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>
The third element in ComfyUI's preview tuple is max_size in pixels, not
JPEG quality. Passing 85 was capping the live loss curve at 85×40px.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
torch.enable_grad() alone is insufficient: operations on inference tensors
(created inside ComfyUI's outer inference_mode context) produce inference
tensors even inside enable_grad, breaking autograd. inference_mode(False)
exits the inference context so the deepcopy, apply_lora, and training loop
run with a fully clean autograd context.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
torch.enable_grad() re-enables grad tracking but nn.Parameters created while
torch.inference_mode() is active are inference tensors that can't enter autograd
regardless. Splitting into _train_inner() and calling it inside enable_grad()
ensures the deepcopy, apply_lora, and the training loop all run with a clean
autograd context.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ComfyUI executes all nodes inside torch.no_grad(), which prevents gradient
tracking and makes loss.backward() fail. torch.enable_grad() overrides it.
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>
Recent torchaudio defaults to torchcodec as the audio backend, which requires
FFmpeg shared libraries. Falls back to soundfile for envs where torchcodec
can't load (e.g. containerised ComfyUI without system FFmpeg).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
torch.stft requires float32 input — casting vae_utils to bf16 caused silent
failures during dataset pre-loading. Also adds traceback.print_exc() so future
clip-load errors are visible in the ComfyUI log.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
At every save_every steps, run a quick 8-step no-CFG inference pass on
a random training clip and save the decoded waveform as
sample_stepXXXXX.wav next to the checkpoint. Uses the existing
generator.unnormalize + feature_utils.decode + vocode pipeline from
the sampler. Failure is non-fatal (logged and skipped).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Send updated loss curve to ComfyUI frontend every 50 steps via
pbar_train.update_absolute() with a JPEG preview tuple — same
mechanism as KSampler's denoising previews.
- Fix x-axis step labels for resumed runs (previously always started
at 0; now correctly shows start_step + offset).
- Split _draw_loss_curve (returns PIL Image) from _pil_to_tensor
(converts for ComfyUI IMAGE output).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Runs the full training loop inside ComfyUI. Reuses the already-loaded
CLIP model from the inference model for text encoding; loads only a
minimal VAE encoder separately (freed after dataset pre-loading).
Outputs:
- SELVA_MODEL with LoRA applied (ready to connect directly to Sampler)
- adapter_path STRING (for SelVA LoRA Loader in future sessions)
- loss_curve IMAGE (PIL-rendered line chart of training loss per 50 steps)
Progress is shown via ComfyUI ProgressBar (two phases: dataset loading,
then training steps). Resume is supported via resume_path input.
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>
When name is provided, features are saved as name.npz (or name_001.npz,
name_002.npz etc. if the file already exists) instead of a content hash —
useful for building a named training dataset. Hash-based caching is
unchanged when name is left empty.
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>
- Replace zero-fill with neutral gray (0.5) fill so masked background
pixels stay in-distribution: 0.5 maps to ~0 in CLIP normalized space
and exactly 0 after sync's [-1,1] normalization
- Add mask_strength float (0–1) for partial background suppression
- Add mask_clip / mask_sync booleans to toggle masking independently
on the CLIP (384px) and TextSynchformer (224px) encoding paths
- Fix temporal mask sampling: use fps-accurate index formula (same as
_sample_frames) instead of proportional int(i*M/N)
- Include mask_strength, mask_clip, mask_sync in cache hash when mask
is connected, so changing any param correctly busts the cache
- Log lines now report masked/skipped state and strength per path
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Both nodes moved models to GPU before work then back to CPU after.
Any exception (OOM, cancellation, bad input) would skip the cleanup,
leaving models on GPU permanently until ComfyUI restarts.
Wrap the entire work block in try/finally so offload_to_cpu cleanup
always runs regardless of how the node exits. Also removes the unused
`mode` variable in SelvaSampler.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- selva_sampler: wrap decode+vocode in their own OOM catch — previously
OOM during mel decode or vocoding gave a raw CUDA traceback instead
of the actionable hint
- selva_feature_extractor: sync frames log line now shows (masked) when
a mask is active, matching the CLIP log line
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Allows per-frame or static segmentation masks to be applied before CLIP
and sync encoding, zeroing background pixels. Useful when multiple objects
compete for the same sound and text prompting alone is insufficient.
- _apply_mask(): resizes mask spatially (nearest-exact), samples temporally
to match sampled frame count, multiplies into frames
- _hash_inputs(): includes mask bytes in cache key (begin/mid/end sampling)
- INPUT_TYPES: mask added to optional inputs with tooltip
- extract_features(): mask=None parameter, applied after _resize_frames for
both CLIP (384px) and sync (224px) paths, before normalization
- Log line notes when masking is active
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Model Loader:
- bf16 support check — auto-falls back to fp16 on unsupported GPUs
- DESCRIPTION and OUTPUT_TOOLTIPS
Feature Extractor:
- Store variant in features dict and .npz cache
- Progress bar (3 steps: CLIP encode, T5 encode, sync encode)
- Expand cache hash to 32 hex chars
- DESCRIPTION and OUTPUT_TOOLTIPS
Sampler:
- Variant mismatch validation against extracted features
- Cancellation support via throw_exception_if_processing_interrupted()
- OOM catch with actionable error message
- normalize toggle (optional BOOLEAN, default true) for peak normalization
- Remove empty optional: {} block
- DESCRIPTION and OUTPUT_TOOLTIPS
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- nodes/__init__.py: fix [PrismAudio] leftover label in error print
- selva_feature_extractor: hash beginning, middle and end of video tensor
instead of just first 1MB, avoiding collisions on videos with same opening frames
- selva_sampler: derive SequenceConfig from model template via dataclasses.replace
instead of hardcoding sampling_rate/spectrogram_frame_rate per mode
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This branch registers only the three SelVA nodes. PrismAudio nodes stay
on master/feature/lora-trainer.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Actual filenames in jnwnlee/SelVA: generator_*_44khz_sup_5.pth.
download_utils.py had the wrong names so those MD5s are unverified — set to
None to skip MD5 check for 44k generators. All other files verified/unchanged.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously _ensure() trusted any existing file. Files downloaded by the
broken requests-based code (HTML error pages) would be silently reused.
Now checks MD5 on every load; deletes and re-downloads on mismatch.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
download_utils.py used requests without auth — jnwnlee/SelVA returned an
HTML error page which torch then failed to unpickle ('E' / opcode 69).
huggingface_hub.hf_hub_download() handles HF_TOKEN auth automatically,
validates downloads, and retries. Files are still copied to models/selva/.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
PyTorch 2.6 changed the default to weights_only=True. SelVA checkpoints
contain non-tensor types (numpy scalars etc.) that fail strict unpickling.
All weights come from trusted sources (jnwnlee/selva HF repo).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- SelvaSampler: multiline:false puts negative_prompt inline above sliders
- SelvaModelLoader: VAE filenames in download_utils are v1-16.pth/v1-44.pth,
not v1-{mode}.pth (mode includes the 'k' suffix)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Move negative_prompt to required inputs, right after prompt, so it appears
above duration/steps/cfg/seed in the ComfyUI node layout.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Users can now wire the prompt output directly to SelvaSampler's prompt input,
making the data flow explicit instead of relying on the implicit features fallback.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ComfyUI renders required inputs above optional ones. Moving negative_prompt
to optional puts prompt first (natural order) and negative_prompt at the
bottom where it belongs as a power-user input. Also guards against
negative_prompt=None when not connected.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Uses selva_core/utils/download_utils.py (already has URLs + MD5s for all
weights). Models download to models/selva/ on first load. Synchformer reuses
models/prismaudio/synchformer_state_dict.pth if already present (no duplicate
download for PrismAudio users), otherwise downloads to models/selva/.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
SelvaFeatureExtractor now stores the prompt in SELVA_FEATURES (both in the
returned dict and the .npz cache). SelvaSampler's prompt is now optional —
when left empty it falls back to the prompt stored in features. A non-empty
override can still be passed when CLIP text should differ from the sync text.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- selva_feature_extractor: cache hash now includes resolved duration;
same video + different duration override no longer returns stale features
- selva_sampler: MPS-safe noise generation (torch.Generator on CPU then
move to device, same pattern as PrismAudioSampler)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Calls update_seq_lengths with actual feature dimensions (not seq_cfg) to
avoid rounding assertion mismatches. Progress bar tracks each Euler step.
Supports negative prompts for steering, normalizes output to [-1,1].
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
CLIP frames at 8fps→384px (normalize inside FeaturesUtils).
Sync frames at 25fps→224px, normalized to [-1,1] externally.
T5 text encoded via FeaturesUtils, sup tokens prepended, then text-conditioned
sync features extracted via TextSynch.encode_video_with_sync(). Results cached
as .npz keyed by hash(frames[:1MB] + prompt + fps + variant).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The fps output was only returned on cache hits. Fresh extractions
returned only features, leaving fps null.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Setting duration to 0 in PrismAudioSampler now reads the duration
stored in the PRISMAUDIO_FEATURES dict (set by the feature extractor).
Default changed from 10.0 to 0.0 so V2A workflows are wired up
automatically.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Remove synchformer_ckpt input — always resolved from models/prismaudio/
(errors early with clear message if missing)
- Replace python_env string input with dropdown: managed_env (isolated
auto-created venv, default) or comfyui_env (current Python, with warning)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When the VHS LoadVideo video_info output is connected, loaded_fps is
used automatically instead of the manual fps input.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Saves frames as uint8 .npy instead of H.264 MP4, eliminating the
lossy codec roundtrip. extract_features.py loads .npy directly and
skips decord when given a numpy file. Passes --source_fps for
correct temporal sampling.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Ethanfel