The anime model I tried on Civitai this summer used tag lists, and for hard poses, model-library workflows offer pose-guide add-ons such as ControlNet, which PixAI also offers on its other models. Tsubaki.3, PixAI's newest anime model, takes full sentences and also handles references and edits in the same model. I wanted numbers instead of impressions, so I treated each prompt like a test case: a fixed spec, a batch of four, and a pass count for every requirement.
Setup
- Model: Every run used Tsubaki.3, with batches of four for new scenes and single images for edits.
- Preprocessing: I switched off Prompt Helper, PixAI's automatic prompt rewriter, so each result reflects my own wording.
- Extras: I used no style preset, no Recipe and no LoRA.
The fixture
The character is Sorane, a street dance crew captain I created for this test. Her spec had seven checkable traits:
hair: tangerine bob, undercut on the right side
eyes: teal
face_mark: small black star tattoo under the left eye
jacket: oversized lime track jacket, exactly 3 white chest stripes, off the left shoulder
top: black tank top
legs: charcoal cargo joggers, red bandana on the right thigh
shoes: white high-tops, red laces
In the first batch, eyes, star, tank top, bare shoulder, bandana and sneakers passed 4/4. The stripe count passed 1/4. The undercut side failed 4/4: it landed on her left, next to the star. I kept the image with three stripes as the fixture for every later test.
Test 1: a hard pose from text alone
In the next prompt, I described a two-person breaking battle: Sorane upside down on one hand, a silver-haired opponent mid-jump on the right, a floor-level wide-angle camera, a dark crowd, two crossing spotlights and puddles.
| Assertion | Passed |
|---|---|
| upside-down one-hand pose, knees bent | 4/4 |
| opponent mid-jump, knees tucked | 4/4 |
| left and right placement | 4/4 |
| colors stay with their own character | 4/4 |
| no feature copied to the other character | 3/4 |
| crowd as dark silhouettes | 3/4 |
| spotlight beams crossing | 1/4 |
| shadows falling toward the camera | 0/4 |
The pose and placement passed from text alone. The lighting assertions failed most often.
Test 2: nine assertions in one prompt
I stacked nine requirements into the staircase prompt. I asked for the seven traits, a mid-air jump down the stairs, an open-mouthed laugh, a high camera and a corner-to-corner staircase. I also asked for exactly five cones in a row, a blue "EXIT 7" sign, rain on wet steps, and one yellow streetlamp near the top-right corner throwing her shadow toward the lower left.
| Assertion | Passed |
|---|---|
| jump, laugh, high angle, rain | 4/4 |
| lamp near the top-right corner | 4/4 |
| shadow toward the lower left | 4/4 |
| wet, reflective steps | 3/4 |
| sign with clean "EXIT 7" lettering | 1/4 |
| exactly five cones | 1/4 (counts: 4, 4, 5, 4) |
| staircase running corner to corner | 0/4 |
| lamp as the only light source | 0/4 |
Compare the shadow rows across both tests. In Test 1, I named two spotlights on the pillars but gave them no place in the frame, and the shadow assertion failed 0/4. In Test 2, I named one lamp, its position in the frame and the shadow direction, and it passed 4/4. PixAI's Tsubaki.3 prompt guide recommends naming the light source, and these two runs are consistent with that.
I also tried isolating the failing assertion. On the image with the clean sign, I ran one edit, asking only for five cones in a straight, visible row. The row came out straight and visible, and it still had four cones.
Test 3: the reference image as a fixture
Tsubaki.3 takes a reference image in the prompt box, called @image1 in the text. I passed the fixture and described only the change, following PixAI's own advice: a rooftop at sunrise, then a waist-up studio shot in blue evening light, four images each.
| Check across 8 images | Result |
|---|---|
| star tattoo under the left eye | 8/8 |
| undercut at reference size | 2/8 (grew into a half-shaved head in 6) |
| jacket hue close to reference (studio set) | 0/4 (darker, cooler green) |
| unrequested sleeve stripe from the reference | about half |
The stripe in the last row was never in my prompt, yet it came back in about half of the new images. In my runs, accidental details in a reference were copied like intended ones.
Test 4: edits as patches
I applied two edits in sequence. In the first, I swapped only the jacket for a crimson satin bomber and listed everything to keep. It worked on the first run, with four side effects: the jacket hung open, the tank top became a T-shirt, one hand moved into a pocket, and a drawstring appeared on the joggers. With the second edit, applied to the first result, I moved her into a night scene. The setting, light and shadow direction passed, while the red laces turned white and the drawing style shifted.
What I would tell another developer
- Write scene requirements as sentences, and test each one separately across a batch.
- Make lighting assertions specific: source, position and shadow direction.
- Treat exact counts as unreliable and check them by hand.
- Diff every reference-based result against the reference, including details you never asked for.
- Keep edits narrow and check for side effects after each one, since both of my edits had some.
If you want to run the same kind of test on your own character, start with a written spec and a batch of four on PixAI.
Top comments (0)