Changelog

2026-10-23

Version: 1.0.11

• Feat (New)

  • Add PIPE-only Image Timeline Video and IO Image Timeline to hold one image for the soundtrack's duration or switch multiple images using even spacing, nearby quiet gaps or manual times. Share Wan/MiniMax activity-gap detection and Batch Extend With Overlap's visual blends, including CPU optical-flow morphs with reusable motion estimation, shared size/FPS overrides and a visible report. Stream a lossless temporary VIDEO with the original soundtrack for captions or saving, preserving audio timing without allocating a full decoded video batch.

  • Add Load Image Timeline with reorderable file slots, multi-file drag-and-drop, centered cropping and per-image blends that override renderer defaults. Scroll the file list without canvas zoom. A shared Update previews button refreshes every crop locally or resolves optional width/height inputs through one partial run ending at the loader, with each dimension's source shown. Choose one image's sampler, prompt and seed metadata for downstream PIPE use; preserve per-image options when unchanged images pass through IO and retain the Image Selector route.

• Feat

  • Add Save Video's Audio input toggle to replace native VIDEO soundtracks with connected audio while preserving video duration, frame timing and crop/trim. Trim excess audio, pad shorter tracks with silence, and return the saved VIDEO with the same soundtrack as its preview. Existing workflows keep the override off.

  • Share ghost zoom, flash/light leak, block glitch, zoom blur and dip-to-black transitions across image timelines, Load Video and Batch Extend With Overlap. Load Video adds global defaults and per-clip overrides to its joined preview and output, preserving clip timing, audio and metadata.

  • Add a Global style section to Image Timeline Video and Load Video for whole-sequence colour grading and repeating effects, including combined Black pulse flash, Ghost edit and Glitch edit styles with strength and period controls. Stream effects without retaining full video batches; existing workflows keep styling off.

• Fix

  • Keep mouse-wheel scrolling inside Load Video's clip list in Nodes 2.0, including at its top and bottom, while retaining normal canvas zoom outside the list.

  • Convert native VIDEO previews to browser-compatible H.264 with AAC audio instead of remuxing lossless RGB streams, preventing green/magenta playback from image timelines and video upscaling. Preserve source timing and reuse already compatible Eclipse previews.

• Docs

  • Document both image timeline sources, PIPE connections, framing, per-image/clip blending, global styles, metadata selection and connections to lyric captions and saving.

Version: 1.0.10

• Feat (New)

  • Add Render Lyric Captions_v2 with PIPE-only input/output and compact IO Lyric Captions for original audio, vocals, lyrics/timing, separate manual corrections, background, dimensions, FPS, trim and seed. Keep appearance and alignment choices local to each renderer. Return the used settings and full-song timing alongside video, SRT and report; retain the existing renderer.

  • Add WAN LipSync Conditioning, using native InfiniteTalk with an explicit audio start and a bounded motion tail. Plan Step accepts IMAGE batches or disk-backed timelines; the disk path preserves the full audio clock at 25 FPS without keeping all preceding frames in RAM.

• Feat

  • Add optional Qwen3 forced alignment to Transcribe Audio and both lyric renderers, with a compare-only mode, bounded phrase timing and cached reuse. Download missing official model files on first use, verify the pinned revision and reuse local files offline. Keep Whisper large-v3 as the default and preserve manual timing priority; invalid Qwen estimates retain existing phrase timing.

  • Add Play to Preview Video's Mode bar to run its upstream branch through the preview without starting downstream work or changing Stop. Play remains available when collapsed; existing review and cancel controls are retained.

  • Share transcription and bounded reference caches between Transcribe Audio and the new pipe renderer. Empty corrections keep original timing; recognizable wording edits align only changed phrases while preserving unchanged lines, and populated manual timing takes priority. Preview and final renders can reuse the same resolved pipe.

  • Let Render Lyric Captions accept {} as automatic alignment for edited lyrics. This marker passes through empty-string fallback switches while muted manual inputs still select the original timing; populated timing JSON retains strict validation.

  • Add Play and Stop controls to Transcribe Audio. Play opens the transcript and runs only its upstream branch; Stop pauses normal queues after changed results, with cached continuation when queued again unchanged. The collapsed bar keeps Play beside its summary. Existing workflows keep Stop disabled.

  • Add optional image-change warmup to WAN LipSync Timeline Planner. Plan Step supplies trim_image for overlap plus warmup and audio_start for conditioning; its crop output applies to both IMAGE batches and disk timelines. Hidden lead-in frames use preceding guide audio without shifting visible transitions or the soundtrack. Warmup stays within the fixed render window and defaults to disabled.

• Fix

  • Filter the recurring “Satsang with Mooji” attribution artifact from automatic transcripts and lyric-recognition windows, including its word timestamps. Preserve surrounding lyrics and explicit manual text/timing; report filtered occurrences and invalidate older transcription caches.

  • Recover short unmatched lyric sections, including openings and endings, with confined recognition retries on vocals and the original audio. Align recovered words using their matching audio source while preserving thresholds and neighboring timing; invalidate older alignment cache results.

  • Preserve panel node widths when expanding or collapsing option bars and sections, including Add Filename Prefix, Smart Prompt, sampler and media panels. Heights continue adapting to controls, while resizable text and preview panels retain their existing height behavior.

• Docs

  • Explain lyric and timing correction priority, optional Qwen dependencies and model setup, transcript/preview controls, and Wan warmup and disk-timeline connections.

2026-10-21

Version: 1.0.8

• Feat (New)

  • Add PIPE-only Eclipse KSampler Advanced with standard/two-stage modes, independent step ranges and noise controls, optional between-stage latent upscale, a shared base seed and Stage 2 offset, and final-only decoding. Defaults match the configured Nova sampler in the Moody Moss workflow, including 2× bicubic upscale.

  • Add IO KSampler to control both sampling stages, seeds, noise toggles, latent upscale, preview and decode through a PIPE, using the standard sampler fields for Stage 1 and separate Stage 2 fields, with local fallback for missing values and preservation of other metadata.

• Feat

  • Support Image Selector seed locking for the new advanced sampler.

  • Add adjacent model_1 and model_2 connections to IO KSampler for high/low-noise model pairs, with latent-format checks, independent model patches and preview-decoder replacement at the handoff. Stage 2 automatically reuses model_1 when model_2 is absent from the PIPE.

  • Add optional Qwen schedule, DiT skip guidance and Detail Daemon sections with value tooltips and local-only controls independent of PIPE overrides. Recalculate Qwen sigmas per stage resolution, isolate model patches, and include the Detail Daemon curve and sampler wrapper.

  • Give both Eclipse samplers a shared live/final preview that resizes with the node and supports batch thumbnail selection, click-to-next navigation and returning to the grid in both renderers.

• Fix

  • Restore local preview visibility immediately when Use pipe settings changes, and preserve wrapped sampler inpainting options in Detail Daemon.

  • Use 0.1 increments for general float controls, retaining finer precision where fractional defaults or bounds require it, including Qwen schedule and Detail Daemon controls, sampler noise, MiniMax timing, audio alignment and resolution controls so valid values are not marked invalid.

  • Preserve structured LoRA and audio payloads in Concat Pipe Multi while retaining arbitrary PIPE fields and advanced sampler controls. Reject malformed tuple inputs and unsupported merge options. Document loader-to-sampler PIPE extraction and overwrite ordering for resolved generation data.

  • Bound sampler live-preview resolution before TAESD decoding and reuse one decoder across advanced sampling stages. Preview-only out-of-memory errors disable live frames while sampling and final decoding continue.

2026-10-20

Version: 1.0.7

• Fix

  • Refresh Save Images and shared image previews when filenames are reused, preventing stale browser-cached pictures while retaining cached navigation within each result.

  • Ignore late preview dimension responses from previous results after a newer batch arrives or the preview is cleared.

Version: 1.0.6

• Fix

  • Image Selector confirmation locks both Dual Seed values to the displayed execution, preserving full 64-bit precision and saved workflow selections.

Version: 1.0.5

• Feat (New)

  • Add Dual Seed_v2 with independent Image and Prompt controls and outputs, one shared 64-bit toggle, and reusable paired results.

• Feat

  • Keep up to 50 recent results in Seed and Dual Seed; histories persist with workflows.

2026-10-19

Version: 1.0.4

• Fix

  • Give each intentional MiniMax H3 camera cut one early timestamp inside the discarded warmup, allowing the new shot to settle before visible output while preserving frame ranges and audio timing.

  • Align MiniMax H3 technical cuts to activity gaps in multi-image timelines while respecting warmup, endpoint and render limits.

  • Detect softer activity onsets before sustained high activity, and retain surrounding gap evidence when constraining cut positions.

• Docs

  • Explain hidden camera-cut timing, activity-gap placement and rendered transition checks.

Version: 1.0.3

• Feat

  • Add Neutral Jade, Neutral Ocean, Neutral Amethyst and Neutral Amber, pairing Neutral node backgrounds and titles with each palette's colored controls. Eclipse's Theme groups option applies matching Neutral group colors.

• Docs

  • Document the new variants, linked selection across the Nova packs and group appearance.

Version: 1.0.2

• Fix

  • Preserve conditioning attention masks in IO Slice & Dice and list/batch conversion, including single-image selections and shared masks.

Version: 1.0.1

• Fix

  • Preserve category title colors for pipe, router and Get/Set nodes when applying themes or reopening workflows. Previously automatic gray titles return to their category color; custom title colors remain unchanged.

2026-10-17

Version: 1.0.0

Initial release of Eclipse Nova

Eclipse Nova 1.0.0 is the initial release of a separate repository, built from the original Eclipse pack with a redesigned frontend. This entry describes the features and improvements included in that first release.

Canvas performance was a major reason for the move from Eclipse to Nova: large workflows made the cost and complexity of managing many widgets particularly noticeable. This release focuses on reducing that overhead and making complex nodes easier to use in both classic canvas and Nodes 2.0.

The design and panel-based approach are inspired by the DaSiWa node set and adapted to Nova's tools and workflows.

• Design and usability

  • Redesigned nodes use compact panels with structured configuration, replacing the earlier approach of managing large sets of individual widgets through visibility helpers.

  • Shared controls provide collapsible mode bars, pinnable sections, searchable menus, dynamic rows, color pickers and resizable text/media areas, adapted to each tool's purpose.

  • Use Sliders adds live sliders to bounded numeric fields in the redesigned panels while retaining exact number entry and seed precision.

  • Neutral is the default node design for new installations, with gray controls and no selection glow. Saved theme choices remain unchanged.

  • Editable JSON themes provide Jade, Neutral, Ocean, Amethyst and warm Amber (matching the README tour) plus user designs in themes/custom. Appearance settings apply colors to controls and backgrounds, including inside subgraphs.

  • Automatic titles use the theme's +20 gray offset and follow the selected design through workflow loads, including missing or older automatic title colors. Custom titles and black routing bodies are preserved.

  • Theme regular nodes and Theme groups are independent Appearance options, both off by default. Group colors use RGB 60 with dark themes and RGB 80 with Neutral. Turning either option off restores previous colors, including after a workflow reload.

  • Link Nova themes optionally keeps installed SmartLLM Nova and Smart Model Loader Nova packs on Eclipse's selected theme across changes and restarts. Each pack saves its own preference; unlinking restores independent choices. Missing custom themes or failed saves are reported without substituting another theme.

  • Bundled themes are extracted once from .defaults/themes into the user-owned themes folder. Existing palettes, custom edits and deleted files are preserved across restarts and updates; a newly bundled theme is installed once.

  • Panel values and layout choices persist through workflow saves, reloads and renderer changes.

  • Text editors share the available node space between open fields. A single visible field fills the space, and dragging a field's resize handle adjusts the split while preserving the outer node size.

  • Fixed-input processing and utility tools use the shared design as their standard interface, with flat controls for small operations and conditional sections for more complex settings.

• Canvas performance

  • The new panels use localized updates driven by edits, lifecycle events and content-size changes. Size measurements are batched, and unchanged dimensions do not trigger another resize.

  • The old widget-visibility and system-performance helper layers, global preview/viewport culling and resize patches are removed. The redesigned interfaces handle their own layout through ComfyUI's DOM hosting.

  • Prompt-library catalog and file reads are shared during workflow loading and page refreshes, avoiding duplicate requests when many prompt nodes use the same files.

  • Shared appearance loads once at startup and updates on explicit changes, without per-node polling or draw-time color checks. Retired frontend scripts and duplicate interfaces are excluded from the initial release.

• Workflow experience

  • Shared seed controls preserve exact 64-bit values and support local or connected seeds where applicable. Partial execution scopes seed advancement and automatic choices to the requested workflow branch.

  • Play actions use ComfyUI's native execution up to the selected node. Review controls can pause changed results while unchanged cached inputs continue.

  • SEGS Preview Simple passes its input through the first SEGS output, with IMAGE second, so downstream SEGS processing can include the preview in its execution path.

  • Detailer uses Eclipse's final-image grid with click-to-next browsing and a shared area for live sampling previews. Preview immediately hides or restores available images without another queue, and the preview area resizes in both renderers.

  • Image Filter Adjustments Advanced responds to Interrupt/Cancel during resizing and supersampling. It computes supersampled regions on demand instead of allocating the whole enlarged intermediate, preserves the selected scale and filter, and checks available RAM before allocating large outputs.

  • Image Selector keeps the seeds used for the displayed results when continuing with Nova prompt, sampler and companion panels. Confirmation is tied to every image's contents, including changes within a batch.

  • Image uploads respect the shared mutation restrictions and reserve separate filenames for simultaneous uploads.

  • Dedicated override inputs and optional settings pipes connect compact editors to larger workflows. A pipe-priority toggle lets supplied sampler settings and seeds override local controls while retaining local values for missing settings. Searchable file browsers, generation metadata and resizable previews support inspection and reuse of results.

  • Processing panels retain existing data and list contracts, expose connected setting overrides, and make execution reports available beside their controls. Model choices refresh on demand.

  • Related tools are consolidated where they share a purpose, with reusable processing engines and UI components supporting a smaller, more consistent interface.

• Independent Nova release

  • Nova has its own node IDs, menu, settings, API routes and configuration, allowing installation alongside the original Eclipse pack.

  • Linux CFFI startup compatibility is retained.

  • Nova exposes its prompt library through smart_prompt_nova. The original Eclipse wildcard link is preserved, and Nova ignores stale links from removed or disabled checkouts.

  • The README and node guides describe the available tools and their usage, with a visual-tour index and annotated screenshots of the current panels, themes and routing dialogs.

• Fix

  These fixes address issues carried over from the original Eclipse repository and are included in Nova's initial release.

  • Keep primitive conversion, switch and pipe links scoped to their owning graph, including nested workflows and restored typed connections.

  • Evaluate all 64 lazy-switch slots in priority order, fall through resolved empty results, preserve zero and false values, and skip unused branches. Saved high-slot connections survive workflow loading while fresh nodes show only active rows.

  • Preserve separate prompts, every conditioning entry, pooled outputs and masks when converting or selecting batches. Conditioning from encoders without pooled embeddings remains valid. Compatible prompts batch along the sample axis; incompatible dimensions or metadata produce clear errors.

  • Preserve latent metadata, noise masks and batch indices through list/batch conversion, including broadcast masks, mixed masked and unmasked samples, and matching video batches. Small-width masks retain their batch axis.

  • Keep scalar pipe values whole, retain enabled zero-valued generation settings, and keep Slice & Dice conditioning aligned with selected images and text.

  • Reject impossible Image Align Size Shrink requests with Grow/Resize guidance, keep reported image and mask dimensions consistent, and assign cumulative SEGS frame indices across mask batches.

  • Preserve selected newline delimiters in Join and Merge Strings, and preserve floating-point frame data when Frame Consistency effects are disabled.