pale flat daylight, wet pavement, empty street. Medium shot.
Tsubaki.3: 6 of 6 anchors, including both that failed in T1. The scar appears through the right eyebrow. The glove stays on the left hand, which is the upper hand on the cup, with the right hand bare beneath. Braid, cord, amber eyes and ticket punch all match. Single miss: the coat stayed on her shoulders rather than folded over one arm.
Left: Tsubaki.3 daylight. Right: CivitAI daylight.
T5: continuation under a camera reversal
The harder continuation case. Camera turned to face the subject.
Method: reference-based, T1 image attached.
The same conductor from the reference image, photographed from the far end
of the track looking back towards the platform, so the camera now faces her
instead of following her. She has raised the signal lamp to shoulder height
and is leaning forward slightly, calling out to someone behind her on the
platform. Close framing from the chest up. Keep her face, the scar through
her right eyebrow, the amber eyes, the silver-white braid with the red cord
and the high collar of the midnight-blue coat. Lamp light strong on her face
from below, blue platform light behind her, fog. Close shot.
Camera placement succeeded. The viewpoint dropped onto the track, reversed, framed chest-up, and underlit the face from the lamp while holding cold blue and fog behind. The hand on the lamp handle is clean.
Both asymmetric anchors mirrored. The braid moved to her right shoulder and the scar to her left cheek. This is the standard reverse-angle failure: the reference image carries the feature but not its handedness once the viewpoint flips. The mitigation is to restate the side explicitly in the new prompt. The character consistency guide documents the reference workflow in more detail.
DreamShaper lost identity in both T4 and T5. In daylight: bare hands, a black belted coat, no scar, standing against a doorframe instead of seated on the steps. In the reversal: twin pigtails in place of the single braid, a street light in place of the signal lamp, semi-realistic style throughout.
Left: Tsubaki.3 reverse angle. Right: CivitAI reverse angle.
T6: in-place editing
Two sequential edit instructions applied to the T1 image on each platform.
Method: edit, both applied to the T1 image.
Edit A: garment swap.
Change her midnight-blue conductor's coat to a heavy charcoal winter greatcoat
with the same brass buttons and high collar. Everything else stays exactly as
it is: her face, the scar through her right eyebrow, the silver-white side
braid with the red cord, the fingerless black glove on her left hand, the
brass ticket punch at her hip, the signal lamp in her right hand, her pose,
the camera angle, the drifting tickets, the fog and the lighting.
Tsubaki.3: target changed, all else preserved. Charcoal grey with a heavier material read, brass buttons and high collar intact. Face, braid, cord, glove, lamp and punch identical to the source. Camera angle, tickets, fog and both light sources unchanged.
Edit B: environment swap.
Change the weather to heavy falling snow settling on the platform and on her
shoulders, and switch the platform lights from cold blue to warm amber. Keep
her, her greatcoat, her braid, the glove, the ticket punch, the signal lamp,
her pose, the camera angle and the track behind her exactly as they are.
Tsubaki.3: environment changed, subject preserved. Heavy snow through the frame, platform lighting shifted from cold blue to warm amber, every character anchor plus pose and background geometry locked to the source.
Two procedural notes. Edit B was applied to the original T1 image rather than to the Edit A output, so the coat remains blue in that frame. Snow accumulated on exposed ground and track rather than on her shoulders, consistent with the platform canopy overhead.
PixAI Tsubaki.3: Left: original. Middle: after the coat edit. Right: after the weather edit.
CivitAI: regenerated rather than edited, both times. Edit A changed the coat to charcoal, then produced a different face with no scar, replaced the braid with short white hair, dropped the right arm and the signal lamp, removed the ticket punch, and replaced the track with generic fog and a gas lamp. Edit B repeated the pattern, adding snow and warm light to what was by then a different character.
CivitAI: Left: original. Middle: after the coat edit. Right: after the weather edit.
Results summary
| Tsubaki.3 on PixAI | CivitAI on DreamShaper | |
|---|---|---|
| Setup before first image | None | None, but the default model isn't anime |
| Prompt style | Plain sentences | Responds better to tags and negative prompts |
| Anime style out of box | Yes, every test | Drifted to semi-realistic in every test |
| Multi-subject scenes | 4 of 4 actions landed | 3 of 4 passengers dropped |
| Counting instructions | Failed both counts | Failed both counts |
| Character continuation | 6 of 6 anchors in daylight | Lost identity both times |
| Editing an existing image | Changed one element, kept the rest | Regenerated a new image |
| Model choice | A handful of in-house models | Thousands of community models |
The largest effect was in editing. Tsubaki.3 treats an edit instruction as a delta on the current image, while DreamShaper treated each instruction as a fresh prompt.
Threats to validity
This section matters more than the scores.
The comparison model was not an anime model. DreamShaper is general-purpose. Pitting it against an anime-specialised model on anime prompts guarantees a lopsided result, and a tuned anime checkpoint would close much of that distance.
CivitAI's core value proposition was switched off. The library of community checkpoints and LoRAs is the entire reason to use it. Testing it with none loaded measures the platform in its least representative state.
Single runs per test. No test was repeated for variance, so individual scores carry noise.
Cost cannot be compared directly. Credits and blue buzz have no stable exchange rate.
Conclusion
The hypothesis holds within its stated scope. As a civitai alternative for anime art, Tsubaki.3 produced anime on every generation with zero configuration, parsed plain sentences, composed four characters with four distinct actions, carried a character into daylight with all six anchors, and performed two edits without disturbing the rest of the image.
Its failure modes are specific and reproducible: exact counts fail, asymmetric features mirror on camera reversal, and small accessories render soft on a first pass.
The result is real but narrow. It measures the distance between opening a tab and getting a usable anime image. That distance is short on Tsubaki.3 and longer on CivitAI until configuration is done.
Among sites like CivitAI, this makes the model a reasonable pick for anime-focused creators who want output without setup, work in sentences, compose busy scenes, reuse characters, or edit rather than reroll. It is a poor alternative to CivitAI for anyone who needs a specific community model or LoRA, works across styles beyond anime, already has a tuned setup, trains and shares models, or values library breadth. None of those needs were tested here.
A larger sample of Tsubaki.3 outputs covers more of the model's range if you want more evidence before choosing.
Top comments (0)