Skip to content

GUI v.1.6b

Latest
Latest

Choose a tag to compare

View all tags
@naxci1 naxci1 released this 24 May 10:46
· 4 commits to main since this release
GUI_v.1.6b
b2db936
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

This documentation must not leave out any parameters, UI toggles, or backend mechanisms that exist within the codebase (derived from layouts, processing forms, and execution contexts). Write it in highly organized, clear Markdown layout.

Include every single feature from this structural blueprint:

  1. ARCHITECTURE OVERVIEW & CORE DESIGN
  • Detailed explanation of how this PyQt6 Native UI bridges the complex ByteDance SeedVR2 3B DiT-VAE pipeline smoothly into an intuitive standalone workspace.
  • The 30/70 Splitter System: Left panel (30% width) for absolute parameter configuration / Right panel (70% width) dedicated to input/output media playback, split views, and full-screen comparison.
  1. MASTER COMPONENT & FEATURE CATALOGUE

A. Topaz Video AI Style Timeline & Trimming Engine:

  • Visual Timeline Control & Filmstrip: Dynamic sync between main video frame view and the lower frame preview tape.
  • In/Out Trimming Brackets: Standard buttons [ (In-point) and ] (Out-point) placed directly next to the Media Control Row (Play, Pause, Prev, Next Frame buttons).
  • Timeline Range Highlighting: Once markers are pinned, the chosen kadr boundary tracks are painted in a high-visibility vibrant blue color.
  • Clear Trimming Selection (X button): Erases visual markers, sets the blue span layout back to baseline track dark-grey, and zeroes out the frame constraints.
  • CLI Bridge: Auto-maps In-Point directly to --skip_first_frames and calculates exact kadr duration interval length for --load_cap.

B. Advanced Precision Quality Control:

  • Input Noise Scale (--input_noise_scale): Double spinbox/slider, bounds 0.0 - 1.0, precision step 0.05. Directly targets and dampens structural digital artifacts when processing dense multi-pass 4K outputs.
  • Latent Noise Scale (--latent_noise_scale): Double spinbox/slider, bounds 0.0 - 1.0, precision step 0.05. Smooths extreme high-frequency neural sharpness down in latent spaces to anchor organic cinematic texturing.

C. Full Processing Settings & Form Fields:

  • Pre-Downscale Factor: Combo configuration (1:1, etc.) to stabilize memory overhead before scaling.
  • Resolution Profiles: Dynamic Mode selection combined with Standard aspect sizing keys (700, 720 (HD), 1080 (FHD), up to Ultra High Resolution ceilings).
  • Max Resolution Constraint: Spinbox boundary to hard-cap dynamic scaling limits.
  • Batch Size Controls: Numeric stepper panel accompanied by "+" and "-" layout elements, linked with the "Uniform Batch Size" boolean toggle switch.
  • Temporal Overlap & Prepend Frames: Precise spinbox adjustments to manage continuous frame-stitching cross-fades and backward frame injections.

D. VRAM Protection & Memory Recycler Logic:

  • Only Frames Chunking (--only_frames): Sets explicit chunk sizes inside the VAE decoding loop to process frames in strict steps.
  • Empty Cache Automation: Executes aggressive torch.cuda.empty_cache() garbage collection directly at every single loop slice block completion to eliminate CUDA Out Of Memory (OOM) errors.
  • Additional Memory Toggles: "Cache DiT", "Cache VAE", "VAE Encode Tiled", "VAE Decode Tiled", and "Tensor Offload Device" selections mapped across target hardware (CPU/CUDA offloading).

E. Export Media, Profiles & Audio Configuration:

  • Video Container Choice: Formatted drop-down selections matching .mp4, .mov, etc.
  • Video Codec Matrices: Standard and high-fidelity structures including H.264, H.265 (HEVC Main 8-bit / 10-bit), and Apple ProRes configurations.
  • Custom Video Arguments Layout: Directly appends system profiles like ["-c:v", "libx265", "-pix_fmt", "yuv420p", "-preset", "veryslow", "-crf", "12", "-b:v", "40M"] to FFmpeg pipes.
  • Audio Mode Multiplexer: Choose between "Copy Audio" streams or "No Audio" parameters (-an).

F. Windows Native Telemetry & Naming:

  • Intelligent Prefix Deduplication: Outputs drop as seedvr2_[input_stem].[ext]. Uses an internal OS system loop to automatically append iterative sequence integers (_1, _2) if the output directory already contains a filename conflict.
  • WinSound Notifications: Plays an Asterisk success beep for clean completion, or a Hand error beep for processing failures. Includes a toggleable Sound Notifications checkbox under the Debug group.
  1. HARDWARE SPECIFICATION TARGETS
  • Rigorous configuration documentation for: Intel Core i7-14700KF, 96 GB DDR5 RAM, NVIDIA GeForce RTX 5070 Ti (16GB VRAM, Blackwell architecture).
  • Software framework mentions: CUDA 12.8/13.0, cuDNN 9+, PyTorch 2.7+, FlashAttention-2, SageAttention v3, Triton engines.
  1. COHESIVE WORKFLOW PRODUCTION GUIDE (HB2k's ADVANCED MANUAL)
  • Include clear tactical field manuals regarding mapping frame batch sizes close to video chunk bounds to achieve superior temporal coherence, adjusting "Only Frames" limits down to 41/81 for absolute memory safety, and applying subtle 0.10 input noise to lock structural fidelity when working with historical archives.

Generate the entire, finalized README.md output block without omission. Do not summarize sections. Write everything explicitly out.

Assets 3
Loading