ComfyUI-SelVA

Ethanfel/ComfyUI-SelVA

Fork 0

Commit Graph

Author	SHA1	Message	Date
Ethanfel	211494a91c	fix: DITTO gradient never reached x0, remove unused imports and dead code DITTO critical bug: x was reassigned on every ODE step, so by the time loss.backward() ran, x pointed to the final output tensor (grad_fn, not a leaf) and x.grad was always None. The manual gradient transfer never fired — x0 was never updated. The optimization was a no-op. Fix: use a straight-through estimator after the no-grad prefix: x = x + (x0 - x0.detach()) This adds zero value but creates a grad_fn back to x0, so backward() propagates ∂loss/∂x (at the Phase-1/2 boundary) directly to x0.grad. Equivalent to truncated BPTT with ∂x_prefix/∂x0 ≈ I. Also remove unused imports (SelvaSampler, _inject_tokens, random) that caused cascade ImportError risk, and remove dead trainable_count variable in BigVGAN trainer. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-09 12:10:02 +02:00
Ethanfel	1e9551152e	feat: add DITTO optimizer, upgrade BigVGAN trainer, document all nodes BigVGAN trainer (selva_bigvgan_trainer.py): - Add snake_alpha_only train mode: tunes only ~27K per-channel α params (0.024% of 112M) — physically cannot cause harmonic smearing - Add lambda_l2sp: L2-SP anchor regularization toward pretrained weights - Add optional discriminator_path: frozen MPD+MRD feature matching loss replaces mel L1 when a BigVGAN discriminator checkpoint is provided - Inline MPD + MRD discriminator implementations (no extra dependencies) DITTO optimizer (selva_ditto_optimizer.py): - New node: inference-time noise optimization (arXiv:2401.12179) - Optimizes x₀ via mel Gram matrix style loss against BJ reference clips - All model weights frozen — zero quality degradation risk - Truncated BPTT through last n_grad_steps of the ODE (configurable) - Gradient checkpointing on each differentiated step Docs: - README: document all 20 nodes (was 3), add workflow diagrams - STYLE_TRANSFER.md: new guide — DITTO, vocoder fine-tuning tiers, why LoRA/TI fail, combined approach, dataset prep Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-09 12:04:05 +02:00

Author

SHA1

Message

Date

Ethanfel

211494a91c

fix: DITTO gradient never reached x0, remove unused imports and dead code

DITTO critical bug: x was reassigned on every ODE step, so by the time
loss.backward() ran, x pointed to the final output tensor (grad_fn, not
a leaf) and x.grad was always None. The manual gradient transfer never
fired — x0 was never updated. The optimization was a no-op.

Fix: use a straight-through estimator after the no-grad prefix:
  x = x + (x0 - x0.detach())
This adds zero value but creates a grad_fn back to x0, so backward()
propagates ∂loss/∂x (at the Phase-1/2 boundary) directly to x0.grad.
Equivalent to truncated BPTT with ∂x_prefix/∂x0 ≈ I.

Also remove unused imports (SelvaSampler, _inject_tokens, random) that
caused cascade ImportError risk, and remove dead trainable_count variable
in BigVGAN trainer.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2026-04-09 12:10:02 +02:00

Ethanfel

1e9551152e

feat: add DITTO optimizer, upgrade BigVGAN trainer, document all nodes

BigVGAN trainer (selva_bigvgan_trainer.py):
- Add snake_alpha_only train mode: tunes only ~27K per-channel α params
  (0.024% of 112M) — physically cannot cause harmonic smearing
- Add lambda_l2sp: L2-SP anchor regularization toward pretrained weights
- Add optional discriminator_path: frozen MPD+MRD feature matching loss
  replaces mel L1 when a BigVGAN discriminator checkpoint is provided
- Inline MPD + MRD discriminator implementations (no extra dependencies)

DITTO optimizer (selva_ditto_optimizer.py):
- New node: inference-time noise optimization (arXiv:2401.12179)
- Optimizes x₀ via mel Gram matrix style loss against BJ reference clips
- All model weights frozen — zero quality degradation risk
- Truncated BPTT through last n_grad_steps of the ODE (configurable)
- Gradient checkpointing on each differentiated step

Docs:
- README: document all 20 nodes (was 3), add workflow diagrams
- STYLE_TRANSFER.md: new guide — DITTO, vocoder fine-tuning tiers,
  why LoRA/TI fail, combined approach, dataset prep

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2026-04-09 12:04:05 +02:00

2 Commits