Foundry Preview 24 - Single Image model fix

Preview 24 fixes a first-run error in Single Image or Text. The wizard asked for an FP16 MoGe 2 checkpoint, while Foundry prepared a different MoGe file. Foundry now verifies and downloads the exact checkpoint the wizard uses.

Thanks to @knomadaero for sharing the full error log. The report helped track down the mismatch, and there is a thank-you in Foundry's About tab.

Pinokio users can use Update. Both translated READMEs remain in the release ZIP, in the location accepted by earlier updaters. The canonical wizard was tested through geometry and Hi3D refinement in Comfy Desktop.

Changelog

Foundry preview.23 — Latin American Spanish

Latin American Spanish

Choose Español (Latinoamérica) in the Foundry header or Settings. This preview includes translated wizards, help, accessibility labels, errors, all 12 CCORE examples, and exported package documentation.

The shared localization fixes also cover Single Image progress labels, quad mesh badges, and tab loading text. The language manifest now supplies both the picker and the release README inventory.

Spanish was verified in 15 live Foundry views at default and 200% text. Switching languages preserved edited text and the rendered preview. Catalog and archive checks passed.

Spanish README · Changelog

Foundry preview.22 — English and Simplified Chinese

English and Simplified Chinese / English 与简体中文

Foundry now offers English and 简体中文 across its guided workflows, controls, help, status messages and new package READMEs. Switch languages in place while preserving your inputs and loaded previews. The header has a globe beside the language picker, with local Chinese font fallbacks.

• All twelve CCORE examples include localized names and complete editable descriptions.
• Both CCORE wizards explain the author's English-prompt guidance and offer Translate to English, separate from creative enhancement. Review or restore the text before generating.
• Choose Q4 for lower translation VRAM use or Q8 for higher numerical precision. Translation requires full GPU placement, keeps your original prompt when it cannot fit, and offers clear smaller-model guidance. A shared default-policy hook prepares for future memory profiles.
• The rewritten English README and new Chinese README explain setup, workflows, exports, community contributions, the MCP and companion skill.
• App/skill compatibility metadata helps agents check current capabilities and updates. If you copied the skill into your agent's own directory, refresh that copy too.

现在可以使用简体中文界面了。 切换语言无需重新加载工作流，已输入的内容和预览会保留。CCORE 示例也会填入中文描述；生成前可用本地助手翻译成英文，检查后再继续。

Validation: 168 Foundry backend tests, 34 browser-module tests, five CCORE description-contract tests and seven release/installer tests passed. Live Comfy Desktop checks covered language switching, preserved previews and inputs, 200% text, and actual Q4/Q8 translation. Existing user workflows and assets were preserved.

English README · 简体中文说明 · Changelog

Foundry Preview 21

• Share the same CCORE presets, official HairDeck examples, Gallery handoffs and model sort preferences between the GUI and MCP.
• Let agents retrieve saved image previews as pixels, look up Image Creator's installed models and LoRAs, inspect a named mesh stage and open sections in the selected Foundry window.
• Expose the garment editor's supported operations and controls, and reject a stale Gallery revision before handing it off.

Mumblebox Foundry v0.1.0-preview.20 — Agent tools and quad solvers

• Expand the Foundry MCP to the mesh, 3D-Guided, Splat & Solid, HairDeck, both CCORE modes, Garment Particles and image workflows. Discover live options and advanced settings, attach saved runs to the app, inspect ComfyUI graphs, and retrieve previews and exports through the same services as the wizards.
• Bundle the Mumblebox Assets companion skill. It helps coding agents choose a Foundry workflow, run local generation while other work continues, then inspect and use the exported asset in Blender.
• Choose QuadriFlow, Instant Meshes, or CyberRemesher CUDA in Single Image and Turnaround. Each solver has advanced controls. Run receipts identify the requested and effective solver and any repair or recovery.
• Show the names of the models that are currently resident, including models offloaded to system memory. The indicator updates as workflows load and release them.
• Add experimental Garment Particles controls for users who have configured its isolated runtime and author weights. Add optional Idol Turbo and SMD-TRELLIS.2 choices to their existing workflows.
• Add Tripo, Meshy and expanded Hi3D integration to the roadmap.

Mumblebox Foundry v0.1.0-preview.19 — CCORE Image + Text

• Add a separate CCORE Image + Text wizard. Drop in one reference or make one in Foundry, then describe what to keep or change. Preview the prop in 3D and save GLB or ZIP with model and projector credits.
• Enable dedicated GPU memory checks in both CCORE wizards. Choose Standard (Q4_K_M) or Higher precision (Q8_0). If the selected budget will not fit, Foundry offers measured retry choices and keeps your inputs. Image + Text also accounts for the author’s BF16 vision projector.
• Show elapsed time during processing, including downloads and multiple stages of the same run. The timer continues when you reopen an active run.
• Remember Installed models sorting by name, date added or size. Open the primary ComfyUI models directory from Settings.
• Add matching Image + Text artwork, align workflow cards in each row, move Create Images after the CCORE cards, and update the About and Roadmap tabs. Rename the disabled Donate control to Coffee.

Mumblebox Foundry v0.1.0-preview.18 — Create a Prop from Text

Foundry now includes Create a Prop from Text, a guided workflow for making stylized props locally with CCORE GLB-3D-Coder.

• Open Choose a Workflow → Open Prop Wizard. Start with one of 12 editable examples or write your own brief. Use Enhance description to develop it with the local Qwen assistant, then edit the result before generating.
• Inspect the finished prop in the rotatable 3D preview, download the GLB or a ZIP with file details and credits, or reopen the result from Foundry History.
• The Windows installer prepares the isolated Chromium exporter. Foundry downloads model weights on first use.

The installed standalone ComfyUI Desktop wizard loaded all 12 examples, completed two live enhancements, and rendered a 3,224-triangle crate from an edited brief. CCORE can sometimes generate invalid scene code. The wizard reports the failed export and keeps the description available for another seed or a simpler brief. See the bundled release validation for test scope.

Use Pinokio’s Update control after finishing your work, or follow INSTALL.md for standalone ComfyUI.

Mumblebox Foundry v0.1.0-preview.17 — Single Image and grounded exports

Foundry now has Single Image or Text, the most direct mesh workflow: supply one image—or create one from text inside the wizard—then use the official ComfyUI Pixal3D single-image recipe, Hi3D refinement, cleanup, final TRELLIS.2 texturing/baking and your selected exports.

• Per-object floor placement: automatically calculate a near-upright support pose, place on the floor without rotating, or keep the original orientation. The rigid correction reaches the actual exported files and final preview. Changing placement reuses saved AI stages. It does not flatten geometry or make uneven generated feet perfectly coplanar.
• A clearer toolbox: matching artwork for all four mesh methods, compact layouts and help, consistent icons, larger preview space and expandable technical errors. Turnaround, 3D-Guided, HairDeck and Splat & Solid retain independent workflows.
• Accessible text: a glasses-icon Font menu with Default, 112%, 125%, 150%, 175% and 200% sizes.
• Maintenance controls: release-status flag and safe cache review/confirmation with selected categories and active-job guards. Saved models, references, assets and settings are protected.
• Optional quads: explicit CPU guidance and a reported topology-preserving fallback when QuadriFlow opens an already closed surface. Textured triangle exports still work with quads off.
• Installer/release fixes: declare and probe support-pose dependencies; include WebP method artwork; reject partially validated Comfy jobs; preserve Single Image drafts and Resume; correct selected-output documentation and model credits.

The Windows / RTX 4080 SUPER 16 GB / 64 GB RAM reference system completed a real Single Image asset and a packaging-only grounding repair. Detailed/compact/quad runtime GLBs and editable BLEND were checked, with native previews and reload/Resume. Targeted tests and software-archive checks passed. Other export/settings combinations and a fresh installation of this release were not exhaustively tested. The installed Pinokio copy was deliberately not updated during this release; see the bundled validation records.

Use Pinokio’s existing Update action after finishing/stopping work, or follow INSTALL.md for standalone ComfyUI. The download contains software only: no test assets, model weights or credentials.

Thanks to the researchers and community behind Pixal3D, TRELLIS.2, Hi3D, Idol, HairDeck, Qwen, FLUX, MoGe, ComfyUI and Blender. Foundry connects their work into guided local workflows; model and tool terms remain their own.

Mumblebox Foundry v0.1.0-preview.16

Fix normal Update after a closed Foundry backend leaves a stale runtime record. The check now confirms the runtime has exited and the connection is refused; live jobs and uncertain states remain protected.

Includes the preview.15 community and installer improvements below.

Foundry is a growing community toolbox of useful local workflows, with guided steps, clear choices and automatic software setup.

• About now explains that approach and welcomes workflow suggestions, improvements and pull requests.
• A running Thank You section follows the model/researcher credits. Thanks to Hal Rottenberg (@halr9000) for the installer improvements adapted in this release; his original commits are linked and the installer commit carries scoped co-author credit.
• HairDeck dependency builds use setuptools’ bundled distutils instead of inheriting an unavailable stdlib version from the shell. Existing Torch constraints, requirements and real-entrypoint probes are retained.
• Blender software downloads use its mirror service and an identifiable Foundry client, with the same pinned version, checksum enforcement and verified cache behavior.

The full HairDeck wizard was exercised in standalone ComfyUI with a new reference, two variations, saved-chain generation, interactive previews, sheets, fitting/colors/materials, avatar validation, playable motion, exports and saved History/resume. Focused installer/release checks and an isolated dependency build passed. See the bundled validation document for precise proof and limits.

In Pinokio, choose Update and then reopen/reload Foundry. Models, saved images, assets and settings are preserved. Software downloads exclude test assets, weights and developer state.

Powered by HairDeck and its researchers, with FengLairh’s community ComfyUI integration. Models and tools retain their respective terms. HairDeck makes stylized hair cards and spring-bone VRM; the motion preview is a short head-turn demonstration.

Post-publication verification: the actual Pinokio updater installed preview.16 from preview.14. A fresh indigo-hair reference completed the full installed wizard, visible review/material sheets, avatar validation, playable motion and selected exports. History and Resume restored the final preview. Community guidance and the contributor acknowledgment were visually verified in About.

Mumblebox Foundry v0.1.0-preview.15

Foundry is a growing community toolbox of useful local workflows, with guided steps, clear choices and automatic software setup.

• About now explains that approach and welcomes workflow suggestions, improvements and pull requests.
• A running Thank You section follows the model/researcher credits. Thanks to Hal Rottenberg (@halr9000) for the installer improvements adapted in this release; his original commits are linked and the installer commit carries scoped co-author credit.
• HairDeck dependency builds use setuptools’ bundled distutils instead of inheriting an unavailable stdlib version from the shell. Existing Torch constraints, requirements and real-entrypoint probes are retained.
• Blender software downloads use its mirror service and an identifiable Foundry client, with the same pinned version, checksum enforcement and verified cache behavior.

The full HairDeck wizard was exercised in standalone ComfyUI with a new reference, two variations, saved-chain generation, interactive previews, sheets, fitting/colors/materials, avatar validation, playable motion, exports and saved History/resume. Focused installer/release checks and an isolated dependency build passed. See the bundled validation document for precise proof and limits.

In Pinokio, choose Update and then reopen/reload Foundry. Models, saved images, assets and settings are preserved. Software downloads exclude test assets, weights and developer state.

Powered by HairDeck and its researchers, with FengLairh’s community ComfyUI integration. Models and tools retain their respective terms. HairDeck makes stylized hair cards and spring-bone VRM; the motion preview is a short head-turn demonstration.