I wanted to know whether an anime image generator could keep a character's small details intact across a whole workflow, so I treated the character like a test fixture and ran the same suite on two platforms: PixAI with Tsubaki.3, and SeaArt with a popular community anime checkpoint plus SeaArt's own reference and editing tools. On both, I only used features that free accounts can also use.
I expected to separate the two platforms by image quality. Instead, I kept tracing failures back to one question: how does each prompt format bind attributes to the right subject?
The fixture
The character is Aldric, a clumsy wizard who cooks for an adventuring party. His spec:
character: Aldric
hair: ginger, short ponytail
beard: short, ginger
eyes: pale blue
robe: mustard yellow, hood down, sleeves cut at the elbow
belt: wide, brown, exactly 3 wooden spoons
prop: crooked wooden wand
legs: cream trousers, rust brown boots
The spoon count is the assertion I trusted most. A face can look right by accident, while three spoons are either on the belt or missing.
The suite
I fixed the budget before the first run: one batch of four images per new scene, one allowed prompt revision for the whole suite, and up to two tries per edit step.
| Case | Input change | Assertion |
|---|---|---|
| A. Create | baseline | all spec fields present |
| B. Crowded scene | second character, eight requirements | every attribute on the correct person |
| C. Reuse | two new scenes from a reference image | spec fields unchanged |
| D. Edit twice | outfit change, then setting and light | change applied, everything else unchanged |
One spec, two serializations
Tsubaki.3 reads plain sentences. The checkpoint on SeaArt expects comma-separated tags. So I wrote the same spec in two formats.
PixAI: "...a wide brown leather belt with exactly three wooden
spoons tucked into it, cream trousers, and rust brown boots..."
SeaArt: "...wide brown leather belt, three wooden spoons tucked in
belt, cream pants, brown boots..."
Results
| Assertion | PixAI, Tsubaki.3 | SeaArt, community checkpoint |
|---|---|---|
| 3 spoons, case A | 4 of 4 | 0 of 4 |
| Crooked wand, case A | 4 of 4 | 0 of 4, walking staff instead |
| Attributes on the correct person, case B | 4 of 4 | 1 of 4 |
| 3 spoons, case C | 5 of 8 | 1 of 2 |
| Edit 1 applied, rest unchanged | yes, first try | yes, first try |
| Edit 2 applied | yes, relit for the new scene | background changed, lighting unchanged |
Case C on SeaArt has two images instead of eight because I ran its Character Reference with one image per run.
Why tag lists leak
SeaArt had the most trouble with case B. In three of the four SeaArt images, the model mixed up the two characters: one image split Aldric into a gray-bearded man and a ginger-haired woman in his robe, and another put the robe on a black-haired stranger.
As an analogy, think of a tag prompt as one flat array with no object boundaries:
// tag prompt: every attribute in one list, no owner
["1boy", "1other", "ginger beard", "mustard yellow robe",
"black bob hair", "grey hooded cloak", "holding empty wooden bowl"]
// sentence prompt: attributes scoped to their subject
{
wizard: { beard: "ginger", robe: "mustard yellow" },
companion: { hair: "black bob", cloak: "gray", holds: "empty bowl" }
}
In the flat version, the model has to infer which attribute belongs to which subject. In the scoped version, I answer that question in the prompt itself. This is a mental model for writing prompts, and I am not describing either model's internals.
The one revision
Tsubaki.3 still missed two requirements in case B, in all four images: the cauldron was supposed to float above the fire, and noodles were supposed to hang from Aldric's beard. The Tsubaki.3 Prompt Guide recommends describing spatial relationships in full sentences, so I gave each missed detail its own sentence and left everything else untouched:
The cauldron hovers in mid-air by magic, with a clear gap of empty air
between its round bottom and the tips of the flames, and nothing holds it up.
Both assertions went from 0 of 4 to 4 of 4. One unrelated assertion regressed: the three carrot slices on the cave wall appeared in 2 of 4 images before the change and in 0 of 4 after it. I logged that as a side effect of the rewrite.
Reuse copies state you did not ask for
For case C, I gave PixAI the first image as a reference in the prompt box and described only the new scene. Face, beard and bare head matched in all eight images. A leather pouch that I never specified in case A also appeared in seven of the eight, because it was in the reference image.
SeaArt's Character Reference showed the same behavior in a different form. In the reading scene, it copied the reference background, chests and fire included. In the river scene, the belt count went from three spoons to five or six.
I now treat a reference image as a snapshot of state. Everything in it is input, including accidents.
The edit chain as a pipeline
Case D had two steps: add a stew stain to the robe, then move Aldric to a sunset hillside. I used the stain as a canary value. If step 2 had used the original image instead of the result of step 1, the stain would have disappeared.
On PixAI, I linked the steps as nodes in PixAI Studio, with Tsubaki.3 in every node. The stain stayed through step 2, and Tsubaki.3 relit Aldric for the sunset, with a small drift: a new smear on his cheek and a slightly different wand.
On SeaArt, the stain also stayed, and Aldric came through step 2 almost pixel-identical. The lighting on him did not change, so he looks out of place in the new scene. SeaArt also returned the first edit at a smaller size than the input, 864 pixels wide instead of 1152, and its reference and editing windows have no negative prompt field.
What I would reuse
- Write each subject as its own sentence block when the model reads sentences.
- Put one countable detail in every character spec. With it, "looks about right" becomes a pass or fail.
- Use a canary value in multi-step edits to confirm which image each step used as input.
- Check the reference image for accidents before reusing it.
If you want to run the same suite on your own character, you can start on PixAI.
Top comments (0)