Choose a model by the constraint you need¶
AIsketcher is not a checkpoint leaderboard. Its job is to keep the same
prepare → explore → pick → vary → export → replay contract while a backend
changes. The model choice should therefore describe a capability, not imply
that one model is best for every input.
Current product roles¶
| Role | Current path | Use it when | Important limit |
|---|---|---|---|
| Guided tour | bundled, hash-verified study | you want to understand the workflow immediately | read-only; no new image is generated |
| Fast Edit | FLUX.2 Klein 4B | speed, photo restyling, flexible sketch interpretation, and instruction edits matter | reference-image editing is not Canny and does not guarantee exact line locking |
| Structure Lock | SDXL + Canny ControlNet | preserving a line drawing or replaying an existing study matters more than current-generation edit quality | stable legacy fallback, not the quality showcase |
| Structure candidate | Z-Image Turbo + Union 2.1 Lite | modern Canny, HED, Scribble, or related structure control | not a default until the benchmark below passes; large first download and a separate VideoX-Fun runtime |
| Pro candidates | Qwen Image Edit 2509/2511 | high-end structure, product, character, or identity-sensitive work | 20B-class models; not appropriate for a T4-first experience |
FLUX.2 Klein 4B is the current concrete Fast Edit model. Its official interface accepts a reference image and editing instruction. AIsketcher no longer labels that path as an intelligent Auto router or describes it as strict structure control.
Z-Image Turbo and the Alibaba-PAI Union 2.1 adapter are the leading Structure candidate because their public Apache-2.0 path supports Canny, HED, Scribble, depth, pose, and other controls. The adapter is not yet a native Diffusers preset, and the reviewed runtime payload is too large to silently become a Simple-mode default.
Qwen Image Edit 2509 is the Pro Structure candidate; 2511 is the Pro Consistency candidate. They stay out of the local default until a larger-GPU backend and the same replay contract are validated.
The default-model gate¶
A candidate can replace the current Fast Edit or Structure role only after a reproducible study covers all of the following:
- At least 12 inputs across sparse pencil lines, dense child drawings, product outlines, graphic compositions, people, and ordinary photos.
- Four recorded seeds for every model/input/prompt combination.
- The same reviewed English model prompt, with the Korean source retained separately when translation is involved.
- Blind designer preference, structure similarity, prompt adherence, visible distortion/failure rate, first-download bytes, warm latency, peak VRAM, and cancellation behavior.
- Separate results for T4 FP16 and a BF16-capable L4, A10, A100, or H100.
The winning model must improve both human preference and the constraint it is meant to serve. A higher edge score alone is not enough, and a beautiful image that discards the sketch is not a Structure winner.
Why the package still matters when models change¶
The portable value is the study around the model: explicit seeds, a chosen
parent, controlled variations, model revisions, source and output hashes,
prompt provenance, and strict or compatible replay. Custom local or hosted
backends implement the small Backend protocol and keep those records without
forking the workflow.
If you can run a candidate model, use the model benchmark report form to contribute a comparable result rather than a screenshot without settings.