title: Textual Latent Interpolation: Model-Bounded Control description: "Textual latent interpolation is a model-specific conditioning experiment: preserve non-target inputs, bind it to an exact encoder and adapter, sweep the requested range, and prove controllability and preservation instead of assuming semantic linearity." category: techniques tags: [interpolation, text-embeddings, continuous-control, expression-editing, evaluation]
Textual Latent Interpolation: Model-Bounded Control¶
Textual latent interpolation constructs a conditioning vector between two text encoder outputs. For embeddings e₀ and e₁, a common form is e(α) = (1 - α)e₀ + αe₁. This arithmetic is well-defined; a smooth or semantically meaningful visual transition is not guaranteed by the arithmetic alone.
The technique becomes a usable control only when a particular model, tokenizer, adapter, and training/evaluation setup demonstrate the requested behavior. It is not a universal property of text encoders or diffusion checkpoints.
Bind the experiment¶
For a reproducible run, record:
- the base model, text encoder, tokenizer, adapter/checkpoint digest, and runtime revision;
- two prompts with all non-target language held constant;
- source image, seed, edit mask, and preservation requirements where editing an existing image;
- the requested
αrange and increments; and - a no-interpolation control using the same model and prompt contract.
Changing several words, swapping an encoder, or loading a different adapter changes the experiment. Do not call results comparable merely because both runs use a number named alpha.
What to test¶
Use a small sweep across the requested range rather than selecting a single attractive output. Review each step for:
| Question | Evidence |
|---|---|
| Requested control | a task-specific, predeclared annotation or review rubric |
| Preservation | unchanged protected regions, subject traits, scene structure, and product details |
| Monotonicity | whether the requested visual attribute changes in the intended direction over the tested range |
| Stability | repeated seeds and nearby values do not cause abrupt unrelated edits |
| Boundary behavior | values outside the trained/tested range are labelled exploratory, not normal controls |
For facial images, an expression edit is a requested visual transformation, not evidence of a person's real emotional state. Do not turn an interpolation control into an identity, health, age, or emotion classifier.
Relationship to PixelSmile¶
PixelSmile and its paper describe textual latent interpolation as part of a specific facial-expression editing system. That evidence supports evaluating the released PixelSmile contract; it does not prove that an arbitrary CLIP, Qwen, or other text encoder exposes the same continuous direction.
Keep the model's own expression vocabulary, conditioning path, and release status separate from a generic interpolation article. A different checkpoint requires its own controls, data boundary, and evaluation.
Release boundary¶
Do not promote interpolation to a product slider until the tested model can:
- reproduce the selected range from a pinned configuration;
- preserve the protected attributes defined for the task;
- surface the source, adapter, value, and result for review; and
- fail visibly when the requested control is unsupported or outside its validated range.
An output may be visually plausible while still changing unrelated content. That is a failed preservation result, not proof that the embedding path is useful.