Carry artist intent through generated assets - #252
Open
OmarB97 wants to merge 21 commits into
Open
Conversation
Rigged models (from the rigging extensions, or any GLB with clips) render frozen in bind pose today: the viewer loads the glTF but never creates an AnimationMixer, so the clips it just parsed never run. That reads as a failed rig rather than a missing feature. Create a mixer for models that carry clips, tick it from useFrame, and dispose it with the model. When a file carries more than one clip, offer a small picker so each motion can be reviewed without leaving the app. Co-Authored-By: Claude Fable 5 <[email protected]>
Asset rows now render their pre-rendered <name>.thumb.png beside the filename instead of showing bare text, falling back to the existing text-only row when no thumbnail exists. The Library panel can also be resized by dragging its edge, with the chosen width persisted across restarts.
…' into feat/desktop-ergonomics
npm ci fails on main because the committed lockfile is stale against package.json; npm install (what upstream CI itself falls back to) brings it back in sync. Committing the result so ci is usable again on this branch. Co-Authored-By: Claude Fable 5 <[email protected]>
Ctrl/Cmd with +, -, or 0 now scales the whole window like a browser, clamped to [-2, 4] and persisted to <userData>/ui-zoom.json so it survives restarts. This extends the existing before-input-event handler (where the macOS quit shortcut already lives) and calls webContents.setZoomLevel directly, deliberately not CSS zoom: a CSS transform scales the rendered pixels without telling the 3D viewport, so pointer math there (raycasting, gizmo dragging) goes wrong as soon as the page is scaled. Co-Authored-By: Claude Fable 5 <[email protected]>
The panel's resize handle was a 1px strip with no resting affordance — easy to miss entirely. Widen it to 2px with a visible resting tint plus a stronger hover/active highlight, mark it up as a proper separator (role, aria-orientation, title) matching the Library panel's handle, raise the width cap from 520 to 900 so wide parameter forms have room, and persist the chosen width to localStorage (modly-panel-width) the same way the Library panel already remembers its own width. Co-Authored-By: Claude Fable 5 <[email protected]>
- Parameter labels used a truncating 5rem column that clipped anything longer than a couple of words; widen it to 8rem and let it wrap. - The prompt textarea was a fixed 3-row, non-resizable box; make it 14 rows by default and user-resizable (resize-y, 14rem min-height). - Say plainly which Generate inputs are required vs optional: "Your sketch or photo -- required" and "Extra details -- optional". - Add a text param type (a plain input, unlike the existing string type folder picker) so free-text fields like the exporter new name, project, and tags params render as usable text boxes instead of falling through to a numeric input. - Render clickable tag-suggestion pills beneath any tags param: slugified words from that node typed name/project fields, plus a small fixed game-asset vocabulary, deduped and capped at 10. Clicking a pill toggles it in the comma-separated value; no network calls. Co-Authored-By: Claude Fable 5 <[email protected]>
The clip picker was a raw select mounted via drei Html at the model origin, recreating the whole AnimationMixer every time the clip changed (a hard cut, not a blend) and easy to lose behind the scene. Replace it with a proper bottom-left overlay component (MotionBar), matching the existing ViewerToolbar/stats-overlay idiom: - The mixer and active AnimationAction now live for as long as the model does; switching clips fades the old action out and the new one in over 0.25s instead of recreating anything. - GltfMeshModel reports its clip list up to Viewer3D (the same onStats/onObject callback pattern already used here), which owns the active index and renders MotionBar as a plain sibling of the Canvas. - Zero clips renders nothing, exactly one clip shows its name read-only, more than one gets a labelled dropdown plus a Next button that cycles with a cross-fade. All playback loops. Co-Authored-By: Claude Fable 5 <[email protected]>
The built-in exporter named every file export-<epoch-ms>.glb, so a folder of work told you nothing. Add three optional params: model_name (used to slug the file name, falling back to a dated model-YYYY-MM-DD-HHMM name), project (files land under Exports/<project-slug>/ so a game groups together instead of one long list), and tags. Output is always <slug>-<4-char-unique>.<ext> so re-exports never collide. Tags are the operator typed list plus what the file itself can tell: rigged (has skins), animated plus clip-<name> per animation, textured, and one of low-poly/mid-poly/high-poly by triangle count, plus the project slug. Tags are written both as a <file>.tags.json sidecar and embedded under the glTF asset extras (extras.modly), so they travel with the file either way. Ported from the already-working implementation at /mnt/tank/modly/app/extracted/resources/builtin-extensions/mesh-exporter on the operator machine. Co-Authored-By: Claude Fable 5 <[email protected]>
The agent had no idea what was actually on screen beyond a mesh path and triangle count -- it couldn't tell you which clip was playing, what the model was named, or what game it belonged to. Extend both ends of that pipeline: - ChatPanel now mirrors the viewer's selected clip and full clip list (Viewer3D reports them into appStore alongside the existing meshStats/meshSelected pattern) into the /agent/chat context, so requests like "make the stroll slower" resolve without re-asking. - agent.py resolves the current mesh's name/project/tags from its .tags.json sidecar, falling back to the GLB's own embedded extras.modly when no sidecar exists, and renders them into the system prompt in the same style as the existing scene-context lines. Lineage: when mesh-exporter's input traces back to an existing workspace asset, record derived_from in the sidecar -- the source's workspace- relative path plus its name if it had one. The workflow runner threads a sourceAssetPath through NodeOutput/ProcessInput from wherever a mesh first resolves (the Load 3D Mesh node) down to whichever node eventually exports, so multi-step chains still carry it. Re-deriving from an already -derived asset keeps just the immediate parent plus the chain's root (read from the parent's own derived_from, if any) instead of growing an unbounded list. Surfaced in the chat context alongside name/project/tags. Also fixes an open question: the rig-and-animate extension's "Describe the movement" param sets multiline: true, a flag ParamField never checked -- it fell into the plain "string" case and rendered as a single -line input with a stray folder-browse button. ParamField now renders any multiline-flagged param as a resizable textarea (matching the existing "Extra details" box), independent of its base type. Co-Authored-By: Claude Fable 5 <[email protected]>
Search now force-expands every section that contains a match instead of requiring a manual Show click per group, and clearing the query falls straight back to whatever was collapsed before (collapsedSectionKeys is never mutated by the search itself). The listing area now flexes to fill a resizable popup instead of being capped at max-h-64 (roughly 3-4 rows) - a new bottom-edge drag handle mirrors the existing width handle, with its own clamp/persist pair so the height survives restarts the same way the width already does. Also fixes the search input rendering as an oversized box with the placeholder vertically centered in the middle of it - a known Chromium input[type=search] quirk, fixed with appearance-none like the rest of the app's inputs effectively already assume.
The panel width floor, default, and ceiling were all far too small (260/320/560) for a panel that is now full-height - a thumbnail, a readable name, its full workspace path, and its type badge could not share a row without truncating. Raised to 380/640/960 and, since the floor moved too, a panel already dragged down toward the old 280px-ish minimum on a prior run now clamps back up to something usable on the next launch instead of silently staying cramped. Selecting a Library row and then separately clicking "Open selected asset" is gone as a required two-step. A row click now selects and opens in one motion via a shared openLibraryEntry() helper; the button stays as a secondary affordance (useful to retry after an error) but is no longer the only path. Non-openable entries still just select and show their reason, same as before. Hover previews: on hover, a row with a rendered preview manifest swaps its static thumbnail for a looping WebP clip, with a small mini-select control to cycle clips when more than one exists - lazily fetched and cached per row, never fetched for rows with no manifest. This reuses the existing workspace:library:thumbnail IPC channel rather than adding a parallel one: an optional previewClip request field fetches a named clip's WebP instead of the static thumbnail, and the default (no previewClip) response now also carries the manifest's clip list when one exists. The manifest's file field is trusted only as a filename (basename-only) and re-resolved through the same normalizeWorkspaceAssetPath guard every other workspace read uses, so a malformed or stale manifest can never point outside the asset's own directory. No preview manifests exist yet (that pipeline is separate, in flight elsewhere) so today this is inert: previews stays empty and every row behaves exactly as it did before this change.
Text typed into "Extra details" was assigned straight onto the node's params and spread after them, so it overwrote the selected style's prompt entirely. Choosing a Creature style and then asking to keep one tentacle short deleted every word about being a photographic creature; the generator received a cropped notebook drawing and faithfully rebuilt it as a flat card. The flatness was blamed on decoder resolution and mesh budgets for a long time. Neither was involved: the failed mesh measured 0.030 flatness with 29,745 faces, cascade resolutions reached only 0.068, and a properly styled input at the same settings and seed reached 0.369. mergeNodeText now combines the two, and keeps the fuller text when one already contains the other — collapsing to the shorter note would throw the style away again, which a test caught before this shipped. Co-Authored-By: Claude Fable 5 <[email protected]>
Thin features like blades, wires and tentacles come out shredded at the normal setting, and the detector for them was silently raising the setting on the artist's behalf. It is not reliable enough for that: it fired at 0.99 on a squid sketch and gave opposite verdicts on two crops of the same lamp. It now offers the choice, says why, and states what it costs — about a minute instead of about twenty-five seconds — and stays quiet when it is unsure. It never blocks a run. Co-Authored-By: Claude Fable 5 <[email protected]>
Answering the rigger's clarification question re-ran the workflow from the top. The style step redraws rather than edits, so the artist answered a question about a model they would never see again — and the rigger had already worked out the answer before it stopped. The machinery to resume from a single step was already there and unused. Wording is now about the step at hand rather than branches, which is not vocabulary a game designer should have to learn. Co-Authored-By: Claude Fable 5 <[email protected]>
Ctrl/Cmd +/- and Settings' "Interface Scale" were separate systems for the same thing, each with its own stored value and neither aware of the other. The main process restored the saved zoom level when the window loaded, then the renderer mounted and overwrote it with the scale preset — so the keyboard setting was lost on every restart and had to be redone by hand each launch. There is now a single saved factor. The main process still catches the chord, because Chromium claims it before the page can see it, but it only forwards the direction. Settings shows the real percentage rather than a preset switch that would have to lie whenever the keyboard landed between presets. Co-Authored-By: Claude Fable 5 <[email protected]>
Painting a texture in an external editor now updates the model in the preview without a rebuild. Measured at about 63ms from save to visible change, with nearest-pixel sampling preserved through every reload. A malformed or wrong-sized save keeps the previous image and shows a plain explanation rather than silently going stale, because a stale texture that looks fine is the worst outcome here. PSD was investigated and left out: Electron decodes only PNG and JPEG, and partial PSD support risks confidently showing the wrong image.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
modly.intent.v1record before the first Style/model stagederived_fromlineage while retaining the full intent historyThis is stacked on #251 because it branches from the newest installed desktop build.
Verification
npm run buildnpm run lint -- --quietnpm test(23 Python tests and 94 Node tests)npm run package -- --linux dirneeds_clarification, and surfaced the exact question in the desktop UI without exporting or guessing