feat(ui): Tons of workbench refinements and tweaks

[joshistoast]

Summary

The workbench ran on Chakra's marketing-sized scales: md controls were 36–40px and body text was 16px. As a result, nearly every call site opted down by hand with size="xs"/"2xs" and fontSize="2xs"/"xs". This PR rebases the scales, polishes a set of workbench controls, makes dialogs and menus animate out instead of vanishing, and tidies the Model Manager's starter bundles.

Size scale

• md is now the working size: 28px controls and 12px text. One control scale runs xs–3xl in 4px steps (20–44px), and the type scale extends to 9xl.
• Stock Chakra recipes and explicit size/typography call sites are renamed onto the new scale, preserving their geometry during the rename.
• Theme types are generated with chakra typegen, which makes retired recipe sizes such as size="2xs" a type error. A separate architecture check rejects retired fontSize/textStyle="2xs" literals, which Chakra's permissive CSS types cannot catch.
• Default controls use the working scale and body text defaults to md; redundant size and font-size props are removed. The intentional default-size pixel changes are listed under Compatibility / Rollout.
• Hand-built control heights read the shared scale, including controls integrated with the floating-window and center-recovery UI.

Workbench polish

• Generate:
  • Right-clicking from Steps to CFG keeps exactly one context menu open. Switching between a concept row, its weight scrubber, and a settings field also replaces the previous menu reliably.
  • Dividers separate the Krea-2 conditioning rebalance, seed, and seed variance fields; the rebalance divider has even spacing.
  • The seed variance switch sits at the end of its label row.
  • The guidance badge shows active out of total.
• Concepts in Generate, Upscale, and Video:
  • All three use the same concept (LoRA) row, with a tighter snug list density, weight scrubber, toggle, remove action, and context menu. A permitted action opens the model in Model Manager.
  • The picker says when every compatible concept is already added. Upscale labels its LoRA picker as Concepts, and unsupported concepts remain removable.
  • Weight edits stay local for 250 ms. Commits, additions, and menu actions use current settings, so adding, removing, or toggling one concept cannot revert another concept's pending weight or an intervening settings update. Removing a row mid-edit writes nothing beyond the removal. Video keeps its accelerator-default reconciliation and notices.
  • Enter/Space opens row actions, Escape restores focus, and removing a row moves focus to its neighbour, or to the concept picker after the last row. Project changes dismiss menus without flushing retained drafts into the new project. Weight scrubbing avoids rerendering sibling rows on every pointer move.
• Canvas: "Scale before processing" defaults to Auto. An explicitly saved choice, including None, is kept; existing projects with no saved method also take the new default.
• Layers:
  • The swatch shelves have dividers between them.
  • The properties pane title is gone; a running operation names itself in its own section header.
• Upscale: dividers separate the setting groups, and the redundant bundled-VAE note is removed.
• Appearance: a new "Prompt text size" preference (small / default / large / larger) applies to the prompt fields, with highlighting aligned to the chosen size.
• Launchpad: What's New opens on the first click; before, the first open dismissed itself under StrictMode. The connection state uses a Badge.
• Image Map:
  • The missing-encoder link opens Add Models searched for that encoder instead of downloading it. Reopening the map asks the server again while the encoder is still missing.
  • An open map detects encoder removal during a server session and shows "Embedding model not installed" with the Add Models link instead of only a stall message. It recovers after reinstall.
  • Encoder replacement no longer depends on the map being open. Once in-flight work drains, indexing and gallery semantic search resume on the replacement and the retired encoder's resources are released. Old labels are invalidated even when the browser missed the removal.
  • While a replacement drains, the map shows Computing rather than the install prompt, gallery search stays offered, and search explains that the model is switching.
  • The encoder is found by its configured name: description or cover edits keep indexing running, while deleting or renaming it makes it missing. The Image Map guide no longer says a restart is needed.
  • Status polling pauses in hidden tabs, and a repeated missing-encoder status no longer re-renders the map or cancels Check again.
• Shared controls and chrome:
  • Segment tabs hide the dividers beside a hovered tab. Interactive list rows do the same, retain their highlight while a context menu is open, and support detail content beneath the primary row. Static rows keep their static appearance.
  • List-row spacing is tightened and pointed rows use the accent-tinted bg.hover token. Queue lists gain dividers aligned with their tabs.
  • Scrubber fills extend past the thumb with a rounded end, then sit inset with the shared hover tint.
  • The splash screen uses the preview dot grid; the grid styling lives in the theme. Toasts wrap long model names, and badge styling is polished.
  • The Invoke button keeps rounded corners within its attached cluster, with neighboring borders running beneath it; the routing tail carries hover/active fills. Unknown progress uses the shared Spinner, and cluster styles stay with their components.
  • The project switcher uses a down chevron. Combobox default search and empty-result text is generic instead of naming schedulers.
  • Hover and highlight fills lean further toward the theme accent. Menu, select, combobox, picker, command palette, Add Node, prompt autocomplete, and preset/template rows highlight with the same translucent bg.hover as ghost and outline buttons, so an open trigger and its highlighted option match. Inactive layout presets, widget rail items, and status-bar rows hover with it too; selected fills are unchanged.
  • A list row with only a context menu opens it from keyboard, screen-reader, touch, and pen activation; a mouse click does not (right-click does). Rows whose only controls are action buttons get the pointed hover.
• Dialogs and menus animate out: about 25 dialogs unmounted the moment they closed, so they vanished instead of playing their exit animation. Hosts now keep them mounted until the exit completes (useExitRetainedValue / useExitPresence), and Confirm and Rename dialogs keep their last text while closing instead of falling back to empty names or another variant. Two layer dialogs never opened at all and now do: the group Rename and a layer's Run workflow. Modal hotkey layers release at close rather than after the animation.
• Model Manager:
  • Starter, Hugging Face, and folder-scan rows in Add Models share one right-click menu per list offering the row's next step: Install, View queue, or View model. Rows stay out of the tab order beside their Install button.
  • Starter bundles are renamed SD 1.5, FLUX.1, FLUX.2, Z-Image, Krea 2, and LTX 2.5. The Wan 2.2 text-to-video and image-to-video bundles are combined. FLUX.2 adds Klein 9B, the PiD 2K decoder, and FLUX.2 [dev] (GGUF Q4_K_M) with their encoders.

QA Instructions

Webv2 checks used Node 22.14 and pnpm 10. Run bare pnpm commands from invokeai/frontend/webv2; commands with -C and Python commands run from the repository root:

• pnpm lint: passed (formatting, oxlint, theme typegen, TypeScript, architecture).
• pnpm test: 9,509 passed.
• pnpm test:browser: latest full run 2,208/2,208 passed. Earlier runs intermittently failed the SINGLE/COLLECTION InvocationFlowNode export pixel assertions, which fail the same way on base b05ecdc; their expectations are unchanged.
• pnpm test:fixtures: 28 passed.
• pnpm test:performance:architecture: passed after re-recording the baselines. Request counts are unchanged on every route (minimal editor startup remains 89 requests / 84 scripts); the exit-retention hook is grouped with ConfirmDialog so it adds no startup request. Editor startup-owned JavaScript is 148,443 B: the dialog exit fix adds about 600 B, and the 2026-10-02 baseline predated about 4 KB of earlier growth on this branch.
• pnpm test:project-files:browser: passed. pnpm test:accessibility:browser: all 18 journeys passed against the fresh build.
• Fixtures, performance, project-file, and accessibility results above were rerun on the final commit.
• pnpm check:release stops at the known workflow-export browser failure; the remaining release stages listed above passed separately.
• Shared API generation, pnpm -C invokeai/frontend/api lint, and its 2 generator tests passed. Legacy pnpm -C invokeai/frontend/webv1 lint:tsc reports 16 existing Qwen 3.5 model-union errors, reproduced with identical diagnostics against b02f548.
• OMP_NUM_THREADS=2 uv run --no-sync pytest tests/app/services/image_index tests/app/routers/test_image_map.py -n 4: 322 passed. Starter model and bundle tests (tests/backend/model_manager/test_starter_models*.py, router starter tests): 23 passed. Ruff 0.11.2 check and format check passed for all changed Python files.

Size scale: the original Playwright geometry/typography sweep covered about 28 states against the mock backend. The rename commit was pixel-identical. Later default changes include the visible Badge, Switch, Checkbox, and ColorPicker changes documented below; the rename result does not establish that those defaults are pixel-identical.

Polish items: checked in the browser against the mock backend. Screenshots are not attached; local captures cover the concept menus, preserved weight, and missing-encoder prompt. To try the flows:

• Generate: with a Krea-2 model, open Render. Right-click Steps, then CFG; exactly one menu stays open. Check the conditioning, seed, and seed variance dividers.
• Concepts: adjust A's weight, immediately open B's context menu, wait for the draft to settle, then remove B; A keeps its new weight. Switch between row, weight, and settings-field menus. Use Enter/Space to open row actions and Escape to restore focus. Switch projects with a pending draft and open menu; neither leaks into the new project.
• Video/Upscale: scrub a concept weight and remove or toggle another concept; other committed settings remain intact. The integration tests also flush two drafts in one turn and verify Video accelerator defaults/notices. The shared-row regression measures ten pointer moves with one eventual settings commit and no sibling-row renders during those moves.
• Preferences → Appearance: set "Prompt text size" to Larger; the prompt becomes 17px and its highlighting stays aligned.
• Image Map: with no encoder installed, click the model name; Add Models opens searched for it and nothing installs. Availability/recovery tests cover an encoder removed during the session, reinstall, replacement while the browser missed the missing state, replacement detected by the worker with no map request, and shutdown while a replacement drains.
• Launchpad: What's New opens with one click.
• Dialogs: close the layout preset save-as/edit dialogs, What's New, the command palette, Add Node, a model's delete confirm, or a layer's Rename; each animates out with its text intact. The browser tests record each dialog reaching its closed state before it unmounts.
• Add Models: right-click a starter, Hugging Face, or scan row; the menu offers Install, then View queue while installing and View model once installed. Tab from an Install button skips the rows.

Review

Resolved findings include stale concept-menu callbacks and captured-array updates, the row/field right-click race, inert keyboard stops, static-row hover fill, missing pending-draft project-switch coverage, per-move Video/Upscale writes, and Image Map availability/label recovery after encoder deletion or replacement. A later pass fixed indexing and gallery search staying down after a replacement until the map was opened, a draining replacement reading as a missing model in the map and as disabled in the gallery, map churn from repeated missing-encoder polls, a captured list in Generate concept adds, focus dropping to the page after removing the last concept, and no-op Upscale/Video writes from removed rows. The earlier Add Models reopening and rebalance-spacing fixes remain included. Accessibility setup now selects the Uncategorized board row explicitly, avoiding the disclosure displaying the same board name.

A later round fixed dialogs that unmounted on close (and two that never opened), Confirm/Rename text changing during exit, hotkey layers held through the exit, stale forms when reopening during an exit, layer-menu exit races, keyboard-only detection that missed screen-reader and touch activation, and per-row install menus that doubled tab stops and subscriptions.

The main-branch concept conflicts are resolved, retaining unsupported-model messaging/removal and remapping the incoming retired tokens. The typography guard covers literal JSX/style/recipe values, including conditionals and responsive values. It does not evaluate variables/calls; still-valid names such as sm require semantic review when merging older code.

Remaining risks and verification limits:

• Shared concept rows: Upscale and Video also receive the tighter density.
• What's New: the fix stops the modal closing on focus-outside in every build, although the original cause was dev-only. Escape, the backdrop, and the close button still dismiss it.
• Checks: the existing workflow-export pixel failures prevent a fully green combined release command; legacy TypeScript has the unchanged baseline errors described above.
• Encoder lifecycle: tests use lightweight fake encoders and real temporary records. Deletion/reinstallation of an actual GPU encoder was not exercised. A replacement whose hash differs still deletes the old embeddings under the activation lock, as before this PR, now without a map request.
• Browser verification: the later fixes were verified through Chromium browser tests, not an interactive session.
• Dialog retention: a dialog closed before its lazy chunk renders open never reports an exit, so it stays mounted (closed and inert) until the next open; the hook documents this bound. Focus after deleting a model or uninstalling a node pack still falls to the page, as before.

Compatibility / Rollout

• Canvas scaling is intentionally changed for existing projects: an unset method now means Auto. Older projects that kept the former default None generally have no saved choice, so they also switch to Auto. With a sub-optimal frame (for example 512×512 on SDXL), processing dimensions can grow; output for the same seed, generation time, and VRAM use can change. No migration pins these projects to None. Select None to process at the frame's exact size; an explicitly saved None remains None. The Canvas guide documents this behavior.
• Default-size pixel changes are intentional: unsized Badges shrink from 20px to 16px tall, with unsized text from 12px to 10px; the affected Switch defaults shrink from 40×20 to 32×16, and Checkbox from 20px to 16px. ColorPicker triggers/inputs shrink from 32px to 28px and swatches from 20px to 18px, so swatch grids reflow. The recipe rename alone preserves sizes; changing defaults does not.
• New preference: promptFontSize is optional and normalized; a missing or invalid value falls back to the default size.
• Theme types: chakra typegen runs from prepare and before tsc, writing into node_modules, so it needs pnpm install to have run. @chakra-ui/cli is a new devDependency. The typography policy also runs in architecture:check and lint.
• Encoder API: /image_map/points and /image_map/status add an optional, nullable model_id fingerprint. While an installed replacement drains they report computing with no model_name, and /search and /search_by_image return 409 explaining the switch; search requests now also reconcile the encoder. Generated OpenAPI and TypeScript contracts are updated; older clients can ignore the additive fields. No database or project migration is needed.
• Starter bundles: starter_bundles now returns one wan key instead of wan_t2v and wan_i2v. Both frontends render whatever the response lists, so only scripts reading those keys are affected. The FLUX.2 bundle grows from about 11 GB to about 70 GB, and FLUX.2 [dev] is non-commercial.
• Performance baselines: re-recorded for the reviewed startup growth above; request budgets are unchanged.
• Locale keys (en only; no other locale had them):
  • Removed: widgets.properties.title.*.
  • Renamed: common.searchSchedulers / common.noSchedulersFound → common.searchOptions / common.noOptionsFound.
  • Added: settings.catalog.promptFontSize and its option labels.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[lstein]

Thanks for the thorough follow-up. I re-reviewed at 055c439 and everything from the previous round is resolved or disclosed:

• Retired typography tokens: fixed. The merge with main is re-mapped, and the new architecture check rejects fontSize/textStyle="2xs", including object, responsive and css={} forms.
• Stale concept actions: fixed. Add, remove, toggle and weight now apply to the latest settings in Generate, Video and Upscale.
• Canvas scale default: disclosed in Compatibility / Rollout, which is what I was after.
• Image Map encoder removed mid-session: fixed on both sides. The server swaps away from the removed encoder and evicts it from the model cache. The client polls status and shows the missing-encoder prompt, or Computing while a replacement drains.
• Smaller items: the cross-menu right-click race, the inert row button and the project-switch test gap are fixed, and the default-size pixel changes are documented.

Checks at this head:

• pnpm lint: clean.
• pnpm test: 9510/9510 passed.
• pnpm test:browser: 2206/2209 passed.
  • The two InvocationFlowNode export failures also fail on main.
  • SettingsDialogTarget failed once under full-suite load and passed 3/3 on its own.
• Ruff check and format: clean.
• Pytest for the image index, image map router, model cache and starter models: 789 passed.

Approving. One thing I'd ask you to address, either here or in a quick follow-up:

FLUX.2 starter bundle now pulls gated, non-commercial models (91db6b8)

The "FLUX.2" bundle (starter_models/__init__.py:588-599) used to be the Apache-2.0, ungated Klein 4B set. It now also installs:

• Klein 9B, plus its Qwen3 8B encoder from the gated black-forest-labs/FLUX.2-klein-9B repo;
• the PiD decoder;
• FLUX.2 [dev] Q4 with cow-mistral.

That comes to roughly 74 GB.

The new comment says "[dev] is non-commercial; the Klein models are not", but the repo's own flux-2.mdx:34 lists Klein 9B under the FLUX Non-Commercial License and as gated on Black Forest Labs' repositories. A user without a Hugging Face token, or who hasn't accepted the 9B license, who clicks the bundle gets a failed ~16 GB encoder install and a 9B transformer that can't run.

I'd suggest keeping the default FLUX.2 bundle to the ungated, commercially usable Klein 4B set. The 9B and [dev] models could be individually installable, or a separately labelled bundle that says it is gated and non-commercial. Either way, please correct the comment.

Non-blocking notes, for a follow-up if useful

• Encoder swap holds the activation lock across a model load. _swap_model holds _activation_lock while waiting on MODEL_LOAD_LOCK.write. During a swap, map polls, searches and stop() therefore wait out any in-flight generation model load, so the 10 s stop() bound doesn't hold. This is documented and can't deadlock.
• The drain test doesn't exercise the drain. test_encoder_replacement_waits_for_inflight_embedding_and_discards_retired_batch[*-query-*] still passes with the drain loop in _finish_retirement deleted: the query is released before the worker's next pass. Its _model_config assertion runs on a background thread, so a failure there would only surface as a pytest warning. The request-side _model_users check in try_activate has no coverage.
• A brief false "not installed" during a swap. _replacement_pending is cleared before the new _model_id is published, so a status read in that window reports model_missing for one poll.
• Unconditional status polling. The map polls /image_map/status every 5 s while visible, even with a healthy active model, and each poll hashes every accessible item. Gating the poll on the missing or draining states would avoid that.
• Stale hover labels outside the map. Gallery and preview hover labels are only invalidated when the map itself reloads, so they can show the old encoder's labels until the map is reopened.
• Docs still use the old bundle names and contents. The Wan, FLUX and LTX starter-bundle docs (wan-2.2.mdx, video-workflows.mdx, the Video Generation introduction, flux-1.mdx, flux-2.mdx, z-image.mdx, stable-diffusion-1-2.mdx, ltx-2.mdx) predate the renames and the merged Wan bundle.