TheAnimeScripter

[2.10.0]

Added

• Added -rocm (AMD ROCm/HIP, Linux) counterparts for all PyTorch CUDA paths: upscale, interpolate (RIFE/DistilDRBA/GMFSS), restore, dedup, scene-cut, segment, depth, motion blur, transnetv2 autoclip, and DUT stabilize. Eager, no CUDA graphs; shares .pth weights with CUDA. New linux-rocm dependency profile.
• Added video_limbo-tensorrt and video_limbo_v2-tensorrt: locally exported multi-view engines with 4/8/16/32-frame windows, overlapping depth alignment, and exact-length tails. Source weights, resolution and precision have separate export caches; engines also track window size. Added tools/export_limbo_streaming.py for raw safetensors.
• Added CUDA streaming depth methods video_limbo and video_limbo_v2, using the existing Limbo checkpoints and fixed resolutions with overlapping windows controlled by --depth_window.
• Added CUDA depth methods video_small_v3 and video_base_v3, reusing DA3 Small/Base weights with cross-frame attention and overlapping 4/8/16/32-frame windows controlled by --depth_window.
• Added cyte 2x anime upscaling for CUDA, MPS, TensorRT, DirectML, and OpenVINO; corrected Limbo V2 downloads to use its published upstream files.
• --inpoint/--outpoint accept frame points with an f suffix (e.g. --inpoint 100f --outpoint 500f), resolved exactly without fps rounding; plain numbers stay seconds.
• --depth_method limbo_v2 (plus limbo_v2-mps, limbo_v2-tensorrt, limbo_v2-openvino): Limbo-V2 anime depth finetune with the same fixed-resolution setup and backend coverage as v1; README and PARAMETERS.MD now list Limbo v1 and v2 side by side.

Changed

• Updated PyTorch to 2.14.0 and torchvision to 0.29.0 across all dependency profiles, with Triton 3.8 on Windows and PyTorch-2.14-compatible Nelux 0.19.0. Corrected mismatched CUDA wheel pins in the Linux profiles.
• An *_nelux --encode_method combined with --bit_depth 16bit now stays on the in-process encoder instead of falling back to the plain FFmpeg twin: frames are carried as uint16 at full precision into a 10-bit pixel format (yuv444p10le, yuv420p10le for av1_nelux, p010le for nvenc_h265/h265/av1_nelux), mirroring the twin's 16-bit output. Requires nelux >= 0.17.0; an older installed nelux still downgrades loudly, and nvenc_h264_nelux stays 8-bit like its twin since NVENC H.264 has no 10-bit path.

Fixed

• slow_x265 and slow_x265_nelux now select the profile from the output pixel format, allowing 16-bit input to encode without an incompatible main profile error.
• RIFE CUDA inference now orders input and timestep writes before inference, refreshes captured encoder features at higher interpolation factors, and clears warmup features when graph capture fails. This prevents intermittent incorrect interpolated frames and stale-feature output in eager fallback.
• --dedup/--smooth_dedup output names carry the CLI --dedup_sens value (e.g. -Dedup35) instead of the mapped detector threshold (-Dedup0.965), which varied by backend.
• A fractional --upscale_factor via --json/--preset (e.g. 2.5) is rejected at startup instead of silently passing the check while leaving 2.5 stored for model dims; whole-valued floats (2.0) are coerced to int.
• --segment_method cartoon is rejected at startup with a CliValidationError naming the supported methods instead of failing late with NotImplementedError after metadata probing.
• small_v2-tensorrt and og_small_v2-tensorrt engine builds no longer fail with Dynamic input tensor input is missing dimensions; a failed build now raises instead of crashing on num_io_tensors.
• Video inputs with only 1 detected frame now warn to check the AE work area / prerender range and time remap.

TAS Adobe Edition

New

• AMD ROCm model choices for upscaling, interpolation, restoration, depth, segmentation, duplicate removal, scene detection, automatic clipping, stabilization and motion blur. An AMD ROCm se

[2.9.6] - 2026-08-25

Added

• --depth_method limbo (plus limbo-mps, limbo-tensorrt, limbo-openvino): Limbo-V1, an anime finetune of Depth Anything 3 small, dual-licensed CC BY-NC 4.0 with commercial use granted to TAS users; 239 fps at 1080p through TensorRT on a 3090.

Fixed

• TensorRT engine builds no longer fail with Could not find any implementation for node {ForeignNode[...]} on models whose whole graph fuses into one Myelin region, which adore at 1080p did on some driver/GPU combinations. Every engine now builds with a 4 GiB workspace budget instead of 1 GiB, which was fractionally under what such a build needs; engines already on disk are unaffected and inference VRAM is unchanged.

Removed

• 47 --depth_method choices were removed for licensing reasons, leaving small_v2, og_small_v2, video_small_v2, og_video_small_v2 and the v3 family (20 in total).

Full Changelog: https://github.com/NevermindNilas/TheAnimeScripter/compare/v2.9.5...v2.9.6

[2.9.5] - 2026-08-16

Behaviour change: an *_nelux --encode_method combined with --bit_depth 16bit or --depth now runs the equivalent plain encoder instead, with a warning — losing the in-process encoder's speed, gaining an output that matches what was asked for. --output_scale no longer triggers the swap (nelux 0.18.0 resizes encoder-side); only an installed nelux older than 0.18.0 still downgrades it.

Added

• --preview_port (default 5000) picks the --preview server's port.
• For clips whose slowdown was already baked in before export (a precomp render full of duplicate runs), --smooth_dedup --interpolate_factor 1 regenerates each duplicated slot as a true in-between at unchanged duration/fps; factor 1 is now pinned by tests as a supported path.
• 16 new in-process encoders: every slow_*, *_animation*, 10-bit, vp9, prores, gif and lossless* encode method now has a *_nelux twin (22 of 32, from 6). Each mirrors its FFmpeg twin's settings — verified within 0.1 dB PSNR on real frames, prores within 0.03% of the twin's size. png/jpeg (image sequences) and the QSV/AMF families stay FFmpeg-only.
• --stabilize_method dut: DUT deep unsupervised stabilization (RFDet keypoints + PWC-Net flow + learned mesh motion propagation and trajectory smoothing, MIT-licensed, vendored in src/stabilize/dut/), rendered as a per-region mesh warp at native resolution. On the 480-frame benchmark clip it cuts camera-path jitter 2.5-3.9x vs classic and keeps 87% of content in the worst frame vs 39%; CUDA only, ~19 fps e2e at 360p (24 fps analysis stage) vs classic's ~45.
• --encode_method jpeg writes a JPEG image sequence, mirroring png's folder/pattern and audio handling, as full-range 4:4:4 mjpeg at best quality; 8-bit only, so --bit_depth 16bit warns and encodes 8-bit. A single PNG input with processing keeps an explicit jpeg choice (writes .jpg) instead of being forced to png.
• tests/test_docsDrift.py fails when a CLI flag or an --encode_method choice is missing from PARAMETERS.MD, when the docs name a flag the parser does not accept, or when an example uses a value outside a flag's choices.
• tests/test_depthWriters.py fails if a depth backend encodes with OpenCV instead of the shared writer.
• tests/test_runOutcome.py fails if a standalone capability stops reporting its own outcome (a structural check — it pins the wiring, not the runtime behaviour).

Changed

• FFmpeg now comes from TAS-FFMPEG — one pinned, SHA256-verified 8.1.2 build per platform — instead of BtbN's and johnvansickle's rolling builds, so which FFmpeg you get no longer depends on the day you first ran TAS. macOS still installs FFmpeg with Homebrew and stays unpinned.
• Directory batches now pick up .m4v (which PARAMETERS.MD already advertised) plus .m2ts, .mts, .mpg, .mpeg, .m2v, .flv, .wmv and .ogv, and name any media file they still skip instead of dropping it silently. A source in a container TAS cannot write (.wmv, .flv, ...) now produces an .mp4 rather than an output the muxer rejects.
• --interpolate_factor below 1 is rejected at startup instead of producing a video 1/factor times longer than its audio; 0 and negatives are rejected instead of reaching the encoder as -r 0.0. Factor 1 is unchanged and still supported.
• --smooth_dedup_max_span 0 now disables the cap, interpolating every duplicate run however long; previously values below 1 were clamped to 1. Negatives normalize to 0.
• *_nelux encoders honor --output_scale natively via nelux 0.18.0's encoder-side resize instead of downgrading to the FFmpeg twin; nelux pin bumped 0.16.0 → 0.18.0.
• *_nelux encoders now convert and tag colour like their FFmpeg twins: bt709 by default (byte-identical matrix; previously swscale's default bt470bg/601) and bt2020nc with the source's own transfer for a probed BT.2020 source.
• A .webm output no longer swaps av1_nelux/

[2.9.0] - 2026-08-03

Highlights

Animation on twos and threes. --smooth_dedup now works — real motion across held drawings, same length, audio kept.

RIFE on NVIDIA was degraded. Every rife4.15-and-up model reused stale data. Fixed; output matches the reference again.

RIFE on DirectML is ~2x faster. Part of the model ran on the CPU. AMD and Intel GPUs now run all of it.

--dynamic_scale was worse than off. It picked its quality backwards. Corrected, and 10% faster.

Faster all round. Dedup +33%, decode +12–21%, autoclip −24%, stabilize +33%, detection +10%. TensorRT rebuilds 24s to 3s.

Broken output files. Wrong trim range, empty 16-bit files, 8-bit output from 10-bit NVENC, repeated depth frames.

Runs that lied or hung. Failures reported success, some hung silently, and the non-NVIDIA fallback never fired.

Also fixed. large_v3 depth 404'd on install, AnimeSR broke on odd resolutions, OpenVINO restore ran on CUDA.

TAS-AdobeEdition. Drive navigation in the Media Browser, ordered restore chains, and install failures that say what went wrong.

TAS-Standalone. New Windows desktop app, shipping with its own backend. Every setting, live progress, presets, themes.

Updated. TensorRT 11.2, PyTorch 2.13, Python 3.14.6. Engines rebuild once.

Behaviour change: --smooth_dedup now disables --dedup and enables --interpolate. Presets with both give a full-length interpolation.

Removed: --offline, and six depth choices that were never implemented.

Added

• --smooth_dedup is now real dedup-aware interpolation, with --smooth_dedup_method, --smooth_dedup_sens and --smooth_dedup_max_span. Held drawings get interpolated motion at unchanged duration, with audio.
• ty static type checking in CI (pinned 0.0.63). It fails only on first-party names that do not exist.

Changed

• An invalid --<x>_method now prints every valid choice, grouped by backend, instead of the first 10.
• Bundled Python 3.14.5 to 3.14.6, and the Linux/macOS standalone build tag to 20260728.
• TensorRT 11.1.0.106 to 11.2.1.2 on Windows and Linux, fixing 11.1's memory regressions. Engines rebuild once on first run.
• TensorRT rebuilds reuse a timing cache keyed to your GPU and versions: rife4.25 takes 2.8s instead of 24.0s.
• --smooth_dedup now disables --dedup and enables --interpolate, instead of meaning "dedup, but keep the audio".
• Interpolation drivers refresh the RIFE encoder once per pair, so --interpolate_factor 3 and up no longer reuse the previous pair's features.
• --dynamic_scale keeps CUDA graphs instead of disabling them: +10.5% at 1080p rife4.25, identical output. Factor 2 only.
• PARAMETERS.MD model tables match parser.py again - MPS everywhere it exists, a backend matrix, and the models that were missing.
• PyTorch 2.13.0 and torchvision 0.28.0 across every extra-requirements-*.txt.
• --stabilize caches ORB features instead of recomputing them: +33% end-to-end, bit-identical.
• --obj_detect draws onto the RGB frame directly: +10% at 4K, bit-identical.
• Startup reads metadata with nelux.probe() instead of ffprobe, saving 20-250 ms. Needs nelux>=0.15.1.
• --preview samples the writer's own frames, so the encode command is identical with preview on or off, at no throughput cost.
• --preview works with the *_nelux encoders.
• importRifeArch lived as six copies; it is defined once now, so adding an arch is one edit.
• README.md and PARAMETERS.MD listed Rife Elexor as DirectML- and OpenVINO-capable when the CLI had no such choice. True as of this release.
• test_frameWindow's DistilDRBA guard walks the AST instead of counting occurrences, so a new backend cannot fail it for the wrong reason.
• Second dead-code sweep, no behavior change (-877 lines): shared padding and timestep helpers, plus a dozen unreachable functions and registry arms.
• Dead-code cleanu