Skip to content

Releases

Universal Model Loader v1.3.2

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 22 Oct 02:19

Universal Model Loader 1.3.2

  • Fix native MOSS-Image loading on current ComfyUI: allow the standard MODEL/CLIP/VAE route when core exposes moss_image / CLIPType.MOSS_IMAGE.
  • Resolve automatic CLIP selection from MOSS-specific filenames, MOSS-specific Qwen encoder names, and the native MOSSImage model config. Generic Qwen3.5 filenames alone remain ambiguous.
  • Preserve explicit CLIP types, VAE selections, model_name output, socket order, and independent secondary outputs. Core remains responsible for validating weight formats; no downloads, conversions, or substitutions are performed.
  • Keep actionable update-or-typed-loader errors on older cores and retain the optional typed MOSS integration.
  • Add native MOSS routing regression coverage, real-core enum forwarding coverage, and an opt-in CPU weight-loading smoke test.

Validation: 19 Python unit tests, 7 frontend tests, and real-core contracts across 38 supported model config classes. Real CPU MOSS MODEL/CLIP/VAE loading and text encoding were verified during fix validation. End-to-end sampling/image inference was not verified (sampling failed/timed out); this release makes no image-inference claim.

Universal Model Loader 1.3.1

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 18 Oct 00:38

Fix MOSS routing from existing universal nodes: early actionable rejection before core weight loading, and right-click Add MOSS-Image loader (typed). Copies filenames verbatim into a typed sibling without changing existing links or substituting VAE/model files.

The reported raw/pruned INT8 checkpoint remains unsupported; use independently exported MOSS components with moss-image-t8 official encode/sample/decode nodes. No fictional moss CLIP enum.

Verified: 16 Python tests, 7 frontend tests, real installed core contracts (37 config classes), real Base/SFT component integrity validation and INT8 rejection, Registry validation. Descriptor loading only; no inference claim. Restart ComfyUI and refresh browser after updating.

Universal Model Loader 1.3.0

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 18 Oct 00:26

Adds path-only model_name immediately after MODEL in UniversalModelLoaderV2, backward-compatible legacy API handling and UI socket migration, and a typed UniversalMOSSLoader adapter for moss-image-t8 official component loaders. Real local Base/SFT component validation passed; no inference claim. See RELEASE_NOTES_1.3.0.md in the tagged source for compatibility and upgrade details.

Universal Model Loader 1.2.0

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 23 Sep 08:04

Persistent automatic CLIP resolution using encoder hints, loaded model configuration and model filenames. Respects explicit CLIP/VAE selections, improves token-aware path matching, and removes duplicated frontend heuristics. Existing workflows must select auto to enable automatic resolution.

Validated: 8 Python tests, 4 frontend tests, real-core integration across 37 configuration classes, package validation and isolated ComfyUI startup. No full inference test performed.

Universal Model Loader 1.1.0

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 23 Sep 07:18

Summary

  • Add Qwen Image 2.1/QWN2 detection for diffusion model, UNet, and GGUF model hints.
  • Auto-select ComfyUI CLIP type qwen_image for Qwen Image 2.1 naming variants.
  • Prefer the installed Qwen Image 2.1 VAE for QWN2 model-only loads while keeping existing Qwen Image/Krea2 behavior.

Validation

  • node --check web/universal_model_loader.js
  • python -m py_compile __init__.py nodes.py
  • comfy node validate
  • Backend and frontend Qwen Image 2.1 detection tests
  • Local ComfyUI /object_info/UniversalModelLoader verification

Universal Model Loader 1.0.8

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 26 Aug 17:09

Universal Model Loader 1.0.8

Hotfix for the 1.0.7 saved-metadata change.

Changes

  • Keeps inactive hidden model selectors valid for ComfyUI's execution serializer.
  • Still writes null for inactive hidden selectors in saved workflow metadata.
  • Fixes the queue-time “Invalid input / Some input values are not available for this node” regression while preserving the selected-model PNG metadata cleanup.

Validation

  • node --check web/universal_model_loader.js
  • python -m py_compile __init__.py nodes.py
  • comfy node validate
  • Frontend execution+metadata serialization simulation
  • Local ComfyUI /object_info/UniversalModelLoader check

Universal Model Loader 1.0.7

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 26 Aug 15:46

Universal Model Loader 1.0.7

Saved workflow/image metadata fix for dynamic model selectors.

Changes

  • Hidden inactive model selectors now serialize as null in workflow metadata.
  • The active selector for the current model_type keeps the selected model value.
  • Prevents saved PNG workflow metadata/model scanners from reporting the first checkpoint list entry when using diffusion model, UNet, GGUF UNet, or diffusers modes.

Validation

  • node --check web/universal_model_loader.js
  • python -m py_compile __init__.py nodes.py
  • comfy node validate
  • Frontend widget serialization simulation for inactive selectors
  • Local ComfyUI /object_info/UniversalModelLoader check

Universal Model Loader 1.0.6

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 14 Aug 14:45

Universal Model Loader 1.0.6

MiniMax Music3 compatibility update.

Changes

  • Auto-detects latest MiniMax Music3 naming variants as ComfyUI CLIP type minimax.
  • Adds detection for MM3, Music3, RVQ, qwen-rvq, and qwen_rvq file/folder hints.
  • Documents that Music3 workflows should pair the loaded MODEL/CLIP/VAE with ComfyUI's built-in MiniMax Music3 Text Encode and Empty MiniMax Music3 Latent Audio nodes.

Validation

  • node --check web/universal_model_loader.js
  • python -m py_compile __init__.py nodes.py
  • comfy node validate
  • Explicit MiniMax Music3/MM3/RVQ auto-detection tests
  • Local ComfyUI /object_info/UniversalModelLoader check

Universal Model Loader 1.0.5

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 06 Aug 03:26

Universal Model Loader 1.0.5

Optional VAE visibility and dual-output UI fix.

Changes

  • Keeps vae2_name visible for every model_type.
  • Places vae2_name directly after clip2_name so the optional VAE selector is easy to see.
  • Keeps vae2_name defaulting to none.
  • Hides clip2_type and clip2_device until a real second CLIP is selected.
  • Preserves backward-compatible output order: model, clip, vae, loaded_info, clip2, vae2.

Validation

  • node --check web/universal_model_loader.js
  • python -m py_compile __init__.py nodes.py
  • comfy node validate
  • Local ComfyUI /object_info/UniversalModelLoader verified widget order and outputs.

Universal Model Loader 1.0.3

Choose a tag to compare

@Zoltar358-ComfyUI Zoltar358-ComfyUI released this 04 Aug 07:09

Universal Model Loader 1.0.3

Follow-up Node Pack Info extraction fix.

Why 1.0.3

Comfy Registry had already accepted immutable version 1.0.2, but its comfy-nodes endpoint still returned no node definitions. This patch avoids ambiguity between this package's nodes.py and ComfyUI core nodes.py during Registry schema extraction.

Includes the 1.0.2 metadata work

  • Registry icon and banner assets.
  • Richer pyproject metadata for ComfyUI Extensions' Node Pack Info panel.
  • Clearer node preview description.
  • Hardened fallback behavior for lean Registry extraction environments.

Validation

  • node --check web/universal_model_loader.js
  • python -m py_compile init.py nodes.py
  • comfy node validate
  • self-name collision simulation
  • lean Registry extraction simulation
  • runtime object_info-style import check