Ethanfel and Claude Sonnet 4.6
f99d2666e8
fix: interpolate sync_cond to match audio sequence length in transformer
...
Sync_MLP interpolates sync features based on video duration, but audio
latent length depends on the user-set audio duration. When video != audio
duration, the sequences diverge. Resample sync_cond to x's length before
the gated addition so any video/audio duration combo works.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com >
2026-03-27 21:21:39 +01:00
Ethanfel and Claude Opus 4.6
84c81e0e55
feat: extract prismaudio_core model modules (DiT, conditioners, VAE, diffusion)
...
Fetch and adapt inference-critical model modules from upstream PrismAudio repo:
- dit.py: DiffusionTransformer with debug prints removed
- diffusion.py: ConditionedDiffusionModelWrapper, DiTWrapper, MMDiTWrapper
- conditioners.py: Cond_MLP, Sync_MLP, MultiConditioner with stubbed training imports
- autoencoders.py: AudioAutoencoder, OobleckEncoder/Decoder
- transformer.py: ContinuousTransformer, Attention with flash_attn fallback to SDPA
- blocks.py, utils.py, bottleneck.py, pretransforms.py, local_attention.py, pqmf.py
- adp.py: UNetCFG1d, UNet1d, NumberEmbedder
- mmmodules/model/low_level.py: MLP, ChannelLastConv1d, ConvMLP
All internal imports rewritten from PrismAudio.* to prismaudio_core.*,
training-only imports stubbed, flash_attn made optional with HAS_FLASH_ATTN flag.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com >
2026-03-27 17:31:22 +01:00
EthanfelandClaude Sonnet 4.6