H3 Interface 2.16.0 - Completed release and queue safety

H3 Interface 2.16.0 completes the reviewed video/image integrations and closes release-blocking queue and UI issues.

• Prevent shared-GPU pending-worker deadlock and fail closed on unknown peer queues.
• Keep generation available while native queues wait for the shared GPU.
• Remove unfinished FaceRefine controls; keep unsupported requests explicit.
• Fix grouped-reference reorder buttons covered by drag grips.
• Preserve delayed reference audio, ordered adapters, Pika/Singularity settings and supported refinement.
• Continuously test eleven real-API browser journeys alongside Windows/Linux Python 3.12/3.13, JavaScript and archive integrity checks.

Validation: 623 Python tests, 340 JavaScript tests, 528 + 2,096 connected graph cases and eleven local browser journeys with simulated inference. Platform-specific skips are documented. The repository owner confirmed their rendering workflows; no new GPU measurements are invented by this source audit.

See docs/RELEASE_2_16.md for evidence and supported boundaries. Runtime/model pins are unchanged. Model weights are not bundled.

Download the ZIP, verify it against SHA256SUMS, and follow START_HERE.md. Existing installations should update only when their queues are idle.

H3 Interface 2.15.0 - Named elements, fresh project tabs and storage

2.15.0 - Named elements, fresh project tabs and storage

Released 4 November 2026. Model/runtime pins are unchanged.

• Project-element uploads require explicit names before uploading. Normalized current @mentions must be unique. The live catalog exposes current named mentions, without filename or previous-name synonyms. Direct shot uploads reserve only active reference aliases, preventing renamed catalog filenames from hijacking new references. Stored asset identities and historical take bindings remain intact.
• Remove is visible beside existing Replace controls. Project removal hides the reusable catalog entry; its media remains available to saved shots, results and exports. Shot-reference Remove detaches that card. Reference actions have a separate row; catalog actions use a compact grid.
• A fresh project tab starts with empty prompts, references, frames, guides and continuation. It inherits generation mode, canvas, LoRAs, sampling, Dials, refinement and other independent controls from that project's latest completed take with a stored draft. Soundtrack-source bindings and ControlNet inputs are not copied. Existing tab drafts and explicit Reuse retain their full setup. Other/new projects do not inherit these controls.
• Storage in both page headers reports real free/total bytes on the server output filesystem, without exposing paths. It does not count model disks separately. Updates run once per minute while visible, with bounded requests and no recursive scans.
• Clarify base checkpoints versus managed Pika Turbo adapters. Pika Turbo is distributed as LoRAs for compatible H3 bases. Singularity v2 is a replacement checkpoint; its approximately 5,000-character guidance is an upper-range recommendation, not a minimum. Creative adapters remain optional; supported Pika combinations retain automatic sampling.

Verification: 609 Python tests (608 pass, one platform skip), 327 JavaScript tests, existing 528-case plus 2,096-case graph matrices, and actual local browser/API journeys with simulated inference. The new journey checks required naming, filename isolation, visible Replace/Remove, persistent catalog removal, empty-input fresh tabs, settings inheritance, refresh retention, project isolation and live storage. Existing core and selection/recovery browser journeys also pass.

The launcher continues to follow official main, validate pinned artifacts and reuse verified shared files. No machine-specific credentials or paths were added. ComfyUI stays pinned to 0.43.0 pending GPU validation of 0.44.0. No new GPU render, measured GPU speedup, arbitrary-hardware guarantee or universal error-free claim is established.

H3 Interface 2.14.0 - Verified selection paths and resilient recovery

2.14.0 - Verified selection paths and resilient image recovery

Released 4 November 2026. This release hardens supported UI-to-graph contracts. Model weights and runtime pins are unchanged.

• Output soundtrack reuse has an explicit video source selector. The chosen reference alias survives reorder, save/reload and settings reuse. Duplicate uploads with different trims remain distinct. Missing or ambiguous sources block submission; leaving References clears interactive reuse. Selected source audio must cover the output duration.
• Backend admission verifies declared media filenames, native reference slots and paired audio trims, frame loaders, CLIP/VAEs, exact model/LoRA/Dials/shift/accelerator chains, refinement conditioning and seed, continuation source/trim wiring and relevant capabilities. Unsupported graphs fail before queue admission.
• Reference and guide uploads use separate asset maps, so aliases such as source, mask and control cannot overwrite reserved control inputs. Submitted metadata includes resolved input filenames and the compiled prompt.
• Known LoRA family restrictions use shared selection rules. Busy/project-loading configuration stays locked. Text/Frames generation ignores inactive reference trims.
• Qwen recovery bounds network waits, avoids overlapping pollers and ignores stale responses. Metadata retries retain the finished image, observed render time and unlocked controls without repeated preview downloads. Cancellation cleanup is bounded.

Validation: 605 Python tests (604 pass, one platform skip), 323 JavaScript tests, independent backend/JavaScript/security reviews, the existing 528-case and new 2,096-case supported graph matrices, and real local browser/API journeys with simulated inference. Browser journeys cover explicit soundtrack selection and persistence, submitted source wiring, malformed request rejection, mode transitions, Qwen retry/unlock, storyboard/layout/reference behavior, continuation and connected controls.

These checks establish construction, admission and interface behavior for tested supported paths. No new GPU inference, measured speedup, perceptual quality comparison or universal bug-free guarantee is claimed. ComfyUI 0.44.0 is the next GPU-validation target; the tested runtime remains pinned to 0.43.0. See upstream audit.

H3 Interface 2.13.0 — Storyboard sequences and personal layouts

2.13.0 - Storyboard sequences and personal layouts

Released 4 November 2026. This release changes interface organization and reference preparation. Model and ComfyUI pins are unchanged.

• Storyboard sheet uploads one contact-sheet image. Storyboard sequence uploads separate images into one ordered group. Each member retains its own native image slot; one group mention includes the group's images in Guided compilation. The group order provides narrative progression, not exact timing. Character, environment and style references supply final appearance separately.
• Spatial plans remain semantic camera/subject-route guidance. No diagram, board or sequence becomes an opening frame or timed guide automatically. Explicit Temporal keyframes and Frames mode retain their native conditioning routes.
• Customize layout supports stage/section order, cross-stage moves, visibility and reset. Separate video/image browser preferences persist. Hiding controls preserves values; mode/capability rules still apply. Validation reveal opens relevant hidden sections.
• Reference cards/groups, Qwen references and timed guides support drag handles, Alt + arrow keys and button alternatives. Numeric references follow their original media after reorder. Explicit native Subject definitions preserve their identities. Guide reorder leaves timestamps unchanged.
• SortableJS 1.15.7 is bundled offline under MIT with pinned source and file hashes in web/h3/vendor/sortable-manifest.json; no CDN or new API key is needed.

Validation: 596 Python tests (595 pass, one platform skip), 314 JavaScript tests, independent Python/JavaScript reviews, and actual local browser/API journeys with simulated model execution. Browser checks include real mouse reference dragging, keyboard order changes, sequence upload/grouping, synchronized modes, saved project reload, layout persistence/reset, hidden-control reveal, separate native image inputs and no implicit AddGuide/first-frame insertion. Existing reference, spatial, workspace, image, continuation and connected-control journeys are also checked.

The persisted 528-combination graph regression matrix checks supported selection contracts; it does not prove every arbitrary combination or generated result. No new GPU inference, perceptual quality measurement, calibrated trajectory control or universal bug-free claim is established by this release.

See reference controls and storyboard guidance.

H3 Interface 2.12.0

Spatial plans and storyboards now have distinct reference guidance.

• Optional Spatial plan image role binds to a Picture and transfers only requested layout/route relationships. It suppresses diagram annotations and does not invent calibrated camera control.
• Storyboard sequence structure offers Follow the prompt, One continuous take and Separate shots. Sequence choices persist in drafts, projects, saved settings and local preparation metadata. Complete native prompts retain authority.
• Both inputs remain native image references. No automatic first-frame/AddGuide routing, no new required model or external pack, and no API key required for native compilation.

Verified locally: 595 Python tests (594 pass, one platform skip), 304 JavaScript tests and a real local browser project/upload/graph journey with simulated inference. No new GPU rendering or camera-adherence claim.

Storyboards, spatial plans and timed images · Release details

The ZIP contains source and installers, not model weights. SHA256SUMS and the per-file manifest are attached.

H3 Interface 2.11.0

Production now has four clearly separated, independently collapsible stages. Each section folds individually; tab-local preferences survive refresh. Workspace tools start compact and remain available.

Upload alternate views as one Character, Environment, Element or Style group. Previews appear side by side; one Guided mention connects every picture while retaining individual inputs and instructions. Native structured definitions keep authority. Group-local ordering and visible Picture indices clarify bindings.

Fix restored reference controls remaining locked, reveal the prompt for inserted mentions and validation errors, and reject mixed-role groups and case-insensitive duplicate aliases.

Validation: 594 Python tests (593 pass, one platform skip), 299 JavaScript tests and real local browser project/upload/graph/media journeys with simulated inference. No new GPU rendering. Model and runtime pins remain unchanged.

Release details · Reference controls

The ZIP contains source and installer scripts. Model weights are not included. SHA256SUMS and the per-file manifest are attached.

H3 Interface 2.10.0 - managed upstream workflow update

H3 Interface 2.10.0 - managed upstream workflow update

Date: 2026-11-04. Source/CPU and local browser verification; no new GPU deployment or model render.

• Add Pika two-step Draft (References), four-step Lite (both families), canonical publisher URLs and exact per-file sampling. Mode changes retain Lite; Draft has no Text/Frames counterpart.
• Add Character Swap 1.5, retain V1 and explicitly selected legacy scenes, and prevent both versions being applied together.
• Add opt-in Singularity v2 pruned INT8 with a Windows/Linux launcher option and registered discovery. Pika References can retain this exact selected checkpoint; other unsupported profiles remain blocked.
• Update the independent optional MAINodes pack to 1.2.2 and the Dials source pin. Studio's existing Dials integration remains separate from MAINodes' MAI tools. Independent RelayVSR CPU contracts passed; its GPU workers/models and full pack import remain unverified here.
• Audit all 28 managed base/refine/control/Qwen-profile artifacts against current publisher metadata; retain unchanged immutable weights. Added LoRAs require 4,687,800,976 uncached bytes. Singularity v2 is a separate explicit 20,967,647,456-byte download. Source archives do not contain model weights.
• Document researched motion/segmentation/depth tools and a concrete local-agent architecture with implemented/proposed boundaries. No cloud API key is required for current deterministic coordination; an autonomous semantic reviewer needs a configured local model.

Verification covers all managed Pika graph variants, backend graph mutations, actual Studio checkpoint reconciliation and Character Swap version switching. Local browser checks exercise Swap 1.5 and Draft/Lite/HD on Singularity v2 through real project/asset/job routes with simulated capability/inference fixtures, plus save/reload. Unit/source checks and fixtures do not establish GPU memory, speed, identity retention, perceptual quality or physical correctness. GitHub Actions records Windows/Linux Python 3.12/3.13 results for the exact published commit.

See upstream decisions and orchestration assessment, Pika, Character Swap, selection rules and historical GPU evidence.

H3 Interface 2.9.1 - reference workflow fixes

H3 Interface 2.9.1 - reference workflow fixes

Date: 2026-11-03. Focused reference workflow corrections; no new model or GPU deployment.

Behavior

An image assigned to a Character group could keep its group after the Character Swap shortcut changed its role to Custom. Prompt compilation then rejected the stale group. The shortcut now clears that group, matching the manual role selector. It retains the per-reference instruction, original-resolution override and existing sampling choices.

A Linux FFmpeg 8.0.1 compatibility check also exposed a stall while preparing a video whose audio begins late. Video and audio now share one bounded filter graph. The existing delayed-audio regression completes while preserving opening silence, later sound and exact frame count. The original command stalled locally and in CI; the corrected command completed in approximately 0.35 seconds on the local CPU fixture.

Verification

All 286 JavaScript tests pass, including a regression that executes the actual shortcut and prompt compiler. Independent code and JavaScript reviews cover the shortcut correction. The media correction retains existing timing, fit and frame-count tests; all 11 continuation-upload and bounded-padding tests pass locally on Windows FFmpeg 8.1.

A local browser journey exercised the visible controls: create a Character group, retain an image instruction and resolution override, add a source video, apply the replacement preset, submit the native reference graph with the correct Character Swap LoRA, cancel, save and reload. The graph and restored scene matched the selections; no browser JavaScript errors occurred. Inference was simulated, so this does not measure generated quality or GPU compatibility.

GitHub Actions checks Windows/Linux on Python 3.12 and 3.13 for the published commit. The preceding 2.9.0 verification documents the broader 589-test Python suite, persisted graph matrix and browser coverage. No finite suite establishes every possible configuration or guarantees zero bugs.

H3 Interface 2.9.0 - References, workflows and installer safeguards

H3 Interface 2.9.0 — reference direction and transparent installation

Date: 2026-11-03. Source, CPU and browser verification; no new GPU deployment or model render.

Changes

• Rename the visible product to H3 Interface and publish a separate public documentation site. Preserve h3_studio URLs, storage identities, existing projects, history and legacy official Git origins during upgrades.
• Group multiple Character pictures into one native Subject. Preserve independent default subjects and explicit native subject definitions. Forward group identity and every Picture relation through optional local/remote AI preparation, with regression coverage.
• Add independent instructions for every reference role, separate storyboard panel ordering, per-image original-resolution override and selected-range audio/video preview. Scene save, local restoration and Reuse preserve these settings. Complete Structured/native prompts retain directive authority and expose the card-guidance boundary.
• Recompute dynamic LoRA and capability controls after project loading and source import; prevent disabled loading controls from remaining stuck or overriding continuation guards.
• Reject unknown optional-node edits even when HEAD matches the pinned revision, including ignored executable additions. Recognize only the exact owned MotionCache transformation; preserve unknown user changes.
• Require HTTPS for external keyed prompt providers, reject HTTPS downgrade redirects and strip Authorization on cross-origin redirects.
• Publish installer behavior, privilege/update boundaries and upstream provenance. Preserve Charles Mod's MIT copyright and permission notice; add H3 Interface contribution attribution. Model and external-pack rights remain separate.

Verification

The local Python suite ran 589 tests: 588 passed, one platform skip. All 285 JavaScript tests passed. Existing graph-contract regression includes the persisted 528-combination matrix. Security and JavaScript reviews found issues that were fixed and covered by focused regressions.

The new reference browser journey uses real local project/asset/resolve APIs with simulated inference. It verifies group mapping, instructions, original-resolution settings across reload/Save/Reuse, actual derived input dimensions (2304×1408 original versus 2048×1252 auto-fit), selected source playback stopping and media resource cleanup. Existing general, continuation and Qwen browser journeys were brought into line with explicit project creation and current canvas/upload protocols; they exercise cancellation, response loss, saved input restoration, continuation guides and image creation/edit graph capture.

The public site was inspected at desktop and mobile widths, with index/follow metadata, canonical URL, sitemap and SoftwareApplication structured data. These make public content available for discovery; they do not prove Google indexing or guarantee ranking. GitHub Actions provides Windows/Linux Python 3.12/3.13 checks for the exact published commit.

No finite suite guarantees zero bugs, arbitrary external-LoRA compatibility, GPU speed or perceptual quality. Native reference encoding can still resize original-resolution images. Automatic workflow coordination is deterministic; optional AI preparation needs a configured model. See reference controls, installer transparency and compatibility.

H3 Studio 2.8.1 - English guides and connected workflows

H3 Studio 2.8.1

English-only documentation and interface help, automatic supported workflow selection and stricter connected-input validation.

• Rebuild the README around current setup, supported features and verification boundaries. Translate/rename 21 guides, remove repeated update sections and correct stale Attention/Turbo instructions.
• Apply Pika/Turbo prerequisites automatically across supported mode/profile changes; restore saved setups atomically and block incompatible selections visibly.
• Connect optional pinned Matlow Dials to Studio and project persistence; install MAINodes independently. MotionLab/Contact Sheet are backend tools, not integrated automatic Studio passes.
• Honor saved-image auto-fit in generation and prompt preview while preserving original assets, CMYK input and RGBA transparency.
• Validate connected guide timing/types and timeline bounds, reference detail, frame rate, Full HD crop and refinement resizing.
• Keep separate video/standalone-audio reference budgets, Qwen canvas/LoRA fixes and official upgrade provenance.

Verification

585 Python tests (584 passed, one platform skip), 277 JavaScript tests, 528 persisted graph combinations, 48 independently checked continuation combinations and local browser checks. CPU/local API/browser evidence is separate from inference. The final exact-commit Windows/Linux checks are linked below.

No target server was deployed and no new GPU render or universal visual-quality/performance guarantee is claimed. Weights are not bundled. External optional packs keep their own licenses.

Installation · Selection rules · CI

ZIP SHA-256: 88dc3d2c3ce654f45d97a2054dc57096ad37a9006bcb29fd71878c8c1a0abdaa

Source commit: 14182b755151a9360e62d3aedb766c3fd41735aa