CWK Checkpoints Preset Manager

A ComfyUI custom node that combines a per-model preset system with a full-featured model manager — letting you browse, organize, update, download, and delete checkpoint models directly from ComfyUI, while ensuring every model always loads with its ideal generation settings.

---

Features

🎨 Custom Canvas Node

• Fully custom-drawn node UI using LiteGraph onDrawForeground — no standard ComfyUI widgets
• Model thumbnail displayed directly on the node with cover-fit crop and rounded corners
• Inline editable preset rows with ◀ ▶ arrow buttons for stepping numeric values
• Click the center value to open an inline text input directly over the value cell for direct numeric entry — Enter to commit, Escape to cancel, click outside to commit
• Compact listbox dropdowns for list fields (Sampler, Scheduler, RNG, Resolution Preset, CLIP, Clip Type, VAE) — single-click to open, max 6 visible items with scroll, smart up/down positioning
• Three action buttons: 📂 Load Model, ↩ Reset, 💾 Update Preset
• Node size auto-calculated based on content; minimum width enforced
• Last-used model persistence — the node remembers the last loaded model and automatically restores it on page refresh or ComfyUI restart
• Image metadata drag-and-drop — drag an image with embedded metadata onto the node to automatically load the model that generated it
• Quick-load model dropdowns — two dropdown buttons drawn directly on the node (⚡ Checkpoint and ⚡ Diff / GGUF) for instant model switching without opening the Model Browser; models are grouped and badged by base model type (SD15, SDXL, Pony, Flux, etc.)

💾 Per-Model Preset System

• Save and load generation presets per checkpoint model — each model remembers its own settings
• Preset fields: Sampler, Scheduler, CFG, Steps, Clip Skip, Width, Height, RNG, CLIP, Clip Type, VAE
• Resolution preset dropdown — quickly set width and height from a list of common resolutions for SD 1.5, SDXL, and Flux models (square, landscape, and portrait aspects)
• Batch size control — set batch size directly on the node (1–64); outputs an empty latent tensor sized to the current width, height, and batch
• Auto-populate presets from CivitAI — when adding a new model or fetching thumbnails, the node extracts generation metadata (sampler, scheduler, CFG, steps, clip skip, resolution) from CivitAI example images and automatically saves them as the model's preset
• Presets stored in checkpoint_presets.json and merged with per-execution widget overrides at runtime
• Clip skip applied directly to the CLIP model via clip.clip_layer() — no extra node needed
• RNG control — choose between CPU, GPU, or NV (NVidia Philox) noise generation per model:
  • Self-contained RNG implementation adapted from ComfyUI_smZNodes by shiimizu — no external dependency required
  • Uses the smZ_opts model_options protocol for full compatibility — if smZNodes is also installed, the two coexist transparently
  • Includes vendored Philox 4×32 RNG for NVidia-compatible cross-GPU reproducibility
  • Patches comfy.sample.prepare_noise and hijacks k-diffusion's default_noise_sampler for consistent noise in all samplers
• External CLIP & VAE — select embedded (uses the checkpoint's built-in CLIP/VAE) or pick any external CLIP or VAE model file from ComfyUI's clip/vae folders; the selection is saved as part of the preset so each model can have its own preferred CLIP and VAE
• CLIP type selector — when using an external CLIP model, choose the architecture type (stable_diffusion, flux, sd3, wan, etc.) to ensure correct loading; the type is saved per-model in the preset
• GGUF CLIP support — .gguf text encoder files from ComfyUI-GGUF (city96) are supported as external CLIP models with proper type selection

🔀 Sampler / Scheduler Fallback System

Some models on CivitAI use sampler or scheduler names that don't exist in the user's ComfyUI installation — either because they follow A1111 naming conventions, contain typos, or reference custom samplers that aren't installed. Instead of crashing at generation time, the node resolves names through a multi-layer fallback:

1. Alias table — maps common A1111 and CivitAI display names to ComfyUI internal names: | CivitAI / A1111 | ComfyUI | |---|---| | Euler a | euler_ancestral | | DPM++ 2M | dpmpp_2m | | DPM++ 2M SDE Karras | dpmpp_2m_sde + scheduler karras | | euler_ancestral_cfg++ | euler_ancestral_cfg_pp |

2. Pattern normalisation — ++ → _pp, spaces → underscores, case folding

3. Embedded scheduler extraction — detects when a scheduler name is appended to the sampler (A1111 style) and splits them apart

4. Fuzzy matching — finds the closest installed sampler/scheduler using difflib.SequenceMatcher with substring containment bonus

5. Safe fallback — if nothing matches, falls back to euler / simple with a console warning

Fallback is applied both at execution time and when auto-populating presets from CivitAI metadata. All resolution steps are logged to the console.

🔧 GGUF Model Support

The node natively supports .gguf quantized model files from ComfyUI-GGUF (city96):

• Automatic discovery — .gguf files in checkpoints/, diffusion_models/, and unet/ folders appear in the model list automatically
• Smart loader — dynamically finds city96's UnetLoaderGGUF at runtime by scanning sys.modules, regardless of how the extension was installed or what its folder is named
• GGUF CLIP — .gguf text encoder files are loaded via city96's CLIPLoaderGGUF with the correct CLIP type
• No embedded CLIP/VAE — GGUF models (like diffusion models) require external CLIP and VAE to be set in the preset
• Full feature support — all existing features work with GGUF models: presets, thumbnails, CivitAI metadata, favorites, updates, downloads, and deletion

Requires: ComfyUI-GGUF by city96. If not installed, the node will show a clear error message with install instructions.

📂 Model Browser & Manager

The Model Browser is a full visual panel that opens from the node's 📂 Load Model button. It serves as both a model selector and a complete model manager:

• Visual card grid with lazy-loaded thumbnails fetched from CivitAI
• Search by model name with real-time filtering
• Base model filter — filter the grid by architecture: SDXL, Illustrious, Pony, NoobAI, SD15, Qwen, Flux, Chroma, Wan, ZImage, and others
• ⭐ Favorites — mark preferred models with a star toggle and filter to show favorites only
• NSFW detection — models flagged by CivitAI's nsfw_level are automatically blurred; click the 👁 eye button to reveal individual cards
• Video thumbnail support — .mp4 and .webm thumbnails play inline on cards
• Drag to reposition and resize handle — panel size and position persist across sessions via localStorage

CivitAI Integration

• Fetch Thumbnails/Infos — streams metadata from the CivitAI API for all models with a live progress bar (SSE-based, concurrent requests); also auto-populates presets from example image metadata and auto-refreshes the model list and sidebar when complete
• ↺ Rebuild Cache — force re-fetches all metadata from CivitAI, ignoring the local cache (favorites and manual overrides are preserved)
• Model Info overlay — right-click any card → Model Info to view a detailed overlay with model name, version, base model, file size, file path, tags, example images gallery (with lightbox and 📋 metadata viewer), and the full CivitAI model description (HTML rendered)
• 📋 Image Metadata Viewer — click the 📋 button on any example image in the Model Info overlay to view its generation metadata: positive/negative prompts (with copy-to-clipboard), sampler, scheduler, CFG, steps, clip skip, seed, and size
• Editable metadata — edit the display name, version, and base model directly in the Model Info overlay; manual edits survive CivitAI refreshes
• CivitAI API key management — set your API key directly in the panel with a validation button to confirm it works

Model Updates & Downloads

• 🔍 Check Updates — batch-checks CivitAI for newer versions of all your installed models; cards with available updates are marked with a ⬇ badge
• Version checker — right-click a card → Check for updates to view all available versions of a model, including version name, base model, file size, release date, and early access status
• Download any version — download any version directly from CivitAI with a real-time progress bar per version row; the download is streamed with chunked writes and clean error handling (including early access detection)
• Auto refresh after download — after a download completes, the model list is automatically reloaded and CivitAI metadata is fetched for the new file so it appears with its thumbnail immediately
• Auto refresh after delete — deleting a model (right-click → Delete model) removes the file, clears all cached metadata and presets, and refreshes the model list automatically
• 🔄 Reload Models — manually reload the model list from disk without fetching CivitAI data; also auto-fetches thumbnails for any newly discovered models that don't have them yet

Thumbnails

• Thumbnails are fetched automatically from CivitAI based on the SHA-256 hash of the model file (hashes are cached for performance)
• Pick thumbnail from CivitAI — right-click a card to browse and select from all images in the model's CivitAI gallery
• Set local thumbnail — upload a custom image or video from your filesystem as the thumbnail for any model
• Local thumbnails are served from the local_thumbnails/ directory and persist across sessions

Sidebar Preset Editor

• The right sidebar displays and edits the preset for the selected model
• All preset fields are shown: Sampler, Scheduler, CFG, Steps, Clip Skip, RNG, Width, Height, CLIP, Clip Type, VAE
• Click Edit Preset to unlock the fields, modify values, then click Save Preset to persist
• Sampler and scheduler lists are loaded dynamically from ComfyUI's /object_info endpoint, so they automatically reflect any installed sampler extensions
• Auto-refreshes after Fetch Thumbnails/Infos to immediately display newly populated preset values

---

Installation

Via ComfyUI Manager (recommended)

Search for CWK Checkpoints Preset Manager in the ComfyUI Manager node list.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/cowneko/CWK_Checkpoints_Preset_Manager.git

Then restart ComfyUI.

---

Requirements

• ComfyUI
• A CivitAI API key (required for fetching thumbnails, model info, and downloading)
• ComfyUI-GGUF by city96 (optional — required only for loading .gguf quantized models and GGUF text encoders)

Note: smZNodes is no longer required. As of v1.2.0, the RNG subsystem is fully self-contained using code adapted from smZNodes. If smZNodes is installed alongside, the two coexist transparently with no conflicts.

---

Usage

1. Add the node

Right-click the canvas → Add Node → CWK / CWK Model Preset Manager

2. Load a model

Click 📂 Load Model to open the Model Browser. Select a model card and click Load Model. The node will load the checkpoint and apply its saved preset (or defaults if no preset exists yet).

Tip: The node remembers the last model you used. When you add a new node or refresh the page, it will automatically restore the last loaded model with its preset and thumbnail.

3. Load a model from an image

Drag an image with embedded generation metadata (ComfyUI workflow or A1111-style parameters) onto the node. The node will parse the metadata, identify the checkpoint model, and load it with its saved preset.

4. Edit a preset

On the node, click the ◀ ▶ arrows to step numeric values, or click the center value to type a number directly in the inline editor. Dropdown rows (Sampler, Scheduler, RNG, Resolution Preset, CLIP, Clip Type, VAE) open a compact listbox on single click. You can also edit presets from the sidebar in the Model Browser panel.

5. Use resolution presets

Use the Res Preset dropdown to quickly set width and height to a standard resolution. Presets are grouped by model architecture:

• SDXL / Pony / SD3 — 1024px base (1024×1024, 896×1152, 1344×768, etc.)
• SD 1.5 — 512px base (512×512, 448×576, 672×384, etc.)
• Flux — 1024px base (same as SDXL)

Select (preset) to use the width/height from the model's saved preset instead.

6. Batch size & empty latent

Use the Batch row to set the batch size (1–64). The node outputs an empty LATENT tensor sized to the current width × height × batch, ready to connect directly to a sampler — no separate Empty Latent Image node needed.

7. Save a preset

Click 💾 Update Preset to save the current values as the preset for the loaded model. Next time you load this model, these settings will be restored automatically.

8. Reset to preset

Click ↩ Reset to restore all values to the saved preset for the current model, discarding any unsaved changes.

9. RNG modes

Use the RNG dropdown to select the noise generation source:

• cpu — generate noise on CPU (default, deterministic across all GPUs)
• gpu — generate noise on the active GPU (faster, but results may vary across different GPU models)
• nv — NVidia Philox RNG (produces identical noise to torch.randn(..., device='cuda') but runs on CPU — enables cross-GPU reproducibility matching NVidia's CUDA RNG)

10. Model Sampling type

Use the Model Sampling dropdown to select the prediction type for the loaded model:

• eps — epsilon prediction (default for most models)
• v_prediction — velocity prediction (for v-pred trained models like some NoobAI/SDXL variants)
• lcm — Latent Consistency Model sampling
• x0 — direct x0 prediction
• img_to_img — eps-based, useful as a tag for img2img workflows

The sampling type is applied as a model patch at execution time and is saved per-model in the preset.

11. External CLIP / VAE

Use the CLIP and VAE dropdowns on the node or in the Model Browser sidebar to select an external model file, or leave as embedded to use the checkpoint's built-in CLIP/VAE. If an external file fails to load, the node falls back to the embedded version with a console warning.

When loading an external CLIP model, use the Clip Type dropdown to select the correct architecture:

• stable_diffusion — SD 1.5, SDXL, Illustrious, Pony, NoobAI
• flux / flux2 — Flux.1 models
• sd3 — SD3 and SD3.5
• wan — Wan Video models

This is especially important for .gguf text encoders, which require the correct type to load properly via city96's GGUF CLIP loader.

12. Manage your models

Use the Model Browser to organize your collection:

• ⭐ Star your favorite models and toggle the favorites filter
• Right-click any card for options: view model info, pick or set thumbnails, check for updates, refresh CivitAI data, or delete the model
• Check Updates to see which models have newer versions on CivitAI, then download any version directly from the version checker

13. View image metadata

In the Model Info overlay, click the 📋 button on any example image to view its generation metadata — prompts, sampler, scheduler, CFG, steps, clip skip, seed, and size. Use the 📋 Copy buttons to copy prompts to the clipboard.

14. Sampler / scheduler fallback

If a preset or CivitAI metadata contains a sampler or scheduler name that isn't installed (e.g. A1111 naming like "Euler a" or "DPM++ 2M Karras"), the node automatically resolves it to the closest available match. If no close match exists, it falls back safely to euler / simple. Check the ComfyUI console for resolution logs:

[CWK] Sampler alias: 'Euler a' → 'euler_ancestral'
[CWK] Sampler ++ fix: 'euler_ancestral_cfg++' → 'euler_ancestral_cfg_pp'
[CWK] ⚠ Sampler 'unknown_sampler' not found — falling back to 'euler'

15. GGUF models

To use .gguf quantized models:

1. Install ComfyUI-GGUF by city96
2. Place .gguf model files in your ComfyUI/models/unet/ or ComfyUI/models/diffusion_models/ folder
3. The models will appear automatically in the Model Browser
4. Set an external CLIP (with the correct Clip Type) and VAE in the preset, since GGUF models don't include embedded CLIP/VAE
5. Click 💾 Update Preset to save these settings — they'll be restored automatically next time

16. Infos output & Infos Extractor

The main node outputs an infos STRING pin containing a JSON dict with all resolved preset values. Connect it to the CWK Infos Extractor node to fan out individual values as STRING outputs — useful for display nodes, logging, or feeding values to other nodes that accept string inputs.

Model, VAE, and CLIP names are automatically cleaned: sd15\darkSushiMixMix_colorful.safetensors becomes darkSushiMixMix_colorful.

17. Quick-load model switching

Use the ⚡ Checkpoint and ⚡ Diff / GGUF dropdown buttons on the node to instantly switch between models without opening the Model Browser panel. Models are grouped by base model type with colored badges:

| Badge | Architectures | |---|---| | SD15 | SD 1.4, SD 1.5 | | SDXL | SDXL 1.0 | | SDXL-T | SDXL Turbo | | SDXL-L | SDXL Lightning | | Illust | Illustrious | | Pony | Pony | | Noob | NoobAI | | Flux | Flux.1 | | Chroma | Chroma | | Qwen | Qwen | | Wan | Wan | | ZImg | ZImage |

Clicking a model loads it immediately with its saved preset, thumbnail, and metadata. The model list refreshes automatically whenever you load a new model.

---

Outputs

CWK Model Preset Manager

| Output | Type | Description | |---|---|---| | MODEL | MODEL | Loaded model with RNG setting applied | | CLIP | CLIP | Embedded or external CLIP model with clip_skip applied | | VAE | VAE | Embedded or external VAE | | LATENT | LATENT | Empty latent tensor sized to width × height × batch_size | | steps | INT | Preset or override step count | | cfg | FLOAT | Preset or override CFG scale | | sampler_name | SAMPLER | Preset or override sampler (with fallback resolution) | | scheduler | SCHEDULER | Preset or override scheduler (with fallback resolution) | | width | INT | Preset or override width | | height | INT | Preset or override height | | infos | STRING | JSON string with all resolved preset values (see below) |

clip_skip, rng, clip_name, clip_type, vae_name, and batch_size are applied internally and do not appear as separate output pins (batch_size is reflected in the LATENT tensor dimensions). All values are accessible via the infos output.

CWK Infos Extractor

A companion node that takes the infos JSON string from the main node and fans it out to individual STRING outputs.

| Output | Description | |---|---| | model_name | Cleaned model name (no subfolder, no extension) | | vae_name | Cleaned VAE name (no subfolder, no extension) | | clip_name | Cleaned CLIP name (no subfolder, no extension) | | clip_type | CLIP architecture type (e.g. stable_diffusion, flux) | | sampler_name | Resolved sampler name | | scheduler | Resolved scheduler name | | cfg | CFG scale value | | steps | Step count | | clip_skip | Clip skip value | | rng | RNG source (cpu, gpu, nv) | | model_sampling | Model sampling type (eps, v_prediction, etc.) | | width | Image width | | height | Image height | | resolution | Resolution string (e.g. 1024x1024) | | batch_size | Batch size |

Name cleaning: Model, VAE, and CLIP names have their subfolder prefix (Pony\, Flux\, etc.) and file extension (.safetensors, .ckpt) automatically stripped — e.g. Pony\autismmixSDXL_autismmixPony.safetensors → autismmixSDXL_autismmixPony.

---

API Endpoints

The node registers the following REST routes on the ComfyUI server:

| Method | Path | Description | |---|---|---| | GET | /cwk/models | List all checkpoint, diffusion, and GGUF models with presets and metadata | | GET | /cwk/preset | Fetch preset for a model | | POST | /cwk/preset | Save preset for a model | | GET | /cwk/clips | List available CLIP models (embedded + all files including GGUF) | | GET | /cwk/vaes | List available VAE models (embedded + all files) | | GET | /cwk/last_model | Get the last-used model name with its preset and metadata | | POST | /cwk/last_model | Save the last-used model name | | GET | /cwk/resolution_presets | List all resolution preset labels with width/height values | | GET | /cwk/civitai/meta | Get cached CivitAI metadata for a model | | POST | /cwk/civitai/meta/edit | Manually edit metadata fields (civitai_name, version_name, base_model) | | POST | /cwk/civitai/fetch/stream | SSE stream for bulk CivitAI metadata fetch | | POST | /cwk/civitai/refresh | Refresh CivitAI data for a single model | | POST | /cwk/civitai/refresh/all | Force refresh all models (SSE stream) | | GET | /cwk/civitai/validate | Validate a CivitAI API key | | GET | /cwk/civitai/cache | Get all cached metadata | | DELETE | /cwk/civitai/cache | Clear all cached metadata | | POST | /cwk/civitai/thumbnail/set | Set thumbnail URL for a model | | POST | /cwk/civitai/thumbnail/local | Upload a local image as thumbnail | | GET | /cwk/civitai/images | Fetch image list for a model from CivitAI | | GET | /cwk/civitai/versions | List all versions of a model from CivitAI | | GET | /cwk/civitai/model-description | Fetch full model description and tags | | POST | /cwk/civitai/check-updates | Batch check for newer model versions | | POST | /cwk/civitai/download | Download a model version (SSE progress stream) | | POST | /cwk/model/favorite | Toggle favorite flag for a model | | DELETE | /cwk/model | Delete a model file and all cached data |

---

RNG Implementation

The RNG subsystem is self-contained and adapted from ComfyUI_smZNodes by shiimizu. It provides accurate CPU, GPU, and NVidia Philox noise generation without requiring smZNodes to be installed.

The implementation consists of three internal modules:

• cwk_rng_shared.py — Options class using the smZ_opts protocol key, with all fields smZNodes expects for full interoperability
• cwk_rng_philox.py — Philox 4×32 random number generator that reproduces NVidia CUDA torch.randn output on CPU
• cwk_rng.py — prepare_noise() function that replaces comfy.sample.prepare_noise, with stack introspection to read per-model RNG settings, TorchHijack for batch-consistent noise, and k-diffusion default_noise_sampler hijacking

If smZNodes is also installed, the two systems coexist transparently — CWK stores its options under the same smZ_opts key with all expected fields populated, so smZNodes can read them without errors.

---

Sampler / Scheduler Fallback

When a preset or CivitAI metadata contains a sampler or scheduler name that doesn't match the locally installed ComfyUI samplers, the node applies a multi-layer resolution strategy:

1. Alias table — covers all common A1111 display names (Euler a, DPM++ 2M, DPM++ SDE Karras, etc.) and CivitAI-specific patterns (euler_ancestral_cfg++ → euler_ancestral_cfg_pp)
2. Pattern normalisation — ++ → _pp substitution, space-to-underscore, case folding
3. Embedded scheduler extraction — splits "DPM++ 2M SDE Karras" into sampler dpmpp_2m_sde + scheduler karras
4. Fuzzy matching — difflib.SequenceMatcher with substring containment bonus (threshold 0.6)
5. Safe fallback — euler for samplers, simple for schedulers

This runs at two points:

• Execution time — resolve_sampler_scheduler() is called in the node's execute() method before passing values to the output pins
• Preset auto-population — resolve_sampler() and resolve_scheduler() are called in server.py when writing presets from CivitAI image metadata, so stored presets always contain valid names

---

File Structure

CWK_Checkpoints_Preset_Manager/
├── __init__.py                 # ComfyUI entry point — registers nodes and routes, registers .gguf extension
├── nodes.py                    # Node definitions (CWK_ModelPresetManager + CWK_InfosExtractor), preset helpers, sampler/scheduler fallback, CLIP/VAE/RNG/GGUF loaders, resolution presets, last-model persistence
├── cwk_rng_shared.py           # Self-contained RNG options (smZ_opts protocol)
├── cwk_rng_philox.py           # Vendored Philox 4×32 NVidia-compatible RNG
├── cwk_rng.py                  # prepare_noise, TorchHijack, k-diffusion hijacking
├── server.py                   # Aiohttp REST/SSE routes, CivitAI integration, preset auto-population
├── checkpoint_presets.json     # Per-model presets (auto-generated, not tracked)
├── last_used_model.json        # Last-used model name (auto-generated, not tracked)
├── hash_cache.json             # SHA-256 hash cache for model files (auto-generated)
├── model_metadata/             # Per-model CivitAI metadata cache (auto-generated)
├── local_thumbnails/           # User-uploaded local thumbnail files
├── web/
│   ├── cwk_preset_manager.js   # Main node canvas extension, LiteGraph drawing, drag-and-drop handler
│   ├── cwk_panel.js            # Model Browser panel with grid, sidebar, and management
│   ├── cwk_styles.js           # CSS injection for panel and components
│   ├── cwk_context_menu.js     # Right-click context menu, image picker, version checker
│   └── cwk_model_info.js       # Model Info overlay modal with image metadata viewer
├── CHANGELOG.md
├── LICENSE.txt
└── README.md

---

Credits

The RNG subsystem (cwk_rng.py, cwk_rng_philox.py, cwk_rng_shared.py) is adapted from ComfyUI_smZNodes by shiimizu, licensed under AGPL-3.0. These modules have been vendored and adapted to work as a self-contained integration within this node.

---

License

MIT — see LICENSE