adachi0130-Adachi-Inpaint-Workbench/VERIFICATION.md

Verification record

• 0.5.13: the remote history was read without modifying its workflow or queue. Three actual failed request snapshots contained selector 532 with a selected flag but no data input; local normalization restores its saved tag-subgraph connection to 349:412, leaving selector 533's 16 source intact. Two real local CPU queue probes reproduce the missing-input shape and subgraph mapping with lightweight image sources, verify both selector outputs and upstream changes, and avoid model generation. All 39 backend unit tests pass, including Get/Set recovery, flattened subgraph output slots and refusal to reconnect absent or mismatched source types. The remote's complete Anima generation was not executed by these tests.

• 0.5.12: all 37 CPU unit tests pass. The supplied remote workflow was used as request metadata with lightweight red/blue image probes replacing model generation: positive selector 532 chooses B, negative 533 chooses A, all four Gets remain enabled. Ten real queue cases cover reverse Set/source states, switching back, encoder-only and Set-only bypass, unchanged-workflow preservation, cache reuse, and both-enabled/both-ignored errors. No complete Anima model generation is claimed by these routing probes. Sixteen native KJ Get/Set browser queue cases pass on frontend 1.52.7 in both renderers, and all 15 metadata-free stale-state API queue cases still pass.

• 0.5.11: all 32 backend CPU tests pass, including stale branch flags with either single submitted input, strict empty/ambiguous input errors, both-connected passthrough selection and dynamic subgraph node lookup. Fifteen real CPU queue cases verify single-branch A/B output pixels for flags -1 through 3, both-connected A/B selection, and invalid cases. Sixteen real KJ Get/Set queue cases on frontend 1.52.7 cover both renderers, source/Set/Get ignore and mute states, reverse switching, both-invalid states and workflow reload.

Environment: Windows, local ComfyUI 0.19.3, ComfyUI frontend 1.42.11, PyTorch 2.8.0, Pillow 11.0.0, Chrome 155, using a separate port 8190 CPU service with its own input/output directories and an in-memory database.

• 16 backend CPU tests pass, covering both sampling areas, edge masks, inversion, growth, feathering, surrounding-colour fill, batch mapping, source-pixel protection, empty masks, denoise=0, latent noise/zero modes, dedicated inpaint conditioning, different VAE compression ratios and Refiner staging.

• 12 browser interaction tests pass, covering real mouse painting, right-button erase, rectangle/lasso, undo/redo, import persistence, half-pressure hard brush, active-shape saving, zoom coordinates and 1600/768/390 pixel window layouts.

• Painting on the real canvas, workflow serialization, executed-parameter serialization and reload were verified in both classic nodes and Nodes 2.0: mask pixels match and the workflow correctly enters the modified state, with no plugin JavaScript errors.

• The real ComfyUI API queue ran load → prepare → composite → save-image workflows in both sampling areas, with correct output sizes and RGB pixels outside the effective mask equal value-by-value.

• Both example workflows were generated by the real nodes, reloaded and used to build executed parameters; nodes and links are valid and model filenames are user-selectable placeholders.

• The installer was tested in a temp directory for first install, repeat install and previous-version backup.

• 0.1.1: in classic nodes and Nodes 2.0 the bottom height bar was dragged with a real mouse, growing the canvas by 160 px and shrinking it by 80 px; size and mask both survived save and reopen.

• 0.1.2: real-browser Ctrl+wheel tests covered brush size growth and shrink, the 1–512 px limits, slider value sync and plain-wheel zoom, confirming the four sliders sit below the canvas; 14 browser interaction checks passed in total.

• 0.2.0: added Alt temporary erase and the image comparison node; 18 backend tests pass. Real mouse/keyboard use confirms releasing Alt restores the previous tool.

• 0.2.0: the real API queue in classic nodes and Nodes 2.0 verified left/right pixels, the mouse divider, wheel zoom, Ctrl+wheel brush size, batch result selection, upstream editor discovery with several candidates, click-to-adopt, old mask clearing and reopen after save, with no plugin script errors.

• Picture wheel events in both node modes never trigger whole-workflow zoom by mistake, and adopted results are stored in the input directory and can be embedded in the workflow.

• 0.3.0: in classic nodes and Nodes 2.0 the bottom inline parameters were verified to drive the real sampling node, width/height on separate rows, source-size filling, no repeated growth on workflow reopen, disconnect restoring the original parameters, and the original parameter rows actually hidden in Nodes 2.0.

• 0.3.0: wheel events over the toolbar and parameter area pass through to the outer canvas, a continuous outer wheel gesture does not change image zoom when it enters the picture, and wheel events inside the picture after stopping zoom only the picture.

• 0.3.0: the comparison node was measured with the result on the left and the original on the right, showing the full result on exit; the divider is a single physical pixel, the round handle, corner labels and bottom caption were removed, and adoption plus batch selection still pass.

• 0.3.0: 15 brush browser interaction checks pass, vertical stretching and save/restore pass in both rendering modes, and the two updated example workflows pass link and executed-parameter validation.

• 0.4.0: in classic nodes and Nodes 2.0 the inline canvas was verified to keep only the load button with the original painting DOM controls moved into the WebUI node; real mouse painting can then be undone from the inpaint node, the tools and sliders drive the original canvas and the sampling parameters correctly reach the execution request.

• 0.4.0: canvas margins shrinking to the image ratio, thin height-bar stretching, size retention across save/reopen, source-size filling, disconnect/reconnect, deleting and recreating the inpaint node, and toolbar-area wheel events continuing to drive the outer canvas were all verified.

• 0.4.0: the comparison node no longer has a top toolbar, with result selection, dimensions and fit moved to the bottom; left/right comparison in the real API queue, full result on exit, batch selection and result adoption all pass.

• 0.4.1: a real browser confirms the eraser SVG cursor and erase state after holding Alt, restoring the cross brush cursor on release; temporary erase, regular painting and the 18 backend CPU regression tests pass.

• 0.4.1: the public name was unified to Adachi-Inpaint-Workbench while node class IDs were kept so existing workflows stay compatible.

• 0.4.2: node display names, descriptions, tooltips and frontend labels were converted to English so the nodes are usable for the general Manager audience; the same 18 backend CPU tests pass and every touched script parses.

• 0.5.0: in classic nodes and Nodes 2.0, adopting a real queued comparison result preserves mask PNG pixels, base-mask pixels and the current undo/redo command arrays. Undo/redo still work, workflow reload retains the mask, a mismatched result leaves the source/mask unchanged, and ordinary image upload still clears the mask.

• 0.5.0: group preset tests cover real preset-button clicks, enabled/bypassed modes in executed prompts, exclusion of the virtual controller from backend execution, saved workflow reload, renamed and duplicate-title groups, unmanaged new groups, nested group priority, missing groups and protection of the controller itself.

• 0.5.0: the native right-click preset submenu was clicked successfully in both renderers. The example workflow was loaded and all three presets queued successfully without a model.

• 0.5.1: all 23 CPU unit tests pass, including exclusive input validation and identity-preserving forwarding of conditioning, image/mask tensors, latents and falsey values.

• 0.5.1: 26 real CPU queue cases across classic nodes and Nodes 2.0 cover A/B selection, bypassed transform pass-through, mute, both-enabled/both-bypassed/both-disconnected errors, reroutes, one connected input, group presets, cache changes, workflow reload and switching reloaded presets. Output pixels match the selected image; the unused branch was not requested by the selector. A raw API prompt works without frontend metadata.

• 0.5.1: the 40%-width preset list retains internal scrolling and has no horizontal overflow in a 360px node at a 390px viewport.

• 0.5.2: all 29 CPU tests pass. New cases cover single-frame 5D image latents, per-source/variant masks, both crop modes, noise/zero fill, third-party VAE wrappers, native inpaint conditioning, refiner continuation, and explicit rejection of real multi-frame output.

• 0.5.2: actual ComfyUI VAE + WanVAE encode/decode and prepare_mask were exercised with a small randomly initialized Wan architecture (16 latent channels). Four cases cover whole-image/original, masked-area/noise, zero-latent fill, and native inpaint conditioning, with two independent sources and two variants each. Latents were [4,16,1,8,8], masks [4,1,1,8,8], decoded images [4,64,64,3], and all protected source pixels remained exact.

• Anima checkpoint image generation was not performed: this machine does not have the user's Anima/Qwen VAE checkpoint files. The workflow was inspected read-only, confirming node 501 uses VAE node 358 (qwen_image_vae.safetensors) and the Anima model branch.

• 0.5.3: in classic nodes and Nodes 2.0, the inline Fit button is on the same row immediately to the right of Load image, and is absent from the connected inpaint controls. Actual browser clicks reset zoom/pan and preserve the painted mask PNG exactly.

• 0.5.4: frontend 1.52.7 was downloaded as a wheel and extracted into an isolated test directory, leaving the installed frontend unchanged. With saved node dimensions from the user's workflow, the previous bubbling Fit click changed the classic editor viewport from 396x625 to 396x594. Stopping propagation keeps it at 396x625. Nodes 2.0 also retains its viewport and node dimensions.

• 0.5.4: actual browser clicks pass with the saved workflow geometry on frontend 1.42.11 and 1.52.7 in both renderers. Fit resets zoom/pan, remains immediately after Load image, and the inpaint panel occupies over 94% of its node's logical width. The remote server and its installed extension set were not accessed, so the separate narrow-panel screenshot cannot be claimed reproduced/resolved remotely.

• 0.5.5: injecting a cached widget width equal to 35% of its node width reproduces the narrow-column shape on frontend 1.52.7 classic nodes. Before the fix, a 380px inline node had a 113px editor; afterwards its editor is 360px wide, with the remaining 20px reserved for native margins. Inpaint and comparison widgets recover their full widths as well.

• 0.5.5: browser checks on frontend 1.42.11 and 1.52.7, classic nodes and Nodes 2.0, cover attempted stale width writes, resizing node widths to 800/450/1050px, independent height expansion, graph zoom at 0.3/0.6/1.15, Fit with mask preservation, and saved workflow reload. The simulated failure is covered; the remote server itself was not accessed.

• 0.5.6: inspected the supplied workflow read-only: selectors 532/533 are fed by KJ Get nodes p/p_0 and n/n_0, while the p_0/n_0 setters are bypassed. The previous direct-node check therefore incorrectly classified both branches as enabled.

• 0.5.6: actual local KJNodes Set/Get frontend definitions were loaded only into isolated QA browsers (no KJ code is bundled). Across frontend 1.42.11/1.52.7 and classic/Nodes 2.0, 32 real CPU queue cases verify ignored setters, swapped states, ignored/muted real sources, ignored Get, explicit both/neither errors, and switching after workflow reload. Output image pixels match the selected branch; generating prompts does not change editable connections.

• 0.5.6: dependency-free routing tests cover Get/Set, reroutes, current virtual-output APIs, mute/bypass, missing setters/inputs and cycles. All 29 backend CPU tests continue to pass.

• 0.5.7: actual browser wheel events over a real queued comparison image change workflow navigation while the comparison transform stays fitted and centered. Verified on frontends 1.42.11 and 1.52.7 in classic nodes and Nodes 2.0, with left-result/right-original pixels, hover/leave behavior, batch selection, adoption preserving mask/history, workflow reload, rejected mismatched adoption and ordinary-upload mask reset.

• 0.5.7: three real queued two-image batches accumulate 2, 4 and 6 results and select the newest result. After adoption, old results retain their old original and new results use the adopted original; all six results survive workflow reload in both frontend versions/renderers. Dependency-free tests additionally cover legacy batch mapping, empty outputs and immutable appending.

• 0.5.8: real single-image queues on frontend 1.52.7 in classic nodes and Nodes 2.0 show 1/1 then 2/2. Replaying the same execution output twice does not increase the count or reset a manually selected result. Reloading a workflow containing duplicate preview references removes duplicates. Unit checks distinguish filenames, subfolders and file types.

• 0.5.9: on frontend 1.52.7 in classic nodes and Nodes 2.0, a 1024x1536 comparison was measured at workflow zoom 1x, 2x and 4x. At 4x, 24 synchronous paints with readback dropped from 483 to 36 ms (classic) and 453 to 35 ms (Nodes 2.0). These are isolated browser measurements, not end-to-end workflow FPS guarantees. The divider measured one screen pixel at every zoom. Dirty-strip redraws matched full-redraw pixels exactly across forward/reverse/unchanged divider movements and edge positions; pointer exit hides the divider.

• 0.5.10: frontends 1.42.11 and 1.52.7, both classic nodes and Nodes 2.0, verify the Upload label and pin icon beside Fit, pin persistence across workflow reload, plain wheel navigating the workflow, Ctrl+wheel brush sizing while pinned, and restored image wheel zoom after unpinning. At 4x graph zoom the display allocation remains bounded, the original mask dimensions remain 320x200, and 20 cursor-only redraws never recomposite the mask overlay. All 15 standalone editor regression groups pass, including real painting, Alt/right-button erase, undo/redo, shapes, pressure, serialization and zoom/pan coordinate mapping.

• 0.5.10: actual adoption reduces the comparison list to 1/1, preserves mask PNG and undo/redo identities, and saves the pending new-round state. Reloading before generating retains that state; the next two-image batches start at 2/2 and then accumulate to 4/4. Rejected adoption leaves history unchanged. These checks pass in both frontend versions/renderers.

The sampling path was also checked for noise-schedule continuity against the real common_ksampler/KSampler functions in the local ComfyUI; an actual checkpoint has not been loaded to test GPU image quality or every model family, so CPU/interface tests must not be treated as model-quality verification.

References used during implementation:

• ComfyUI frontend extension APIs
• ComfyUI widget serialization notes
• WebUI inpainting pre-processing semantics
• rgthree image comparison interaction reference

The plugin code is an independent implementation; see LICENSE for the release licence. No source from the projects above is bundled.