title: PixelSmile: Release-Bound Expression Editing description: "PixelSmile is a release-bound facial-expression editing project; pin its published human preview, base model, patched runtime, consented source image, and expression review rather than treating benchmark numbers or adapters as general guarantees." category: models tags: [facial-expression, image-editing, lora, qwen-image-edit, provenance, consent, evaluation] aliases: ["PixelSmile"]
PixelSmile: Release-Bound Expression Editing¶
PixelSmile is a research and software project for fine-grained facial-expression editing. Its upstream repository describes a Qwen-Image-Edit-2511 base model, a published human-preview PixelSmile weight, inference and benchmark artifacts, a demo, and released training code. The repository separately lists a future stable model, so each run must be tied to the exact published artifact it uses.
The associated paper reports continuous expression control through textual latent interpolation. It is evidence for the paper's checkpoint, data, and evaluation protocol—not a guarantee that every face, expression, prompt, or downstream integration will behave linearly.
Release contract¶
Before running or evaluating PixelSmile, record:
- repository revision and the model/adapter file digest;
- the exact Qwen base-model revision, installed runtime, and required local patch status;
- source-image authority, consent, allowed edit scope, and protected regions;
- requested expression and tested control range; and
- seed, input/output hashes, reviewer result, and failure status.
The upstream setup contains model-specific dependency and patch instructions. Follow those instructions for the pinned release, then run a small reproducible smoke test before a broader batch. A different Diffusers, base model, adapter, or community node is a new compatibility target, not an implicit fallback.
Facial-editing boundary¶
PixelSmile changes the depiction of a face. It must not be used to infer a person's actual emotional state, medical condition, identity, age, or intent. Use only authorized source images, respect the agreed editing scope, and make the derived result reviewable.
For each output, review separately:
| Review | Question |
|---|---|
| Requested edit | Does the visible expression match the approved request? |
| Preservation | Are identity-relevant appearance, pose, scene, accessories, and protected regions preserved to the agreed scope? |
| Artifact check | Did the edit introduce changed teeth, eyes, skin texture, background geometry, text, or duplicate features? |
| Provenance | Can a reviewer recover source, adapter/base versions, parameters, and approval? |
Visual identity preservation is a release criterion, not a biometric proof. When the edit changes an important source fact or cannot be reviewed, hold the output rather than presenting it as a faithful correction.
Training and reuse¶
The upstream project now publishes training material, but a local training run needs its own data and rights contract. Do not assume that a public paper dataset, a face collection, generated labels, or a released repository grants rights to reuse person images in another product.
Keep training data, annotations, consent/provenance records, base model, adapter, and evaluation split bound together. Validate on held-out, source-disjoint examples and preserve a manual review path for sensitive edits.
The repository is published under Apache-2.0, but code licensing is not a substitute for checking the base model, adapter/model-card terms, community integration terms, or source-image rights.