- LoRA dropout: applied to the LoRA path only (not frozen base weights),
0.05–0.1 helps regularize on small datasets (arXiv:2404.09610)
- LoRA+: separate optimizer param groups for lora_A and lora_B with
configurable LR ratio; ratio=16 enables LoRA+ (arXiv:2402.12354)
- Curriculum mode: logit_normal for first N% of steps then uniform,
directly addresses early convergence + fine-detail degradation at
boundaries (arXiv:2603.12517)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
logit_normal reaches lower loss but perceptual improvement over uniform
is dataset-dependent. Keeping uniform as default to match original MMAudio
training behavior; logit_normal remains available as an option.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Clips from shorter videos produce fewer CLIP frames (e.g. 2s → 16 frames,
8s → 64 frames). Mixed-length datasets would cause torch.stack() to fail
during batching. Normalize to seq_cfg.clip_seq_len / sync_seq_len at load,
same as latents are already normalized to latent_seq_len.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Uniform timestep sampling undertrained t>0.8 (the final denoising steps),
leaving residual noise that CFG amplifies at inference. Logit-normal sampling
concentrates training near t=0.5 while still covering the full range, improving
high-t coverage and reducing noise floor in generated audio.
Default changed from uniform to logit_normal (sigma=1.0). Previous behavior
available with timestep_mode=uniform.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces single-sample steps with batched sampling via random.choices().
Tensors are stacked to [B, T, C] before the forward pass; t is now [B].
Default grad_accum lowered to 1 since real batching gives stable gradients.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
500 warmup steps is 25% of a 2000-step run — too long. 100 steps lets
the full lr kick in much earlier without sacrificing stability.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
STFT hop-size rounding produces ±1 latent frame vs the expected seq length.
Clamp to seq_cfg.latent_seq_len after transpose so generator.forward assertion passes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Step checkpoints now save optimizer state, scheduler state, and step
number alongside the LoRA weights. Pass --resume path/to/adapter_stepXXXXX.pt
to continue training from that checkpoint. --steps always means total steps,
so resuming from 1000 with --steps 2000 trains 1000 more steps.
adapter_final.pt format is unchanged (state_dict + meta only) so
SelvaLoraLoader remains compatible.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- _resolve_named_path: replace / \ and null in name to prevent path
traversal outside cache_dir (would cause a confusing FileNotFoundError
at np.savez time instead of at path resolution).
- train_lora: load_npz was called twice per clip when prompt was in
prompts.txt; consolidate to a single call before prompt resolution.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Input is now pre-extracted .npz files (from SelvaFeatureExtractor) paired
with clean audio files (same stem). Visual features no longer re-extracted
during training.
- FeaturesUtils loaded with enable_conditions=False (VAE only) — Synchformer
and T5 are no longer loaded, saving ~3-4 GB VRAM.
- CLIP text encoder loaded separately via patch_clip so text prompt can differ
from the one used during feature extraction.
- Prompt priority: prompts.txt override > embedded in .npz > directory name.
- Removed: torchvision video loading, frame sampling/resizing, net_video_enc,
synchformer path check.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- LoRALinear now creates lora_A/lora_B with dtype matching the base
linear's weight, preventing a float32/bf16 mismatch at forward time
when the generator is loaded in bf16 or fp16.
- Remove unused `import math` from train_lora.py.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Teaches the model new/partial sound classes from custom video+audio pairs.
Only ~10 MB of adapter weights are trained vs ~4.4 GB for the full model.
selva_core/model/lora.py
LoRALinear: wraps nn.Linear with frozen base + trainable A/B matrices.
B initialised to zero → zero adapter contribution at init.
apply_lora(): walks named_modules, replaces matching nn.Linear in-place.
Default target: "attn.qkv" (all 21 SelfAttention QKV projections in
large_44k). Add "linear1" to also wrap post-attention output projections.
get_lora_state_dict() / load_lora() for ~10 MB save/load.
train_lora.py (standalone script, no ComfyUI dependency)
Data format: directory of video files + optional prompts.txt
("filename: description"). Falls back to directory name as prompt.
Pre-extracts features for all clips into RAM, then trains from those.
Training loop: encode audio→latent (need_vae_encoder=True), flow
matching MSE loss on velocity prediction, backward on LoRA params only.
Saves adapter_stepNNNNN.pt checkpoints + adapter_final.pt with metadata.
Key verified interfaces used:
encode_audio() → DiagonalGaussianDistribution; .mode().clone() required
normalize() is in-place
forward(latent, clip_f, sync_f, text_f, t) takes raw tensors
nodes/selva_lora_loader.py (SelVA LoRA Loader ComfyUI node)
Loads .pt adapter, deep-copies the generator, applies LoRA, loads weights.
strength param scales lora_B to adjust adapter contribution at inference.
Reads rank/alpha/target from embedded metadata if present.
Returns a patched SELVA_MODEL bundle for use with the existing Sampler.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Ethanfel