Repository navigation
Releases: thaakeno/ComfyUI-MiniMax-H3-Studio
Releases · thaakeno/ComfyUI-MiniMax-H3-Studio
Release list
v0.1.0-alpha.20
release: MiniMax H3 Studio v0.1.0-alpha.20
v0.1.0-alpha.19
What's Changed
Added
- Added a polished Custom LoRAs workflow in the Director with up to six compatible H3 adapters, searchable installed-LoRA selection, browser-local favorites, and per-LoRA remembered strengths.
- Added repository-generated light and dark Star History charts so the README no longer depends on Star History's restricted public stargazer API.
Fixed
- Fixed TAEH3 live-preview checkpoint discovery so taeh3.safetensors resolves correctly from ComfyUI's vae_approx model roots without arbitrary filesystem searches.
- Fixed Custom LoRA controls being replaced by older Director surfaces after state changes, including strength sliders that could reset the visible UI instead of updating live.
- Fixed Custom LoRA picker overflow and theme mismatch by keeping the popup viewport-contained, scrollable, searchable, and visually aligned with the Director.
- Hardened the initial Director shelf/layout mount so the Studio reflows reliably when ComfyUI restores or resizes the node.
Changed
- Changed normal Custom LoRA strength controls to a practical 0-3 range and remember the most recently used strength per LoRA; saved values are shown directly in the dropdown and restored when selected again.
- Simplified Custom LoRA rows by removing explicit reorder controls while preserving the stack's add order internally.
- Refined Manager/UAD onboarding and release automation while keeping the maintained workflow and runtime behavior compatible with existing saved workflows.
v0.1.0-alpha.18
What's Changed
Added
- Added a persistent SQLite-backed Director generation library with prompt/seed/reference search, favorites, sampler filtering, sorting, and output re-indexing from embedded H3 Studio image metadata.
- Added SQLite-first history restore with an indexed per-image lookup and PNG/WebP metadata fallback, keeping restores fast while preserving metadata-backed recoverability.
- Added compact lazy history thumbnails and a bounded browser cache so large generation libraries remain practical without storing full Director states in localStorage.
Fixed
- Fixed duplicate History entries by using the actual output image identity across live runs, browser migration, SQLite indexing, and re-indexed PNGs.
- Fixed History card restore, favorite, expand, hover, and scrolling regressions caused by competing capture handlers and self-triggering MutationObserver/card-reorder loops.
- Fixed localStorage QuotaExceededError failures by moving authoritative history state to SQLite/embedded image metadata and limiting browser storage to a compact cache.
- Fixed slow History restores that downloaded and parsed the full output PNG on every click; normal restores now use the already indexed SQLite state and only read PNG metadata as a fallback.
- Fixed severe TAEH3 live-preview artifacts at higher resolutions while restoring realtime cadence by using a distribution-preserving moderate latent decode budget and resizing only the decoded RGB preview to the requested display size.
- Fixed single-reference Guided T2I conditioning so the 512 px semantic Qwen copy optimization applies to one or more references while full-resolution FL2VA keyframes remain untouched.
- Made automatic reference roles prompt-aware and more conservative so obvious subject/character references no longer get mislabeled as environment as often.
- Improved Director timing telemetry so sampling remains backend wall-clock accurate and average-step reporting uses observed sampling intervals rather than including model initialization overhead.
Changed
- Made embedded PNG/WebP H3 Studio metadata the recoverable source of truth for generated-image state while SQLite acts as the fast searchable index and browser storage acts only as a lightweight cache.
- Kept the original polished strip card hover motion and full-resolution expand behavior while serving small cached thumbnails for History browsing.
- Refined the History strip into a generation-library workflow without changing the existing Demos interaction model.
- Enabled TAEH3 Live Preview by default in the maintained H3 Studio workflow while preserving saved user workflow choices.
v0.1.0-alpha.17
What's Changed
Added
- Added a workload-aware Runtime Optimization layer with Auto, OG / Current, Quality, Fast, Low VRAM, and Extreme Low VRAM presets while keeping Base, LightX, PDD, steps, and LoRA sampling profiles independent.
- Added explainable runtime detection and console reporting for GPU/VRAM, actual packed H3 sequence length, route/reference pressure, selected attention backend, head chunking, VAE path, model assets, fallbacks, and the reason Auto chose its effective configuration.
- Added H3 Studio Smart Benchmark Lab with searchable installed transformer and LoRA discovery, per-LoRA strengths, route-aware sampling profiles, runtime A/B helpers, scenario duplication and validation, same-seed execution, labeled comparison grids, and shareable H3B1/H3B1Z benchmark configurations.
- Added one-click H3S1/H3S1Z preset sharing for runtime settings, sampling profile, resolution, connected logical model choices, and ordered custom LoRAs with exact strengths without including prompts, reference images, absolute paths, or secrets.
- Added a curated Director demos shelf with session history and metadata-backed examples that can be reused without turning the main Director renderer into a large static gallery.
- Added H3-native Face Refine after still selection: YOLO-first face detection, source-anchored FL2VA crop rerendering, Auto and Strong modes, optional SAM masking, adaptive denoise/color matching, and visual before/after inspection.
- Added Face Refine setup/readiness tooling plus a Director-style 512-1536 px refine-canvas slider, presets, landscape inspection stacking, and marker on/off inspection controls.
Fixed
- Hardened Director seed handling end to end by keeping seeds below ComfyUI's safe 2^50 ceiling, synchronizing the queued and visible value, and preserving the exact executed seed through generated-image metadata restoration.
- Fixed Face Refine Auto to select at most one eligible face and made missing dedicated four-step LightX assets fall back to a compatible installed FL2VA LightX profile, including the active eight-step v1 profile, or Base Balanced instead of failing solely because the optional four-step LoRA is absent.
- Made failed Face Refine crops fail safe by preserving the original pixels and reporting detector, selection, refinement, timing, and failure details instead of corrupting the output.
- Restored and preserved the verified Director visual hierarchy after late frontend regressions, including complete column scrolling and stable Face Refine tier/readout sizing.
- Reduced redundant Director MutationObserver subtree work while retaining the legacy visual decorators required by the verified UI, cutting repeated DOM rescans during slider/readout updates without redesigning the interface.
- Reduced near-Director canvas zoom/pan lag by isolating heavy DOM compositing, suppressing navigation-time hit testing, avoiding repeated marquee layout measurements, and removing redundant draw-loop size work.
- Restored horizontal scrolling for the curated image strip and excluded Director interior interactions from the canvas navigation lock so shelf browsing remains responsive after the zoom-performance pass.
- Replaced fragile manual benchmark filename entry with installed-asset discovery and exact per-scenario LoRA configuration.
- Kept OG / Current as an explicit unchanged runtime path, made optimized runtime paths capability-detected, and prevented Auto from enabling experimental FFN chunking by default.
Changed
- Made Auto the recommended runtime default and made its policy workload-aware: short FL2VA still packets avoid excessive chunking, constrained REF2VA workloads escalate memory protection, and larger cards prefer Comfy Kitchen when it is actually available.
- Expanded route-aware LightX handling for current FL2VA and REF2VA profiles and let Face Refine reuse compatible already-selected FL2VA acceleration rather than treating a dedicated four-step checkpoint as mandatory.
- Compressed shared preset and benchmark codes automatically when that produces a shorter payload while preserving backward-compatible uncompressed imports and safe missing-asset warnings.
- Documented Face Refine's optional Impact Subpack, Ultralytics, Impact Pack, Segment Anything, MediaPipe, detector-checkpoint, and bundled Haar-cascade license boundaries; none of the external custom nodes or model checkpoints are included in the H3 Studio source archive.
- Updated release-facing installation and feature documentation for alpha.17, including the current release pin, Face Refine setup and cost behavior, optional detector/SAM assets, and the existing Nodes 1.0 compatibility boundary.
v0.1.0-alpha.16
What's Changed
Added
- Added a workload-aware Runtime Optimization layer with Auto, OG / Current, Quality, Fast, Low VRAM, and Extreme Low VRAM presets while keeping Base, LightX, PDD, steps, and LoRA sampling profiles independent.
- Added explainable runtime detection and console reporting for GPU/VRAM, actual packed H3 sequence length, route/reference pressure, selected attention backend, head chunking, VAE path, model assets, fallbacks, and the reason Auto chose its effective configuration.
- Added one-click H3S1/H3S1Z Discord preset sharing for runtime settings, sampling profile, resolution, connected logical model choices, and ordered custom LoRAs with exact strengths without including prompts, reference images, absolute paths, or secrets.
- Added H3 Studio Smart Benchmark Lab with searchable installed transformer and LoRA discovery, per-LoRA strengths, route-aware sampling profiles, runtime A/B helpers, scenario duplication and validation, same-seed execution, labeled comparison grids, and shareable H3B1/H3B1Z benchmark configurations.
Fixed
- Replaced fragile manual benchmark filename entry with installed-asset discovery and exact per-scenario LoRA configuration.
- Kept OG / Current as an explicit unchanged runtime path and made every optimized path capability-detected so older ComfyUI or missing optional kernels fall back instead of silently assuming an unavailable backend.
- Prevented Auto from enabling H3 FFN chunking by default; FFN chunking remains an explicit experimental override while attention-head chunking is reserved for actual memory pressure.
Changed
- Made Auto the recommended runtime default and made its policy workload-aware: short FL2VA still packets avoid excessive chunking, constrained REF2VA workloads escalate memory protection, and larger cards prefer Comfy Kitchen when it is actually available.
- Compressed Discord preset and benchmark codes automatically when that produces a shorter payload while preserving backward-compatible uncompressed imports and safe missing-asset warnings.
- Kept newer native H3 denoise-mask and optional FaceRefine capabilities visible to detection without silently injecting experimental editing/refinement behavior into normal Studio generation.
v0.1.0-alpha.15
What's Changed
Added
- Added first-class route-aware acceleration profiles for Kijai's pruned MiniMax H3 LightX adapters: FL2VA v1.0 4-step rank-31, FL2VA v1.0 8-step rank-24, and REF2VA v0.1 4-step rank-20, while keeping the existing FL2VA v0.1 rank-21 and official full v1.0 8-step profiles.
- Added automatic restoration of the optional H3StudioComparisonView output node in maintained Unified Image workflows so Director output can switch between the generated image and reference comparison views.
- Added clearer model-download documentation for core FL2VA/REF2VA models, recommended acceleration profiles, smaller pruned alternatives, preview/VAE extras, and PDD artifacts.
Fixed
- Removed the permanent JoJo-specific example from the Qwen prompt-writer system instruction so generic anime prompts remain style-neutral unless the user explicitly requests JoJo.
- Added backend and frontend route validation so FL2V LightX adapters cannot silently run on REF2VA and the REF2V adapter is only offered for valid REF2VA reference-generation requests.
- Restored the Director's generated/comparison output tabs when comparison mode is enabled and preserved the PR #37 workflow migration, decode warning, group sizing, and curated TAEH3 controls.
Changed
- Removed the obsolete H3_Studio_Native_Fast_T2I example workflow and its generator/release-check path so H3_Studio_Unified_Image is the single maintained user workflow.
- Refined the maintained workflow UI and model-profile labels around FL2VA versus REF2VA, 4-step versus 8-step, v0.1 versus v1.0, and full versus pruned LightX artifacts.
v0.1.0-alpha.14
What's Changed
Added
- Added native MiniMax H3 VAE decode controls with Auto and Manual tiling, curated 256/320/384/512 tile-size presets, 64/96/128 overlap presets, adaptive tile batching, OOM backoff, and live decode progress.
- Added low-RAM Lightning storage safeguards, swap-aware host-memory pressure telemetry, and recovery for broken text-encoder cache links without taking model residency away from ComfyUI DynamicVRAM.
- Added exact H3 Studio PNG metadata restoration so generated images reopen with the generation state and seed that actually produced them.
Fixed
- Fixed promoted decode controls on the outer sampling subgraph so stale serialized values such as seed 42 cannot be interpreted as the tiling mode.
- Fixed H3 Studio seed restoration from workflow metadata and prompt metadata instead of falling back to stale 0/1/42 values.
- Quarantined the legacy H3 Studio Hub frontend without disabling its backend node registrations, preventing browser resource exhaustion while preserving required H3 nodes.
- Reduced unnecessary decoded-frame cloning and kept native H3 VAE state fully restored after every decode.
Changed
- Kept H3's proven native 256/64 decode geometry as the Auto compatibility default while allowing bounded manual tuning for high-resolution testing.
- Kept the validated FL2VA/REF2VA, Base Quality, LightX v0.1, Mamad8 PDD, preview, save, and workflow paths intact while hardening repeated L4 runs.
v0.1.0-alpha.13
What's Changed
Added
- Added regression coverage for the real flattened H3 video-plus-audio latent pack, prompt-writer runtime behavior, qualified preview execution IDs, and Director sizing stability.
Fixed
- Restored TAEH3 live preview for current flattened H3 multi-latent sampler tensors while retaining compatibility with the older channel-packed layout.
- Fixed live-preview frontend rerenders and qualified subgraph execution IDs.
- Preserved selected Qwen prompt enhancement when enabled while releasing optional helper-model residency before H3 32B conditioning instead of disabling enhancement.
- Stopped the Image Director from progressively shrinking in width or height during ComfyUI layout recomputation; manual resizing remains supported.
Changed
- Kept the proven staged conditioning caches and ComfyUI DynamicVRAM ownership unchanged while using a compact compile-only runtime prompt.
- Made the release path update all tracked version metadata and regenerate uv.lock before publishing so prerelease archives cannot ship stale package metadata.