Ethanfel/ComfyUI-SelVA
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
126 Commits 6 Branches 0 Tags
8338560600e7bcad89b9e833412a9a7a5d016db8
Commit Graph

8 Commits

Author SHA1 Message Date
Ethanfel 39984f73c2 docs: add observed batching results to training guide 
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-04-06 00:05:16 +02:00
Ethanfel 1f8cd6f930 docs: rewrite LORA_TRAINING.md with real-world findings 
- Added batch_size VRAM table and updated step recommendations for batched training
- Added adapter strength section with practical guidance (0.6-0.7 for noise)
- Added ComfyUI node as Option A for training (not just CLI)
- Noted .mp3 as not recommended, soundfile fallback implied
- Added output files section with sample_*.wav and loss curve PNGs
- Added "LoRA has no effect" troubleshooting (wrong node wired)
- Updated loss convergence targets based on observed training runs
- Clarified linear1 target: 150+ clips recommended

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-04-06 00:00:36 +02:00
Ethanfel 09b3b94ddd feat: add batch_size parameter to training (default 4) 
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>
 2026-04-05 23:36:12 +02:00
Ethanfel 4806daa4ca chore: lower default warmup_steps from 500 to 100 
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>
 2026-04-05 22:51:27 +02:00
Ethanfel 2f4641247a feat: add resume support to train_lora.py 
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>
 2026-04-05 16:59:30 +02:00
Ethanfel 8e9114b92c docs: add clip length and scalable dataset size recommendations 
- Clip length section: fixed 8s duration, padding/trim behavior, per-sound-type
  strategies (continuous, short events, repeating, onset placement).
- Dataset size table: 5-10 / 15-30 / 30-60 / 60-150 / 150-300 / 300+ clips
  with scenario and expected result for each tier.
- Note on diversity vs quantity.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-04-05 16:34:50 +02:00
Ethanfel 63b4391573 fix: named .npz files always start at _001 
dog_bark_001.npz, dog_bark_002.npz instead of dog_bark.npz, dog_bark_001.npz.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-04-05 15:44:26 +02:00
Ethanfel 89af5a468c docs: add LoRA training guide 
Covers dataset preparation (ComfyUI feature extraction + clean audio),
training CLI reference, tuning guide (rank/steps/lr), adapter loading
in ComfyUI, and troubleshooting.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
 2026-04-05 15:43:09 +02:00