Changelog

All notable changes to ComfyUI Smart Model Loader are documented in this file.

Entries follow conventional commit prefixes:

2026-09-08

Version: 1.0.13

• Feat

  • Show the running Smart Model Loader version in its ComfyUI settings and provide a loopback-only, explicitly confirmed update action that replaces tracked files with official main, preserves untracked user data, installs requirements, and reports the required restart.

Version: 1.0.8

• Fix

  • Validate every MiniMax H3 response before returning it, retry an incomplete or audio-only response once with the same reference images and seed plus a focused format-recovery instruction, and deterministically reconstruct any still-missing fields from the requested action and usable generated audio instead of failing the workflow.

  • Preserve structured motion fields containing chronological words such as then by restricting shared planning cleanup to genuine leading meta-reasoning, preventing valid H3 visual prompts from being reduced to audio-only output.

  • Remove the redundant I2VA first-frame reference sentence from H3 prompts and generated output while continuing to send the image to the vision backend, and migrate untouched bundled prompt entries without overwriting customizations.

• Refactor

  • Rename the Smart LM Loader implementation module to RvLoader_SmartLMLoader.py to distinguish it from the separate Smart Model Loader project while preserving the class and serialized node ID.

2026-09-07

Version: 1.0.12

• Feat

  • Add opt-in MiniMax H3 model sampling to Smart Model Loader with independent video/audio flow shifts, ComfyUI-compatible AV sampling, preserved noise scale and transformer options, and template round-trip support.

• Fix

  • Migrate pre-MiniMax positional Smart Loader workflows after normalizing older denoise and audio-VAE layouts, preserving every later widget value.

• Docs

  • Document MiniMax H3 sampling controls, defaults, ranges, and opt-in template behavior.

2026-09-06

Version: 1.0.11

• Fix

  • Load standalone audio VAEs through ComfyUI's general VAE architecture detection, supporting MiniMax H3 and standalone LTX formats while reserving prefix extraction for baked LTX audio weights and failing before pipe construction when a requested audio VAE cannot load.

  • Group Smart Loader audio-VAE controls directly beneath the regular VAE controls and migrate current and pre-denoise positional workflow arrays before configuration so every saved value remains attached to its original field.

• Docs

  • Describe general ComfyUI-supported audio VAEs in Smart Loader and dual-VAE tooltips and guides.

Version: 1.0.10

• Feat

  • Add Select All and Deselect All controls to the Download Manager queue for bulk job actions and bundle export.

• Fix

  • Label verified-download diagnostics with the active provider so Hugging Face failures no longer appear under the CivitAI prefix.

  • Allow confirmed template deletion when its model is already absent, transactionally removing orphaned integrity sidecars only when one configured model root owns them and preserving ambiguous, shared, or symlinked sidecars.

2026-09-02

Version: 1.0.9

• Feat

  • Generate standalone and Smart Model Loader CLIP architecture choices from the installed ComfyUI CLIPType enum, preserving its declaration order, filtering unsupported recipes on older versions, and exposing future recipes automatically.

• Fix

  • Resolve CLIP selections only through the installed enum and reject stale, malformed, or unavailable values before checkpoint, CLIP, GGUF, or Nunchaku loading begins instead of silently falling back to Stable Diffusion.

• Breaking

  • Remove the external-CLIP sdxl alias; existing workflows must select the canonical stable_diffusion recipe.

Version: 1.0.8

• Fix

  • Keep the Smart Model Loader denoise value independent of templates by omitting it from template saves and ignoring it during template loads, while preserving the default fallback for workflows saved before the widget existed.

2026-09-01

Version: 1.0.6

• Feat

  • LoRA Stack and Nunchaku PuLID ownership: Adopt Lora Stack [Eclipse], Lora Stack apply [Eclipse], Nunchaku PuLID Loader [Eclipse], and Nunchaku PuLID Apply [Eclipse] from Eclipse with their serialized IDs and input/output contracts unchanged. Serve the LoRA Stack mode-chip frontend from this pack and bind PuLID to the same vendored ComfyFluxWrapper used by Nunchaku Flux loading, eliminating cross-pack class-identity failures.

• Fix

  • Recognize Krea2 in Conditioning Zero Out logs as a variable-length model, preserving prompt-dependent token counts instead of reporting the connected model as unknown.

  • Make the Nunchaku PuLID CPU InsightFace provider bypass cuDNN during face embedding preprocessing, allowing it to recover from mixed cuDNN sublibrary installations while restoring the previous backend state afterward. Report an actionable environment error when the CUDA provider encounters the same mismatch.

  • Apply a self-detecting tuple-index compatibility fix to affected installed Nunchaku runtimes, resolve the Qwen transformer from both stable and nightly module layouts, avoid warning-only missing model-config probes, suppress only the empty-FP32-list and known PuLID dependency deprecations at their integration boundaries, and report accelerator cache cleanup only for available backends whose cache was actually cleared.

  • Align Nunchaku ZImage with its vendored runtime contract: validate the active CUDA architecture and model quantization before loading, use ComfyUI dynamic VRAM management, hide Qwen-only offload controls, and omit those ineffective values from new templates and backend calls while preserving old workflow schemas.

• Docs

  • Document the transferred LoRA Stack and Nunchaku PuLID nodes, their optional Nunchaku dependencies, Smart Model Loader ownership, a validated Linux repair for duplicate ONNX Runtime installations and mismatched cuDNN/cuBLAS wheel contents, and the architecture-specific Nunchaku ZImage controls, hardware requirements, and successful live INT4 qualification.

2026-08-24

Version: 1.0.5

• Feat

  • Add a 0.0–1.0 denoise control below cfg in the Smart Model Loader sampler chip, persist it in templates, include it in the emitted PIPE, and append it to IO Checkpoint Loader inputs and outputs.

  • Automatically detect and migrate previous compact and seed-button positional Smart Model Loader layouts before node configuration, preserving flux_guidance and every later widget value when older workflows load.

• Fix

  • Let ComfyUI report one field-specific prompt-validation error for a missing Smart Model Loader primary model instead of repeating one node-wide error for every widget.

  • Preserve the Pipe sampler's manually resized height when tiled-decode controls change while previews are enabled so the preview absorbs row changes, while retaining automatic resizing when preview is None.

2026-08-23

Version: 1.0.4

• Fix

  • Bundle the missing Krea2_GPT Smart Model Loader template used by iGEN Simple, including verified CivitAI acquisition metadata for its primary diffusion model.

  • Isolate bundled GGUF/Nunchaku mutable defaults, correct GGUF tokenizer model-type assignment, bind replacement callbacks to their intended blocks, and preserve causal exception chains.

  • Ignore stale hidden integrity modes during execution and expose verified CivitAI re-download recovery when a present model file has a hash mismatch.

  • Update the copied pipe sampler's tiled-decode controls immediately in classic and Nodes 2.0 views when tiled_decode changes.

• Refactor

  • Modernize tracked first-party and bundled Python imports, typing syntax, collections, dictionary iteration, control flow, and formatting without changing node IDs, schemas, routes, or loader contracts.

• Chore

  • Enable Ruff's complete rule set for tracked Python with documented, scoped exceptions for ComfyUI callbacks, serialized identifiers, hardware orchestration, and maintained bundled integrations.

2026-08-22

Version: 1.0.3

• Fix

  • Omit hidden flux_guidance defaults from Smart Model Loader pipes for Krea2 and other non-Flux model/CLIP combinations while retaining it for supported Flux and Flux2 loaders.

Version: 1.0.2

• Feat

  • Add a persisted color picker for loader chip bars and selected chips, with derived hover, border, and contrast colors.

  • Match loader combo-chip popup widths to their rendered trigger bars without stretching or shrinking individual chips.

• Fix

  • Set the Download Manager dialog background to #3a3a3a.

2026-08-19

Version: 1.0.1

• Feat

  • Add workflow-compatible basic and Advanced CLIP encoding, model-aware Conditioning Zero Out, checkpoint pipe IO, and pipe sampling so the standalone pack supplies a complete loader-to-image pipeline.

  • Add standalone queue-time special-seed controls plus tiled-decode and live/final preview behavior for Eclipse KSampler (Pipe) [Eclipse].

  • Place all eleven nodes under their own Smart Model Loader top-level menu while retaining the historical workflow node IDs.

• Refactor

  • Take exclusive runtime and frontend ownership of the five transferred node IDs without cross-pack imports, endpoint probes, configuration sharing, or workflow migration.

• Docs

  • Document the complete pipeline, optional Eclipse integrations, independent installation, correct template path, support links, and Smart Model Loader branding.

  • Add a Nodes 2.0 visual walkthrough with annotated loader modes, feature chips, template loading and missing-model recovery, conditioning, pipe sampling, menu discovery, and Download Manager views.

2026-08-17

Version: 1.0.0

• Feat (New)

  • Extract the six workflow-compatible Eclipse diffusion loaders into a standalone ComfyUI pack.

  • Include the verified CivitAI/Hugging Face Download Manager with persistent queue recovery and portable bundle compatibility.

  • Add copy-only migration for Eclipse templates, relevant configuration values, and queue state.

• Fix

  • Guard every backend-writing setting against ComfyUI's automatic first change callback while hydrating defaults and credential masks only from the pack's redacted config endpoint.

• Refactor

  • Give Smart Model Loader → General exclusive ownership of stable SmartModelLoader.* settings, endpoints, credentials, log level, sliders, retry policy, and legacy-format policy.

  • Upgrade migration markers atomically with value-free examined-key confirmation while preserving non-default destination values on conflicts.

• Docs

  • Document independent Eclipse, Smart Model Loader, and Smart LM Loader settings and configuration ownership.

• Security

  • Retain immutable provider identity, bounded same-origin mutations, registered-root containment, resumable partial validation, integrity sidecars, and atomic promotion.

• Breaking

  • Replace Eclipse-owned HTTP routes and Download Manager identifiers with the /smart-model-loader namespace; API clients must migrate.

ComfyUI Smart Model Loader

A complete, pipe-first diffusion workflow for ComfyUI: load models, build text conditioning, sample, decode, and acquire verified model files without turning your canvas into a wall of connections.

Version 1.0.1 includes eleven Nodes 2.0-ready nodes. ComfyUI Eclipse is optional.

Visual tour

Find every node in one menu

The pack has its own top-level Smart Model Loader category. Nodes are grouped into Conditioning, Loader, Pipe, and Sampler submenus.

Start compact

The default Smart Model Loader exposes the common checkpoint path: model type, baked CLIP, baked VAE, memory cleanup, and one PIPE output.

Open the chip bar

The feature bar controls the node's optional sections. Selected chips reveal their widgets immediately; inactive sections stay out of the way.

Available chips are templates, clip, vae, audio_vae, latent, sampler, lora, model_sampling, block_swap, memory_cleanup, integrity, and seed. Block swap is disabled when ComfyUI's native dynamic VRAM management already owns that job.

Build the loader you need

Enable latent, sampler, seed, LoRA, integrity, or architecture-specific sampling only when the workflow needs them. The serialized workflow retains the selected feature state.

Load a template

Enable templates, choose Load, then select a bundled or saved template. The selection applies the template's enabled loader and generation settings. Delete Template removes only that saved configuration—not any model files.

Templates live in ComfyUI/custom_nodes/ComfyUI_SmartModelLoader/templates/.

Recover a model referenced by a template

Loading a template never downloads files automatically. If its primary model is not installed, the loader keeps the referenced filename selected with a (missing) marker, restores the saved CivitAI AIR or SHA-256 identity, infers the registered destination, and reveals Download from CivitAI.

The bundled Krea2_RedCraft template below restores its Krea2 CLIP, external VAE, sampler settings, and missing UNet reference. Pressing the download action resolves the provider file for the selected precision, verifies it, saves it under the shown model role, and refreshes the node. Until that button is pressed, no transfer starts.

Use focused loaders when you want direct control

The standalone Model, CLIP, and VAE loaders are useful for modular UNet, Nunchaku, GGUF, video, and audio setups. IO Checkpoint Loader merges direct components over an incoming PIPE and exposes the result on individual sockets.

Shape conditioning

Use the basic encoder for ordinary scheduled conditioning. The Advanced encoder adds a multiplier and Krea2 layer-rebalancing presets, while Conditioning Zero Out clears tensors and can choose a known model's base token length.

Sample from the PIPE

Eclipse KSampler (Pipe) reads its model and VAE from the PIPE, accepts positive and negative conditioning, supports image or latent input, and returns a copied PIPE containing the sampled latent and decoded image. It also provides pipe/widget precedence, tiled VAE decode, preview modes, and queue-time seed controls.

Acquire exact model files

Open Download Manager (Beta) from the Smart Model Loader menu. Inspect a CivitAI AIR/hash/URL or Hugging Face repository, select exact compatible files, choose their registered model destinations, and monitor the persistent queue.