A single good anime image is a unit test. A character you want to reuse is an integration test: the same spec has to pass after generation, after a structured layout, and after two edits in a row. I ran that integration test on Tensor.Art and on PixAI with Tsubaki.3, PixAI's newly launched anime model, and logged every assertion.
Setup
| Stage | PixAI | Tensor.Art |
|---|---|---|
| Generation | Tsubaki.3, Prompt Helper off | WAI-illustrious v17 checkpoint, platform defaults |
| Reference | image in the prompt box (@image1) |
IP-Adapter (ControlNet) |
| Edits | Tsubaki.3 nodes in PixAI Studio | FLUX.1 Kontext Dev in the Edit tab |
On both platforms, I used only features that are also open to free accounts. Tensor.Art generated two images per run in this setup, since four-image batches need its Pro plan.
The fixture
hair_and_eyes: copper-red short spiky high ponytail, gray-green eyes
outfit: sleeveless navy coverall, yellow exoskeleton, hazard stripes on shoulder yoke
asymmetric_detail: armor plate missing on the rig's left forearm, blue pistons exposed
small_detail: single gold stud, right ear
scene: exoskeleton lifts a cargo container with both mechanical arms
Stage 1: generation
| Assertion | PixAI | Tensor.Art |
|---|---|---|
| hair and eyes | pass | partial (mismatched eyes) |
| outfit and palette | pass | partial (no stripes on yoke) |
| left-forearm panel | pass | fail (right forearm) |
| right-ear stud | pass | fail (dangling, left ear) |
| rig lifts the container | fail (her own hands) | fail (no lift) |
Neither platform passed the scene assertion, and the results diverged on the trait assertions.
Stage 2: structured output
The layout spec was an expression sheet: two rows of three, six named emotions in fixed cells, no text. PixAI's character consistency guide lists a reference image and a repeated description as separate methods, and I passed both.
| Run | Result |
|---|---|
| PixAI, reference + text | 2x3 grid, 5 of 6 emotions in place, visible traits in all cells |
| Tensor.Art, IP-Adapter 1.0 | single blurred figure (this run also had two setup errors of mine) |
| Tensor.Art, IP-Adapter 0.6 | about 11 faces in overlapping cells |
| Tensor.Art, text only | clean 12-cell grid, all emotions present, costume changed |
When I removed the reference add-on and kept the same text, the image came out clean, which points to the reference step in this setup. With one run per setting, I treat that as a lead, not a proof.
Stage 3: edits as patches
I started both edit chains from each platform's own Stage 1 image. In Edit 1, I asked for a new pose and a laugh and listed every trait to keep. In Edit 2, I used the Edit 1 result as input and switched the floodlights off, leaving an orange beacon as the only light.
Think of each edit as a patch. With a good patch, only the requested lines change, and the rest of the file stays as it was.
| Check | PixAI (Tsubaki.3) | Tensor.Art (FLUX.1 Kontext Dev) |
|---|---|---|
| Edit 1 requested change | applied | laugh applied, container step had nothing to act on |
| Edit 1 unrequested diff | pose to kneeling, rig leg became an arm, pistons switched sides | new crouch, both eyes green |
| Edit 2 requested change | lights off, blue shadow | frame darkened to about a third of its brightness |
| Edit 2 unrequested diff | beacon moved, window reshaped, containers became cabinets | almost none |
I got opposite results from the two chains. Tsubaki.3 applied each change and touched extra lines along the way. The Tensor.Art editing model touched a few extra lines and applied less of the requested change. In a third PixAI edit, I asked for a high camera angle, and the container disappeared entirely. That is the point where I would regenerate from the spec instead of patching again.
Run count
I needed four runs on PixAI for the four steps and six on Tensor.Art, with both extra runs spent on the Stage 2 sheet.
Takeaways for your own pipeline
- Write the spec as data and assert on every property after every stage.
- Include one asymmetric detail, since a left-right mix-up is easy to spot.
- Test reference settings on a throwaway run before a structured layout depends on them.
- Diff each edit against its input, and regenerate from the spec once unrequested diffs pile up.
If you want to run the same three stages on your own character, you can start on PixAI.
Top comments (0)