Only the submit button triggers generation by default. The "Enter to
generate" toggle (available in both app and website settings top right
nav) still allows opting into Enter-to-submit.
- Pressing **Enter** in any prompt box now adds a new line instead of
submitting - fixes complaints about accidental generations
- **Shift+Enter** is the new submit shortcut
- Desktop app gets an "Enter to generate" toggle in Settings → General
(off by default) for users who prefer the old behavior
- Applies to all prompt boxes in the Artcraft desktop app and the
website
## Artcraft website: unified lightbox, recreate, desktop CTA, prompt
fixes
**Lightbox**
- Gallery modal and `/media/:id` page now share one details component -
no more drift.
- Author sticks to the top when you scroll; `X` moved to the details
panel; "Show more" prompt toggle added to the media page.
- Shows aspect ratio, resolution, duration, and audio on/off.
**Recreate (new)**
- Button in the lightbox restores prompt, refs (image / end-frame /
video / audio), model, aspect, resolution, duration, audio toggle, and
input mode from a past generation.
- Routes to Create Image vs Create Video based on the saved model class,
not URL guessing.
- Does not auto-generate.
**Desktop CTA**
- Dismissible banner above the promptbox linking to the desktop app.
Hidden on mobile, reappears 24h after dismissal.
**Model names**
- Model selector uses `full_name` from the OmniGen API so new models
show up correctly without a code change.
- Added Seedance 2.0 Fast to the shared model list for the lightbox
details panel.
**Promptbox `@mentions`**
- Triggers correctly after Chinese, punctuation, or emoji - not just
whitespace.
- Case-insensitive: `@image1` highlights as `@Image1`.
fixes:
- prompt box no longer overlaps bottom buttons or wastes space on big
screens
- long unbroken text wraps instead of escaping the box
- resizing the window updates the layout without needing to refocus
getCurrentAspectRatioIcon() was reading from aspectRatioList (useState),
which is only updated on manual picker clicks, not external store
writes. Reading aspectRatio from the store directly fixes the stale
icon.
## Video reference library picker
- Added "Pick from library" option to the Video Ref slot, with gallery
filtered to videos
- Added loading state to "Use selected" button in gallery modal
- Enforced seedance-2 limits (max 3 videos / 15s total) for both upload
and library flows, with a deduped video-ref-limit toast so users see a
warning on count overflow, duration overflow, or submitting when already
at max
Two bugs were preventing @image / @video / @character references from
being tagged in prompts written with Chinese, Japanese, or Korean text:
1. The mention-highlight regex was built with just the "g" flag, so a
user-typed "@image6" never matched the canonical label "@Image6" and the
reference was rendered as plain text. Make the regex case-insensitive
and do the colorMap lookup via a lowercase-keyed table so "@image6"
still finds the color for "@Image6".
2. detectMention required the character immediately before "@" to match
/\s/, which never matches CJK characters. Typing "@" right after a
Chinese character like "从" therefore never opened the autocomplete
dropdown, leaving users to type the mention by hand (which then hit bug
1). Replace the whitespace check with a "not an ASCII identifier char"
check so CJK, quotes, and punctuation all count as valid word boundaries
before an @mention.
- **Characters for Seedance 2.0** - Character creation/selection modal,
@-mention support with colored inline highlights via contentEditable
MentionTextarea, character tokens sent in video generation API requests
- **Mobile responsive navbar** - Hamburger menu with slide-down panel,
all nav items accessible, credits/library/sign out included
- **Mobile responsive PromptBox** - Tighter padding, stacked toolbar on
small screens, flex-wrap for option buttons
- **Mobile responsive library** - Proper sticky header offset, icon-only
filter tabs on mobile
- **Task queue overhaul** - Prompt text with marquee hover, copy button,
model creator icons, detailed failure reasons on all card types
(matching desktop app)
- **Gallery improvements** - 512px thumbnails (was 250px), 12px grid
gap, model + media type badges on hover, model icon + prompt on
in-progress cards
- **Quality selector** - QualityPicker for image models
(High/Medium/Low), wired to cost estimate and generation API
- **Video batch enqueue** - Multiple videos enqueue as separate jobs so
each completes independently
- **Form consistency** - All auth pages use shared `<Input>` component,
`<form onSubmit>` for Enter key, consistent styling
- **Bug fixes** - Lightbox type error, library carousel disabled,
popover icon alignment, end frame library picker, image ref max default
(1 → 6), task queue mobile overflow, clear session button removed
- **Mobile model selector** - Model picker now appears in the PromptBox
toolbar on mobile/tablet (`lg:hidden`), since the floating bottom-left
selector is desktop-only
- **Responsive image prompt row** - Start frame and end frame sections
stack vertically on mobile with a horizontal divider, instead of
side-by-side
- **Responsive lightbox** - Stacks vertically on mobile (media top, info
below), nav arrows always visible (not hover-only), sidebar goes
full-width, darker backdrop (`bg-black/80`)
- **Safari video inline playback** - Lightbox videos use
`disablePictureInPicture`, `controlsList` restrictions, and
vendor-prefixed attributes to prevent fullscreen/PiP pop-out on Safari
When listUserMediaFiles fails (HTTP error, network error, null
reference, or any other exception), the gallery now shows the error
message with a "Copy error" button instead of silently falling through
to "No items yet".
This is a diagnostic aid for a macOS-only blank gallery bug where the
root cause is unknown (CORS, session cookie, or runtime exception) and
there is no access to production logs.
setInitialLoading(true) and setUsername(name) are batched together so
React never produces a frame where username is set but initialLoading is
still false, which would fall through to the No items yet branch before
refreshGallery has a chance to set initialLoading(true) itself.
- Remove useless forceFilterRef and its stale useEffect; use forceFilter
prop directly in JSX
- Show skeleton immediately while username is being fetched (!username
condition) instead of flashing "No items yet"
- Track username fetch failure with usernameError state and surface an
error message with a working Retry button
- Fix retry: setUsername("") was a no-op (same value → React bails out);
use usernameRetryCount in deps to actually re-trigger the fetch
- Guard error UI with allItems.length === 0 so stale items stay visible
if a retry fails mid-session
Notable behavior:
- Mixing supported and/or unsupported file types will only keep files of
file type that match the first supported file type in the list received
from browser (e.g. PNG, OBJ, EXE, JPG, OBJ order selection of files will
only show PNG, JPG for image upload)
- Removing files until only one is left will result in the sidebar
vanishing and only a single preview showing. More files can be added by
selecting multiple files again through the file selection button
- Failure to upload has been handled but not tested. It is expected any
failed uploads could be retried with a "retry failed" button. Any
individual fail can also be retried via a retry button upon hovering on
the exclamation mark by the failed file's name
- All multi-file uploads are queued async and likely wouldn't be ordered
as shown in the preview list (first-come-first-serve)
- Potentially deadlocks if a task never responds. Likely handled by
backend timeouts. This deadlock state is only partial and not lossy as
successfully uploaded files will appear in gallery even if the task is
cancelled. Cancellation can be attempted at any point during the upload.
Uploads shown in progress may be successful after cancellation depending
on the internal upload progress.
- TopBar upload images button now opens the image modal instead for
consistency and same-component use.
---------
Co-authored-by: bflatastic <bombayapirat@gmail.com>
## Summary
- **Fix blank screen crash** during video recreate flow: typing text,
adjusting time slider, typing again, then clicking the text box caused
an uncaught DOM exception in MentionTextarea's contentEditable caret
operations, which unmounted the entire React tree (no error boundaries
existed)
- **Fix text selection disappearing** on mouse release in
MentionTextarea: click handler and sync useEffect were destroying active
selections by rewriting innerHTML and restoring only a collapsed caret
- **Add React Error Boundary** around PromptBoxVideo as a safety net
## Changes
### MentionTextarea.tsx
- Wrap `getCaretOffset`, `setCaretOffset`, `scrollCaretIntoView`,
`handleInput`, sync `useEffect`, and `getOffsetRect` in try/catch to
prevent uncaught DOMExceptions
- Add null guard for `childNodes` iteration and `parentNode` access
- Skip `detectMention` in `handleClick` when user has active text
selection
- Skip innerHTML rewrite in sync useEffect when editor has selected text
### PromptBoxVideo.tsx
- Stabilize `characterNames` with `useMemo` to prevent `mentionColorMap`
from rebuilding every render
- Move empty characters array to module-level constant for stable
reference
- Add cleanup `useEffect` for `durationTimerRef` to prevent stale timer
after unmount
### New: PromptBoxErrorBoundary.tsx
- React Error Boundary with "Try again" recovery button
### ImageToVideo.tsx
- Wrap `<PromptBoxVideo>` with `<PromptBoxErrorBoundary>`
## Reproduction steps (original bug)
1. Click "Recreate" on a video
2. Type text into the prompt box
3. Adjust the time slider
4. Type text again
5. Adjust the time slider and click the text box → **UI goes blank**
## Test plan
- [x] Reproduced blank screen bug on Mac before fix
- [x] Verified fix: repeated recreation flow no longer causes blank
screen
- [x] Verified text selection persists after mouse release
- [x] Video generation with @mentions, reference images, and duration
changes works correctly
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Instead of looping within the image set, pressing Right on the last
image advances to the next set and Left on the first image goes to the
previous set, consistent with Up/Down behavior.
BFlat
Brandon Thomas
Tanish Manku
frankywashereandClaude Opus 4.6