Parallel Masked Transformer Decoding (64x36 Grid)
Step 18 / 18
Total Tokens
2,304
Typographic Score
0.94
Parallel Speedup
3.8x vs Diffusion
Unmasked Tokens
2,304 / 2,304
Latent Resolution
64x36
Est. Decode Time
1.40s
Runway Multi-Model Shot Sequencing
Gen-3 Connected
1. Meta Muse Image
Seed Pass
Discrete masked non-autoregressive decoding (18 steps)
2. Typography & Inpaint Layer
Locked
Zero-shot high-frequency text token preservation
3. Runway Gen-3 Alpha
Video Out
Motion synthesis conditioned on Muse seed frame
| Model | Paradigm | Steps | Typo Fidelity | Latency |
|---|---|---|---|---|
| Meta Muse | Discrete Masked | 18 | 94% | 1.4s |
| Continuous FLUX | Flow Matching | 28 | 91% | 3.9s |
| SDXL Turbo | Distilled Diff. | 4 | 68% | 0.9s |
| Legacy Diffusion | Iterative DDIM | 50 | 74% | 5.4s |