Textual Latent Interpolation¶

Technique for continuous attribute control in diffusion models by interpolating between text embeddings. Enables smooth transitions between states (e.g., neutral → happy) with an intensity parameter alpha. Core mechanism behind [[PixelSmile]]'s expression control.

Principle¶

Same idea as word2vec arithmetic ("king - man + woman = queen") but in the text embedding space of a diffusion model's text encoder.

# Direction vector in embedding space
e_neutral = text_encoder("neutral expression")
e_target  = text_encoder("happy expression")
delta = e_target - e_neutral

# Controlled generation at any intensity
e_conditioned = e_neutral + alpha * delta
# alpha=0 → neutral, alpha=1 → full expression, alpha>1 → exaggerated

Implementation Strategies¶

Strategy 1: All-token interpolation (PixelSmile default)¶

Interpolate across ALL token positions in the embedding sequence. Simplest, works when prompts differ only in the attribute word.

# score_one_all method
e_cond = (1 - alpha) * e_neutral + alpha * e_target
# Equivalent to: e_neutral + alpha * (e_target - e_neutral)

Strategy 2: Tail-token interpolation¶

Only interpolate the last N tokens (where the attribute-specific words are). Preserves the shared context tokens exactly.

# Interpolate only last 6-7 tokens
e_cond = e_neutral.clone()
e_cond[:, -7:, :] = (1 - alpha) * e_neutral[:, -7:, :] + alpha * e_target[:, -7:, :]

Strategy 3: Direction projection¶

Compute the delta, project it to a subspace, apply scaled:

delta = e_target - e_neutral
delta_norm = delta / delta.norm()  # unit direction
e_cond = e_neutral + alpha * magnitude * delta_norm

Multi-Attribute Blending¶

Extend to pairwise combinations (PixelSmile's expression blending):

e_angry = text_encoder("angry expression")
e_sad   = text_encoder("sad expression")
e_neutral = text_encoder("neutral expression")

delta_angry = e_angry - e_neutral
delta_sad   = e_sad - e_neutral

# Blend two emotions at different intensities
e_cond = e_neutral + 0.7 * delta_angry + 0.5 * delta_sad

9 of 15 basic-expression pairs produce plausible results. Failures when attributes are physiologically contradictory (conflicting muscle states).

Requirements¶

1. Aligned embedding space: the text encoder must produce embeddings where linear interpolation is semantically meaningful. VLM-based encoders ([[Qwen 2.5 VL]]) work better than CLIP for this.

2. LoRA training with the technique: the model must be trained to respond to interpolated embeddings. PixelSmile explicitly trains with varied alpha values — the model learns that intermediate embeddings mean intermediate expressions.

3. Prompt structure: neutral and target prompts should be identical except for the controlled attribute. Keeps the delta clean.

Applications Beyond Expressions¶

The technique generalizes to any attribute controllable via text:

Relation to Other Techniques¶

Key advantage: zero additional parameters at inference (just different conditioning), continuous control, and composable.