Ethanfel

0 Followers · 0 Following

• Joined on 2026-01-04

Repositories

30

Projects Packages Public Activity Starred Repositories

Ethanfel created branch feature/lora-timestep-sampling in Ethanfel/ComfyUI-SelVA 2026-04-06 00:35:44 +02:00

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-06 00:15:34 +02:00

8ae0ba3c7d fix: increment adapter_final filename on resume to avoid overwriting previous final

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-06 00:11:19 +02:00

2b2b438307 fix: set OUTPUT_NODE=True on SelVA Feature Extractor so it runs without connected outputs

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-06 00:05:18 +02:00

39984f73c2 docs: add observed batching results to training guide

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-06 00:00:39 +02:00

1f8cd6f930 docs: rewrite LORA_TRAINING.md with real-world findings

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 23:45:45 +02:00

20f8138146 chore: show batch_size in training step log

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 23:36:14 +02:00

09b3b94ddd feat: add batch_size parameter to training (default 4)

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 23:15:50 +02:00

3f67de694c feat: save loss_raw.png and loss_smoothed.png to output_dir

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 23:05:25 +02:00

423e174b88 debug: print lora_A norm after loading to confirm adapter applied

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:51:30 +02:00

4806daa4ca chore: lower default warmup_steps from 500 to 100

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:48:58 +02:00

16b3eb11cc fix: pass max_size=800 to progress bar preview (was 85px wide)

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:44:05 +02:00

004ea63f62 fix: fall back to soundfile for torchaudio.save when torchcodec unavailable

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:40:52 +02:00

afb3242eca fix: disable inference_mode entirely for training via inference_mode(False)

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:36:31 +02:00

849f31e2a6 fix: create LoRA params inside torch.enable_grad() to escape inference_mode

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:32:01 +02:00

505d445eb3 fix: wrap training loop in torch.enable_grad()

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:17:30 +02:00

8fade1b0e3 fix: initialize LoRA params on same device as wrapped linear

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:12:23 +02:00

ad57432803 fix: pad/trim latent to exact latent_seq_len after VAE encoding

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:08:01 +02:00

43f732f904 fix: transpose VAE latent from [B,C,T] to [B,T,C] before generator

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 22:03:58 +02:00

6b9adf0816 fix: fall back to soundfile when torchcodec FFmpeg libs are missing

Ethanfel pushed to feature/lora-training at Ethanfel/ComfyUI-SelVA 2026-04-05 21:57:27 +02:00

52434a053a fix: keep VAE in float32 for mel/stft; print full traceback on clip load failure