v.1.9.37b

[naxci1]

1-Click SeedVR2.5 GUI — Changelog v1.8b → v1.9.37b

v1.9.37b

• Split video frame offset sync — The split-view widget now calculates the
  frame count difference between source and upscaled videos and applies an
  automatic offset. Source video frame N is displayed alongside upscaled video
  frame N+offset, ensuring temporal alignment in the comparison view.

v1.9.36b

• Postprocess prepend_frames fix — _process_frames_core was passing
  prepend_frames=0 to postprocess_all_batches, causing 4 extra frames in
  the output. Now passes args.prepend_frames to correctly remove prepended
  frames, ensuring output frame count matches input.

Export System Overhaul

v1.9.35b

• Reverted prepend removal — Temporarily reverted to diagnose frame sync
  issues. Prepend frames are kept in the output for debugging.

Comprehensive Codec Support

• H264 / H265 (NVENC GPU encoding) — Hardware-accelerated encoding via
  h264_nvenc / hevc_nvenc with automatic NVENC detection. Profiles: Baseline,
  Main, High (H264); Main, Main10, Main12 (H265). Quality levels: Low / Mid / High
  mapped to CQ values.
• ProRes — Apple ProRes with 6 profiles: Proxy, LT, Standard, HQ, 4444,
  4444 XQ. MOV container.
• DNxHR — Avid DNxHR with 5 profiles: LB, SQ, HQ, HQX, 444. MOV/MXF container.
• FFV1 — Lossless FFV1 codec (MKV/AVI).
• FFV1 Intra RGBA-16bit — Intra-only FFV1 with rgba64le pixel format for
  maximum fidelity (MKV/AVI). Uses -coder 1 -context 1 -g 1.
• AV1 (NVENC) — Hardware-accelerated AV1 encoding (av1_nvenc).
• VP9 — Software VP9 encoding (WebM/MKV).
• QuickTime V210 — Uncompressed YUV 10-bit 4:2:2 (MOV).
• QuickTime R210 — Uncompressed RGB 10-bit (MOV).
• QuickTime Animation — Run-length compressed RGB 8-bit with alpha (MOV).

v1.9.34b

• Initial prepend_frames investigation — Attempted fix for frame count
  mismatch between input and output videos.

Image Sequence Export

• TIFF — 8-bit and 16-bit (deflate compression).
• PNG — 8-bit and 16-bit.
• JPEG — 8-bit.
• EXR — 16-bit (half) and 32-bit (float).
• DPX — 8-bit, 10-bit, and 16-bit.

v1.9.33b

• Pipeline mode attempt (reverted) — Implemented per-batch streaming
  pipeline (_pipeline_process) with async frame prefetch. This was slower
  than the original 4-phase serial approach and was fully reverted. The
  original serial pipeline is retained.
• Batch size override removal — Removed chunk_args.batch_size override
  that was forcing single-batch processing per chunk, causing OOM with large
  frame counts.
• Stream mode cleanup — Removed all pipeline/stream mode artifacts:
  _pipeline_process, lightweight_cleanup, pipeline_mode, async prefetch
  thread, and related code.

Export Dialog Features

• NVENC auto-detection — Automatically probes ffmpeg for NVENC encoder
  availability and selects GPU encoding when present.
• Quality presets — Low / Mid / High quality levels per codec, mapped to
  appropriate CQ/QP values.
• Spatial AQ — -spatial_aq 1 enabled for NVENC encoders for better quality.
• 10-bit output — Optional 10-bit encoding (-pix_fmt p010le for H265,
  yuv422p10le for ProRes, etc.).
• Smart container filtering — Only valid containers per codec are shown
  (e.g., ProRes → MOV only).
• FFmpeg argument builder — build_ffmpeg_command() constructs optimal
  ffmpeg arguments based on selected codec, profile, quality, and container.

v1.9.32b

• Channels-last memory format (ComfyUI --force-channels-last equivalent)
  — The DiT model is now converted to torch.channels_last memory layout
  after materialization. On Blackwell (RTX 50xx) this enables the faster NHWC
  tensor core path for conv and linear ops. Backend-specific: only applied on
  NVIDIA CUDA and AMD ROCm RDNA3+ via supports_channels_last().
• RoPE bf16 (no float32 upcast) (ComfyUI --dont-upcast-attention
  equivalent) — Rotary position embeddings were upcasting q/k to float32 on
  every attention layer, then casting back. Now RoPE runs in native bf16
  throughout (via supports_low_precision_rope()), eliminating the expensive
  fp32 round-trip on every forward pass in both 3B and 7B models.
• VAE attention upcast disabled — VAE attention used upcast_softmax=True
  and force_upcast=True, forcing softmax to float32. Both disabled to keep
  the entire VAE pipeline in bf16. Applied at runtime in model_loader.py
  based on backend detection.
• VAE tiled encode/decode off by default — Tiling was splitting 960x720
  into 2 tiles, doubling the decode work. Now both vae_encode_tiled and
  vae_decode_tiled default to False (matching ComfyUI node defaults). RTX
  5070 Ti 16GB VRAM handles full-resolution VAE without tiling.
• CUDA caching allocator (no PYTORCH_ALLOC_CONF) — Previous attempts with
  expandable_segments:True (unsupported on Windows, PyTorch #123548 conflict)
  and backend:cudaMallocAsync (reserved VRAM never reclaimed) both caused
  fragmentation. The default PyTorch caching allocator works best on Windows.
• _parse_offload_device no longer forces CPU — When offload_device="none",
  the CLI previously auto-promoted to "cpu" if caching was enabled, silently
  re-introducing model ping-pong. Now "none" always means GPU resident.
• Performance parity fix — model ping-pong eliminated — GUI offload
  defaults were vae_offload_device=cpu and tensor_offload_device=cpu,
  forcing every model and all latent tensors to ping-pong GPU↔CPU between
  phases. Defaults changed to GPU (none), matching ComfyUI's --gpu-only.
• GUI offload labels — Dropdown options changed from cryptic "none"/"cpu"
  to friendly "GPU"/"CPU" labels with bidirectional mapping.

---

v1.9.31b

• Lossless preview source frame — Preview now reads the source frame
  directly from the original file via cv2 (preserving full bit depth) instead
  of the 8-bit display pixmap. 16-bit sources keep their full 16-bit precision
  in the comparison.
• Stale preview cleanup — Old preview artifacts (preview_source.tiff,
  preview_fallback.tiff, preview_result.tiff) are deleted before each new
  preview run and when a file is removed from the project list.
• Synchronized video split view — SplitViewWidget rewritten with
  dual-video synchronized playback via dual cv2.VideoCapture + QTimer.
  Supports play/pause, seek slider, keyboard controls (Space/Left/Right/
  Home/End), and click-to-toggle playback.

Settings Panel Enhancements

v1.9.30b

• Performance parity with ComfyUI — CLI now sets the same PyTorch
  environment variables ComfyUI uses (TORCH_CUDNN_BENCHMARK,
  TORCH_CUDNN_SDP_ENABLED, TRITON_INTERPRET=0, CUDA_CACHE_PATH,
  CUDA_MODULE_LOADING=LAZY), plus cudnn.benchmark=True and TF32 matmul
  (set_float32_matmul_precision("high")). Compile-free speedup.
• EWA-Lanczos downscale — Pre-downscale now uses GPU torch.interpolate
  (Lanczos + antialias) instead of CPU cv2 with uint8 quantization.
• FFV1 Intra RGBA-16bit codec — Added lossless intra-only FFV1 codec
  (ffv1 -pix_fmt rgba64le -coder 1 -context 1 -g 1, MKV/AVI).
• Clear list button — Project panel now has a 🗑 Clear List button
  (variant="danger") that clears all files and emits file_removed per file.
• Sound fixes — Success = rising 3-note Beep chime (659→880→1319 Hz);
  failure = descending two-tone (622→415 Hz). All via winsound.Beep.

Resolution

• Resolution modes — Pixel (exact short-side), X-Times (multiplier), and
  Presets (720p, 1080p, 1440p, 4K).
• Max resolution — Hard cap on the longest edge (default: 3840px).
• Pre-downscale — Optional 2:1 or 3:1 pre-downscale before upscaling, with
  EWA-Lanczos GPU interpolation.

v1.9.21b–v1.9.29b

• Auto-Tune OOM-retry system — When CUDA OOM occurs, the CLI automatically
  reduces tile size (1024→512→256), increases BlockSwap (0→16→24→32→36),
  reduces batch size, or enables pre-downscale — then retries without manual
  intervention.
• BlockSwap escalation — Automatic block swap escalation for low-VRAM GPUs.
• Conv3d memory bug workaround — Detects PyTorch 2.9+ with cuDNN ≥91002
  Conv3d 3x memory bug and applies direct torch.cudnn_convolution bypass.
• Device manager backend detection — Unified Backend enum (CUDA/ROCM/
  XPU/MPS/CPU) with supports_channels_last() and supports_low_precision_rope()
  for backend-specific optimizations.
• VAE decode batch cleanup — empty_cache() between decode batches to
  prevent reserved-but-unallocated VRAM accumulation.
• retry_on_oom sleep removal — Removed hardcoded 0.5s sleep in OOM retry
  path for faster recovery.
• release_model_memory VRAM flush — Added empty_cache() after model
  parameter release to immediately return freed VRAM to the allocator.

Batch Processing

• Batch size — Configurable batch size (default: 81). Higher = better temporal
  coherence but more VRAM.
• Uniform batch size — Pads the final batch to match batch_size for consistent
  processing.
• Temporal overlap — Overlapping frames between batches for smooth blending
  (default: 4).
• Prepend frames — Reversed frames prepended at the start to reduce start
  artifacts (default: 4).

VAE Tiling

• Encode tiled — Optional spatial tiling during VAE encoding.
• Decode tiled — Optional spatial tiling during VAE decoding.
• Tile size / overlap — Configurable tile dimensions and overlap for both
  encode and decode.

Color Correction

• Methods — LAB perceptual transfer, Wavelet adaptive, Wavelet reconstruction,
  HSV, AdaIN, or None.

Noise

• Input noise scale — Adds subtle noise to input before encoding (0.0–1.0).
• Latent noise scale — Adds noise during diffusion conditioning (0.0–1.0).

Advanced

• Auto-Tune — Automatically reduces tile size, enables BlockSwap, and adjusts
  batch size on OOM errors without manual intervention.
• Offload devices — GPU/CPU selection for DiT, VAE, and intermediate tensors.
  Friendly dropdown labels ("GPU"/"CPU") with automatic CLI token mapping.
• BlockSwap — Configurable transformer block swapping (0–36) for low-VRAM GPUs.
• Swap I/O components — Additional I/O layer offloading for maximum VRAM savings.
• Chunked processing — chunk_duration_minutes for streaming large videos in
  time-bounded chunks with temporal overlap blending.
• Frame selection — load_cap, skip_first_frames, only_frames for
  processing specific portions of a video.
• Attention mode — SDPA, Flash Attention 2/3, SageAttention 2/3 selectable.
• Model caching — cache_dit / cache_vae keeps models in RAM between runs,
  eliminating SSD re-reads for batch processing.
• Debug mode — Detailed logging with VRAM/RAM statistics per phase.
• Tile debug — Visualizes encode/decode tile boundaries on output frames.

---

Project Panel

• Clear list button — 🗑 "Clear list" button (danger variant) removes all files
  from the queue. Emits file_removed signal per file for proper cleanup.
• Drag & drop — Drag video files directly into the project list.
• Folder batch — Select a folder to batch-process all supported video files.
• Per-file remove — Each file has a ✕ button for individual removal.

---

Split View Widget

• Dual-video synchronized playback — Opens two cv2.VideoCapture handles and
  plays source + upscaled videos in sync via QTimer.
• Automatic frame offset — Calculates frame count difference between source
  and upscaled videos, applies temporal offset for correct alignment.
• Seek slider — Embedded QSlider for frame-accurate scrubbing.
• Keyboard controls — Space (play/pause), Left/Right (±1 frame), Home/End
  (jump to start/end).
• Click to toggle — Click anywhere on the video to toggle play/pause.
• Divider drag — Draggable split divider for adjustable left/right ratio.
• Zoom & pan — Mouse wheel to zoom, drag to pan.
• Static image mode — Falls back to set_images() for single-frame preview
  comparison.

---

Preview System

• Lossless source frame — Preview reads the source frame directly from the
  original file via cv2 at native bit depth (16-bit TIFF preserved) instead of
  the 8-bit display pixmap.
• Stale preview cleanup — Old preview artifacts are deleted before each new
  preview run and on file removal, preventing stale split-view images.

---

Performance Optimizations

Environment & Allocator

• ComfyUI-matching environment variables — TORCH_CUDNN_BENCHMARK=1,
  TORCH_CUDNN_SDP_ENABLED=1, TRITON_INTERPRET=0, CUDA_MODULE_LOADING=LAZY,
  CUDA_CACHE_MAXSIZE=4GB, persistent CUDA_CACHE_PATH.
• cuDNN benchmark + TF32 — torch.backends.cudnn.benchmark = True and
  set_float32_matmul_precision("high") for compile-free speedup.
• Default CUDA caching allocator — No PYTORCH_CUDA_ALLOC_CONF override.
  expandable_segments is unsupported on Windows; cudaMallocAsync causes VRAM
  fragmentation; max_split_size_mb conflicts with expandable_segments (#123548).
  The default allocator is the most stable on Windows.

Model Memory

• Channels-last (NVIDIA/AMD) — DiT model converted to torch.channels_last
  after materialization, enabling NHWC tensor core path on Blackwell/RDNA3+.
  Backend-gated via supports_channels_last().
• RoPE bf16 (no fp32 upcast) — Rotary position embeddings run in native bf16
  on CUDA/ROCm/XPU, eliminating the per-attention-layer fp32 round-trip in both
  3B and 7B models. Backend-gated via supports_low_precision_rope().
• VAE attention upcast disabled — upcast_softmax=False and
  force_upcast=False to keep entire VAE in bf16. Applied at runtime based on
  backend detection.
• Model ping-pong eliminated — When offload_device=none + caching, models
  stay resident on GPU (no CPU round-trip). When offload_device=cpu + caching,
  models are kept in RAM (no SSD re-read).
• _parse_offload_device fix — "none" always returns None (GPU resident),
  no longer auto-promoted to "cpu" when caching is enabled.

VAE Decode

• empty_cache() between decode batches — Prevents reserved-but-unallocated
  VRAM accumulation that triggers OOM on later batches.
• release_model_memory VRAM flush — empty_cache() after parameter release
  to immediately return freed VRAM to the allocator.
• retry_on_oom sleep removal — Removed hardcoded 0.5s sleep for faster OOM
  recovery.
• Conv3d cuDNN benchmark — benchmark=True in the Conv3d memory bug workaround
  for optimal kernel selection.

Backend Detection

• Unified Backend enum — CUDA (NVIDIA), ROCM (AMD), XPU (Intel), MPS (Apple),
  CPU. Automatic detection via torch.version.hip, torch.xpu, etc.
• supports_channels_last() — Returns True for NVIDIA CUDA and AMD RDNA3+/MI300.
• supports_low_precision_rope() — Returns True for CUDA, ROCM, XPU, CPU.
  Returns False for MPS (numerical stability).

Auto-Tune OOM System

• Automatic step-down — On CUDA OOM: tile size 1024→512→256, BlockSwap
  0→16→24→32→36, batch size reduction, pre-downscale activation.
• No manual intervention — Retries automatically with reduced settings.

---

Audio & UX

• Success chime — Rising 3-note Beep (659→880→1319 Hz) on successful export.
• Failure tone — Descending two-tone Beep (622→415 Hz) on error.
• Toast notifications — Non-intrusive popup messages with auto-dismiss.
• Button3D variants — default, danger, ghost styles for different actions.

---

ComfyUI Optimizations Applied

All CLI optimizations were also applied to the ComfyUI custom node at
C:\ComfyUI\ComfyUI\custom_nodes\seedvr2_videoupscaler\:

• RoPE bf16 (3B + 7B)
• VAE upcast_softmax=False, force_upcast=False
• device_manager.py backend detection
• DiT/VAE cache RAM (none + cache_model = RAM resident)
• Cache validation removal (none + cache_model allowed)
• sleep(0.5) removal in retry_on_oom
• Conv3d benchmark=True

---

Version History Summary

Version	Key Changes
v1.9.18b–v1.9.21b	Export system, codecs, NVENC, settings panel, Auto-Tune
v1.9.22b–v1.9.29b	Project panel, Toast, Button3D, sound, performance env vars
v1.9.30b	EWA-Lanczos, FFV1, clear list, sound fixes, performance parity
v1.9.31b	Lossless preview, stale cleanup, synchronized video split
v1.9.32b	Channels-last, RoPE bf16, VAE upcast, offload GPU defaults
v1.9.33b	Pipeline attempt (reverted), batch override removal
v1.9.34b–v1.9.36b	Prepend frames / frame sync investigation
v1.9.37b	Split view frame offset sync

---

This discussion was created from the release v.1.9.37b.