Mix Studio v1.3.0

Mix Studio 1.3.0 adds an experimental photo-to-3D workflow and improves the existing image, video, Library, and local setup experience.

• Photo to 3D (experimental): upload a photo, generate a steady H3 orbit with RIFE 2×, and explore the resulting Gaussian splat. Includes centering, visibility controls, further training that preserves the original, and documentation-video export. Enable it in Preferences → Experimental Features.
• Additional model support: Qwen Image 2.1 generation and editing, with Balance/Quality controls, stronger quality defaults, native 2K sizing, and guidance options.
• Workspace and setup improvements: release channels, reorganized preferences, LM Studio support, LTX 2.5, queue improvements, and optional Comfy Kitchen attention for H3.

Existing update highlights remain available alongside new Qwen and 3D demos made with actual Mix Studio outputs.

3D reconstruction is experimental: generated backgrounds can still contain inconsistent geometry. RIFE smooths playback; it does not remove existing blur or create newly observed geometry. Automatic reconstruction-tool installation is available on Windows x64; other platforms require manual local-tool setup.

Install or update

Existing users: open Mix Studio's Updates inbox and install v1.3.0. New users can use the attached Windows or macOS installer. User galleries and model files are preserved.

Full changelog

Experimental photo to 3D

• Added Create → 3D behind Experimental Features, using the existing desktop input/result/Library panels and a single mobile flow.
• Generate a slower H3 orbit from a photo, apply RIFE 2× at 48 fps, then reconstruct cameras with COLMAP and train Gaussian splats with Brush, entirely on the local computer. Preserve the orbit video if reconstruction fails.
• Added verified Windows tool downloads in Preferences, a shared Balance / Quality toggle, persisted progress, and profile-scoped scene access. Keep the generation form focused on the photo and Generate.
• Open the normal Library viewer with generation settings, then choose its 3D thumbnail to expand the scene. Orbit, pan, zoom, choose a subject center, limit visible distance, and export the original PLY. Further training saves a separate version; view adjustments do not modify the saved scene.
• Added a documentation MP4 export combining the source photo, generated orbit, and an actual camera move through the splat. Reuse the existing aspect-ratio picker and restore the viewer camera afterward.
• Added optional native Comfy Kitchen attention for H3, with capability checks and mutually exclusive backend selection. Existing defaults remain unchanged.

Release highlights

• Added a Qwen generation-and-editing animation using actual Mix Studio outputs, an experimental 3D highlight, and a workspace improvements card. Preserved the LTX 2.5, H3 Turbo, Wan Animate 2 and Mix Packs cards.

Managed ComfyUI (Preview only)

• Add optional direct startup for detected Python and registered Desktop installations when generating, with owner controls in Setup.
• Verify process ownership before requesting idle model-memory release or stopping an idle managed backend; preserve separately started backends.
• Keep management off by default and limit availability to Preview/Development. Portable batch files and services retain their existing launch behavior.

Release channels

• Default community updates to exact published Stable releases, with opt-in Preview and explicit Development branch updates.
• Validate release files before applying, refuse downgrades and local-work replacement, and preserve recovery metadata/configuration snapshots.
• Add owner-controlled channel selection and safe update diagnostics; make fresh installers select published stable releases.
• Prepare draft GitHub Releases after cross-platform checks, leaving publication as an explicit maintainer action.

Qwen Image 2.1

• Added model selection to Create and Qwen Image 2.1 whole-image editing, preserving Krea 2 defaults and independent LoRA stacks.
• Added hardware-aware, token-free downloads of official INT8/W4A8/BF16 variants, native ComfyUI capability checks, and a visible research-license notice.
• Added animated, keyboard-accessible model settings, Balance / Quality controls, improved 40-step CFG 2.5 quality defaults, native 2K and advanced guidance settings, and model-aware gallery reuse and metadata.

Prompt AI

• Added LM Studio as a local Prompt AI source through its OpenAI-compatible chat endpoint, including vision references, structured Smart plans, connection testing, and saved URL/model settings.
• Let long local Smart plans run independently from the five-minute prompt-polish deadline, extend Ollama and LM Studio planning waits, refresh ComfyUI deadlines from activity, and remain reconnectable for up to two hours while planning continues.

Preferences

• Replaced the always-expanded Image and Video model settings with compact, single-open model-family lists. All model names remain visible at once, while variants and filenames appear only when their family is opened.
• Reorganized Preferences around user tasks: General now contains guidance and privacy, Prompting contains local/external prompt AI and searchable LoRA suggestions, and System groups setup, access, hardware, storage, and maintenance into status-aware disclosures.
• Added compact value summaries for generation defaults, merged the Krea 2 Edit override into Edit Image, moved SeedVR2 attention into Video post-processing, and renamed Add-ons to Mix Packs.
• Added a global Preferences search that finds settings across every tab, including controls inside collapsed sections, then opens and focuses the selected result.

Generation setup

• Expanded automatic discovery to standard nested ComfyUI Portable folders in Desktop, Downloads, nearby install folders, and known drive roots.
• Made Windows folder dialogs foreground reliably, kept browsing on the Connection step, saved a chosen models folder immediately, and added a clear manual save action for pasted paths.
• Added an explicit option to continue with an existing ComfyUI setup without installing the Krea 2 starter; curated workflow requirements are still checked only when that workflow is selected.
• Let the setup dialog use more height and a three-column component grid on large desktop displays while preserving the existing laptop and mobile layout.
• Made every image aspect ratio through 21:9 visible in the desktop Create panel instead of hiding later landscape choices in an unmarked horizontal scroller.

LTX 2.5 video generation

• Added LTX 2.5 as a separate curated video model while preserving LTX 2.3 as the default, with text-to-video, image-to-video, first/last-frame guidance, synchronized audio, user LoRA stacks, RIFE interpolation, and RTX 4K post-processing.
• Recreated the official 24 fps distilled workflow with Gemma 4 conditioning, dedicated prompt enhancement, separate audio and video VAEs, dual audio/video CFG, the official two-stage sigma schedules, and x2 latent refinement.
• Added guided installation for all six official model files, native node capability checks, model filename preferences, generation reuse and metadata, and clear NVIDIA INT8 ConvRot hardware guidance.
• Promoted LTX 2.5 from Preview after native support merged into ComfyUI, while retaining graph-capability checks and directing older installations to the current stable releases.
• Made the Generation Setup install action acknowledge the first click immediately, and added an inline Hugging Face access panel that saves a read token and resumes gated LTX 2.5 downloads without sending users to Preferences.

Queue panel

• Refined the centered Queue with a subtle solid Jobs/Downloads selector and a lightly translucent surface, while keeping the tabs, GPU/VRAM status, and history/reset actions fixed as only the job or download list scrolls inside the rounded container.

MiniMax H3 LoRAs

• Added native user LoRA stacks for MiniMax H3 Standard and Full BF16 generation, including ordered strengths, trigger phrases, presets, Turbo, Reference Turbo, SageAttention, long-context chunks, generation reuse, and saved metadata.
• Added H3-aware LoRA filtering and icon-only compatibility guidance, excluded the app-managed Turbo adapters from normal selection, and blocked current fused-QKV H3 LoRAs on the incompatible experimental DynTime layout.

RTX 4K setup

• Fixed the RTX post-upscale setup loop by connecting the installer to NVIDIA's official RTX ComfyUI node package and its Python requirement instead of completing an empty install plan.

MiniMax H3 resolution reuse

• Fixed setting reuse for H3 renders with RTX 4K Pass so the picker restores the original S/M/L tier independently from the doubled output dimensions; older gallery entries without saved tier metadata are inferred from their base canvas.

Desktop Library actions

• Added an app-styled right-click menu for generation cards in both the full Library and the Create/Edit side panel, with cursor-aware positioning and the existing Save, Group, Composite, Move, Ungroup, and Delete actions.
• Preserved an existing multi-selection when opening the menu from one of its selected cards, added Ctrl/Command right-click selection, and kept mobile hold-and-sweep selection unchanged.

Wan Animate 2 reliability

• Added the required single-video batch_size input to Wan Animate 2 conditioning graphs, matching the current native ComfyUI node schema and official workflow template.
• Prevented end-of-generation stalls by normalizing performance clips to the model-recommended maximum of 24 fps and the selected 480p canvas before ComfyUI decodes them.
• Replaced memory-heavy multi-window graphs with sequential 81-frame jobs, one-frame continuity handoff, exact overlap trimming, and an FFmpeg join that restores the performance audio once across the finished video.
• Fixed longer performance clips repeating their opening motion by carrying ComfyUI's native Wan Animate 2 frame cursor across sequential jobs instead of resetting every conditioning pass to pose frame zero.
• Stopped forcing the optional multi-gigabyte pose cache, reserved visible progress for decode/save, and reconciled completed video history even if ComfyUI's final WebSocket event is missed.

Mobile Library responsiveness

• Fixed a Chrome mobile relayout loop caused by applying content-visibility to very large Library grids, while retaining the optimization on desktop.
• Stopped mobile Library cards from mounting hundreds of idle video elements and eager poster requests; only the centered preview selected for playback now creates a decoder.
• Prevented the desktop-only result stage from loading a hidden full-resolution video on mobile, which was especially costly after RTX 4K passes.

Update showcase

• Added swipe gestures, large previous and next controls, keyboard navigation, accessible slide announcements, and an eight-second auto-cycle with a visible Pause/Play control.
• Paused showcase cycling while the tab is hidden, while a desktop pointer hovers the card, and whenever reduced motion is requested; released each preview decoder before changing slides.

MiniMax H3 Turbo compatibility

• Added an explicit shared Frames Turbo selector for the recommended v4/step600 adapter or the prior v1/ckpt850 four-step adapter, with the active choice reflected in generation guidance.
• Preserved saved four-step workspaces and made dependency installation resolve the exact selected adapter instead of downloading v4 under a legacy or custom filename.
• Added the official LightX2V v1.0 eight-step FL2VA/T2VA adapter as a compact shared Turbo setup for Frames generation and experimental community-tested Ref2V, with pinned installation, standard ComfyUI LoRA loading, the 12/3 audio-video schedule, native audio-safe sampling when supported, and an icon-only Reference caution.
• Added the official LightX2V v1.0 four-step 768p FL2VA/T2VA adapter as an optional Preferences-only Turbo setup, with pinned installation, automatic four-step sampling, 6/3 video-audio shifts, and an enforced 1344×768 canvas; Reference use remains clearly marked experimental.
• Fixed an H3 generation regression where Turbo step selection referenced a block-scoped mode variable after validation, causing selectedMode is not defined before the graph could be queued.