Studio patch notes

Daily Studio development history, from the first working prototype onward.

Versions

0.10.19 · Introducing Photo Repair Studio

October 19, 2026 · Alpha

Added

Photo Repair Studio is now available to signed-in Studio users. Upload or paste a photograph and use Auto Repair Photograph to assess, prepare and repair it.
Repair for damaged and faded photographs, with optional black-and-white and sepia colorization. Colorize is unavailable for detected colour photos.
Optional Fix framing to complete people clipped at the top or sides.
Original, Prepared, Result and Compare views, image dimensions, processing status and download.
Advanced controls for manual crop, edge snapping, right-angle rotation and output enlargement. Default output follows the prepared photo size, with a 2048-pixel longest-side limit.

Privacy

Photos are processed temporarily in memory and are not kept in a Photo Repair library. Download your result before starting another image or closing the page.

Studio interface

Added Photo Repair Studio to the shared Studio navigation and header.
Studio Help is smaller and neutral, beside the AI-machine resource panel. Voice Conversion now includes Help and AI-machine resources; Voice Clone includes Help without the AI-machine readings.

Alpha limitations

Repairs and inferred colours need review. AI can change small details, lettering or framing, even when dimensions match. Returned images are shown for comparison rather than blocked by the previous framing check. HD modernization and Enhance are not included in this release.

0.10.17 · Usability, resources, and help

October 17, 2026

Added

Added 15-minute inactivity sign-out for regular accounts, requiring a fresh login. Background polling does not count as user interaction; administrator accounts are exempt from automatic session expiry.
Added Clear page beside reference selection in Refine and Image to image to reset the workspace without deleting saved files.
Added Uploaded images and Generated images tabs in image file management.
Added the OPAI - Server resources panel and a GPU Core percentage with proportional purple fill in image-generation popups.
Added a Patch Notes link directly beneath the slightly enlarged ALPHA label on Image Studio and Combine Images.

Changed

Made chat’s open/closed preference persist through navigation and synchronize across tabs for the account during the login session.
Aligned Combine Images navigation and workspace button sizing with the other image tools and added its generation-progress popup.
Capped Image to image and Combine output dimensions at 2048 pixels per side. Combine shape selections now default to a 1024-pixel longest side: square 1024 × 1024, portrait 704 × 1024, landscape 1024 × 704.
Added an aspect-ratio lock between Combine Width and Height. Locked dimensions scale together; unlocking permits custom proportions. Selecting a standard shape restores its linked recommended size.
Moved Image to image’s large reference preview into the advanced section, available to administrators, while masking remains disabled in that workflow.
Made the Refine label yellow with a warning symbol and emphasized small, targeted edits. Broader changes are directed to Image to image.

Refine usability

Renamed Send to Generate prompt and apply masking. Added bold green Mask applied feedback and green confirmation below the image when automatic selection succeeds.
Added copy-to-clipboard access to request prompts for reuse.
Explained Change strength and moved the detailed guidance behind an expandable More section.
Compacted the manual masking controls into one row where screen width permits, and corrected the reference-selection/Clear page layout.

Help and documentation

Added concise tool introductions, hover explanations, and expandable how-to content with more readable text.
Expanded imaging instructions for AI Full, AI Assist, Manual, editable prompts, generation settings, specialist models, and LoRAs, with links to actual examples.
Moved the imaging guide to imaging.html and made Start Here a Studio introduction and navigation guide.
Added detailed Voice Clone, Voice Conversion, and Combine guides; reviewed existing editing instructions against the current controls.
Made topic indexes compact, grouped related subjects, and added Back to top navigation.
Expanded Built by James OP to describe the controllers, gateways, connected systems, and work behind the wider OPAI project.
Published retrospective daily developer notes using 0.month.day, beginning with the October 9 prototype.
Made patch notes available without signing in and added a Patch notes link on the sign-in page.

Still in progress

Automatic masking and model suitability for targeted edits remain inconsistent. A successful mask application does not guarantee the correct selection or a successful visual change.
The resource meter shows shared server activity, not a particular job’s completion percentage. The faster resource-stream update remains prepared rather than confirmed installed.

0.10.16 · Voice Conversion Studio

October 16, 2026

Added

Introduced Voice Conversion Studio: convert an existing speech recording using a selected target voice, with source playback, result playback, and audio download.
Connected predefined, imported/shared, and personal voices to the conversion workflow through existing account permissions.
Added source and reference uploads with duration, format, size, and audio-validity checks. Sources support up to five minutes; target references use a selected 3–30-second passage.
Added waveform clipping, editable start/end times, and Listen to selection for checking the exact passage before conversion.
Added conversion cancellation, status reporting, and a dedicated Voice Conversion collection in File management.

Changed

Matched the conversion page to Studio’s visual style with its own Ellie header and grouped Image Studio / Voice Clone / Voice Conversion links.
Limited temporary conversion storage to the latest three source uploads, three uploaded references, and three results per account. Cleanup protects active job inputs and leaves saved voice libraries intact.
Added an explanation of source versus target audio, accent and delivery limitations, clipping, cancellation, and temporary storage.

Fixed

Corrected loading of voice references from shared storage with a dedicated audio loader and validated file access.
Hardened submission handling to reconcile uncertain requests instead of blindly submitting another conversion.
Adjusted clip boundaries to stay within measured recording duration and made validation feedback clearer.

Engineering and validation

Installed the conversion engine as a separate GPU service, leaving the image engine and NAS Voice Clone engine separate.
Recorded successful short conversion runs and verified the required backend nodes and protected Studio routes.
Documented the installed voice services, shared libraries, recovery procedures, and reusable Studio branding.

Known limitations at this stage

Complete signed-in conversion/download/cancel acceptance after the final fixes, long-recording continuity, and real-job retention behavior still required further testing. Successful short conversions did not establish perfect voice similarity or delivery preservation.

0.10.15 · Voice Clone, Combine Images, and shared file management

October 15, 2026

Added

Launched Voice Clone Studio with predefined voices, imported/shared voices, and personal voice imports under the existing Studio login.
Added text-to-speech generation, vocal-expression buttons, speed selection, MP3/WAV output, playback, and saved recordings. Long scripts are processed in chunks and returned as a complete recording.
Added private voice imports and permanent community sharing, with community-voice deletion restricted to the Studio owner.
Introduced Combine images as a separate workspace with two required references and an optional third, saved-image selection, replacement/removal controls, scene instructions, and square/portrait/landscape output choices.
Added clipboard image paste to Image to image and numbered Combine slots. Clicking a slot selects the paste target; ordinary text pasting in prompts and chat remains available.

Changed

Replaced My history with File management, bringing Image and Voice Clone collections together with downloads, individual deletion, multi-select, select-all, and bulk-delete progress.
Limited the Voice Clone page to the three most recent job entries, with older recordings available through File management.
Added matching Ellie branding to Image Studio and Voice Clone Studio.
Shared online presence and tester chat across both pages. Updated Open Chat, Close Chat, and Popout chat labels and grouped Studio-switch links beside the chat controls.

Fixed

Corrected a voice-upload routing restriction that rejected an existing reference before it reached the speech engine.
Fixed repeated voice-generation memory failures by bounding the conditioning cache and enabling inference-only execution. All 51 tested voices completed consecutively after the fix; only five had passed the earlier sequence. Original samples were retained.
Corrected deployment checks that misread the NAS mount or checked the wrong Studio health address. The failed Combine deployment rolled back before a corrected retry.

Engineering and validation

Moved Studio speech generation onto the NAS GPU through the Chatterbox Turbo server, separating it from image-generation hardware.
Connected multi-reference generation to the existing image queue with reference ownership, distinct-image, and revision checks.
Verified the clipboard update’s installation record and tested selected-slot replacement, filling slots out of order, busy-state protection, and text paste.

Known limitations at this stage

The voice-engine memory patch survived restart but needed to be preserved when rebuilding its container.
Removing a voice from Studio did not yet guarantee removal of the speech engine’s separate reference-cache copy.
Combine created a new interpretation of the reference images; exact faces and details were not guaranteed.

0.10.13 · Dedicated application installation and sign-in improvements

October 13, 2026

Changed

Moved Studio and its own Python runtime, image-processing dependencies, and local masking models out of the staging release into a dedicated application installation on the system drive.
Switched the Studio service to the new installation while retaining the original files for rollback. Existing accounts, image storage, generation models, and external engines remained in place.
Updated sign-in to accept either username or email without case sensitivity. Passwords remain case-sensitive.

Editing work

Developed the Manual masking tools grouping, clearer automatic-mask wording, brush controls, and Undo / Ctrl+Z behavior.
Investigated a dress edit that failed despite manual painting. Inspection confirmed the submitted mask included the painted beltline; the requested color change failed in generation rather than being lost by the brush.

Validation

Checked the copied runtime and masking-model loading before switching the service. Local and public page checks passed, and the owner subsequently confirmed Studio was working.
Verified the installed username/email sign-in update and corrected the browser form to accept a username without an email address.

Known limitations at this stage

Automatic masking could still take roughly a minute, and correct masks did not ensure reliable recoloring. Higher strength sometimes changed garment shape without achieving the requested color.
Moving the application did not establish a measured masking-speed improvement; the masking worker still ran on the CPU.

0.10.12 · Help launch, image management, and Refine usability

October 12, 2026

Added

Published the authenticated Studio help site with Start Here, Refine Image, Image to Image, More help, and Image Examples.
Built a reviewed gallery of 45 examples across AI Full, AI Assist, Manual, model comparisons, and Qwen edits, with prompts and actual results.
Added illustrated editing walkthroughs and replaced the Refine success example with the owner’s before, automatic-selection, and successful blue/translucent-wing screenshots.
Added image multi-selection, select-all, permanent deletion, and completed-item progress during bulk cleanup.

Changed

Made output size more prominent beside Generate and compacted the request/output controls.
Improved model-choice labels and buttons and made positive/negative Prompt Details editable while collapsed by default in assisted modes.
Moved reference selection before the Refine request and blocked preparation until an image was selected.
Renamed Refine’s Denoise control to Change strength, displayed it as a percentage slider beside brush size, and kept it available in assisted and manual modes.
Set Refine’s general starting strength to 60% and made side-by-side comparison the default when no saved preference existed.
Refined the shared chat dock, pop-out behavior, unread indication, and login workspace reset while retaining saved history.

Fixed

Changed wing localization to identify existing wings rather than requiring them to already match the requested new color or transparency.
Improved plural-target selection by finding separate visible objects and reviewing mask candidates one object at a time.
Retained bounded mask-failure diagnostics so selection failures could be investigated after temporary processing files were removed.
Extended permanent-image cleanup to registered files and cached previews instead of merely hiding history entries.
Clarified content-screening outcomes and added a second review for inconclusive classifications rather than treating uncertainty as a definite violation.

Validation

Confirmed installation of login-protected help and rejection of anonymous help/image access.
Recorded successful owner tests for blue/translucent wings and multiple mushroom recoloring.

Known limitations at this stage

Stronger edits could leave old outlines or change the selected shape. The reported 1536-square generation error remained unresolved; no cause was established.
Successful examples did not establish reliable automatic masking for every subject.

0.10.11 · Saved references, prompt handling, administration, and chat

October 11, 2026

Added

Retained new uploads with ownership and attribution records and added My images for reusing uploaded or generated references.
Added lightweight gallery thumbnails and cleanup of obsolete thumbnails while preserving original images.
Added fresh random seeds for new generations unless Lock seed for repeatable tests is enabled.
Expanded administration with Users, Recent activity, and Logfiles, downloadable logs, administrator promotion, and account removal that revokes access while preserving attribution.
Added persistent group-chat history with older-message pagination, Message of the day, administrator moderation, pop-out chat, and incoming-message notifications.

Changed

Simplified dedicated Image to image so the selected reference and literal edit request could go directly to the Qwen editing workflow.
Reworked Refine preparation to retain the latest edit request instead of replacing it with long generic style instructions; necessary LoRA triggers were handled separately.
Adjusted whole-image style preparation to offer fresh compatible model choices unless the request explicitly retained the current model.
Used 60% as the tested whole-image style starting point after lower settings were too weak and higher settings caused excessive reconstruction.
Restored Create’s use of the established bundled creative compiler after overly broad prompt restrictions reduced useful enhancement.

Fixed

Corrected menu/rendering behavior that could leave controls incorrectly enabled or the interface unresponsive.
Added a scoped negative-prompt fallback for Refine when an SDXL model supplied no negative prompt, preserving existing nonempty negatives.

Validation and development findings

Recorded an owner-confirmed purple-wing recolor that retained the surrounding scene.
Rejected the long forced anime-prompt approach after severe visual drift. A proposed vision-captioning replacement was staged but not installed.
Content-screening work and an eye-mask/status patch were prepared; the day’s record did not establish deployment of those particular packages.

Known limitations at this stage

Eye/glow selection remained unreliable, and some style transformations changed the subject or scenery. An applied mask or passing code test was not treated as proof of a successful visual edit.

0.10.10 · Hosted alpha and multi-user foundation

October 10, 2026

Added

Deployed the Studio alpha on the AI server and made it available through the authenticated public Studio address, removing the need to keep the Windows preview running.
Added registration with administrator approval, login/logout, account administration, password-reset links, and protected sessions.
Added account-owned workspaces, private image history, and saved generation metadata.
Added the shared tester lounge and online-user display.
Added a persistent generation queue with frozen job settings, owner/workspace association, and handling for existing backend activity.
Connected actual backend sampling progress to the generation indicator while retaining an indeterminate state when progress was unknown.

Changed

Preserved imported reference dimensions and refined the separation between the generation mask and final softened blend.
Simplified Refine’s selection rule: an existing mask limits the edit; an empty mask permits a whole-image edit.
Packaged Studio’s masking dependencies separately from the existing generation environment.

Fixed

Allowed spaces in display usernames during account setup.
Corrected a launcher check that could report an occupied port after the previous process had stopped.
Corrected public-origin configuration that rejected domain-based logins despite valid credentials.

Validation

Confirmed the hosted service and public address, checked the login redirect and health endpoint, and verified anonymous history access was rejected.

Known limitations at this stage

Refine remained inconsistent for clothing replacement, hair effects, and pigmentation edits. Increasing strength could distort anatomy or pose.
Email ownership verification and automatic password-reset email delivery were not connected; administrator-issued reset links supported the alpha.
Broader administration and clipboard changes were prepared late in the day; their preparation alone was not proof of installation.

0.10.9 · First working Studio prototype

October 9, 2026

Added — local prototype

Built the first OPAI Image Studio test interface after evaluating existing frontend options and testing the installed image-editing backend.
Introduced Create image, Refine image, and Image to image as separate workspaces with their own requests, prompts, settings, and results.
Added AI Full, AI Assist, and Manual modes, model recommendations, editable prompts/settings, and visible applied LoRA names and strengths.
Added painted-area refinement, reference upload, brush sizing, and Use last generated image across the workspaces.
Connected a separate Qwen instruction-edit workflow and tested transfer of a generated image between workspaces.
Added preparation/generation indicators, elapsed time, busy-state control locking, Enter to send, and Shift+Enter for a new line.

Fixed during prototype testing

Corrected model filtering and selected output dimensions in the initial frontend evaluation.
Reduced stale request, effect, and model state leaking into new image requests.
Added Clear / New request and aligned recommendation messages with the current preparation.
Connected compatible effect LoRAs to Refine requests while retaining the selected model where appropriate.

Validation and scope

Tested image creation, masked refinement, and Qwen editing, including reference reuse and isolated workspace state.
This was a local development preview using the existing AI server for model work—not the public multi-user alpha.
Accounts and durable per-user history remained future work. Automatic selections and several visual effects still needed substantial tuning.