| Piece | What it does | |---|---| | Alchemist | Prompt + LoRA-caption nodes for Krea 2 and MiniMax H3, any LLM provider | | Prompt Rotate | Paste a prompt list → STRING list socket, or queue one job per prompt | | Model Hub | One window, two tabs: Civitai and Hugging Face — browse, filter and download straight into the right models/ folder |

Sources bundled from:

https://github.com/babydjac/ComfyUI-Alchemist
https://github.com/babydjac/ComfyUI-PromptRotate
https://github.com/babydjac/ComfyUI_HF_ModelDownloader

Install

ComfyUI-Manager / Registry

Search crypswolf69 and install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/babydjac/crypswolf69.git
pip install -r crypswolf69/requirements.txt

Restart ComfyUI and hard-refresh the browser.

Hugging Face downloads use aria2c. If it's missing, click Install aria2 in the Model Hub header (Homebrew, winget, Chocolatey, Scoop, apt-get, dnf/yum, pacman or apk — whichever the machine has). Civitai downloads don't need it.

Model Hub

Open it with the draggable Model Hub launcher, Alt+M (Option+M on a Mac), the Model Hub sidebar tab, or the command palette. Ctrl/Cmd+Shift+B opens it straight on the Hugging Face tab. The header shows the aria2 status on both tabs — a green ✓ when it's installed, an Install aria2 button when it isn't.

Civitai tab

Browse — masonry grid with fast CDN thumbnails and hover-to-play video previews (H3, Wan…), infinite scroll, comfortable / compact density.
Filter — model type, base model (Krea 2, MiniMax H3, Flux.2, Illustrious… — the list grows as you browse, and any value can be typed), sort, period, NSFW (sensitive previews stay blurred until you choose to show them).
Details — version picker, media viewer + full-screen lightbox (← →), click-to-copy trigger words, description, and every file with size / format / precision.
Download — per-file Save to picker with a sensible default (Krea 2 / Flux / Wan / Qwen checkpoints and any GGUF → diffusion_models), sorted into base-model subfolders (loras/Krea 2/…, loras/MiniMax H3/…). Buttons show live progress; "owned" files are marked.
Paste link — a Civitai model page, Civitai download link, or Hugging Face file link; name, type and folder are worked out for you.
Downloader — up to 3 files at once, 8 parallel segments each, automatic retry and resume, never writes outside models/, tokens are only ever sent to their own site.
⚙ Settings — Civitai API token, Hugging Face token, connections, segment size, subfolder sorting, aria2 status / install.

Keyboard: / search · Esc closes the top-most panel · ← → in the viewer.

Hugging Face tab

Browse — curated index from trusted owners, tabbed by category (Diffusion Models, Checkpoints, Text Encoders, VAE, LoRAs, ControlNet, Upscale Models…), installed files marked.
Live HF — search the whole Hub; quick chips (Upscalers, ESRGAN, Krea 2, ControlNet, LoRAs, VAE, Encoders, GGUF); pick a repo, tick files (the destination folder is shown on each), Download N from the files header.
Downloads — per-job and per-file progress with cancel.
Compact — a denser layout for small screens (remembered).

Keyboard: / search · ↑ ↓ move through repos / files · → / ← jump between repos and files · Enter open a repo · Space select a file · Esc clears the search, then closes the hub.

✦ Advanced LoRA Loader

One node for a whole LoRA stack (CrypsAdvancedLoraLoader, in model/loaders).

Rows — on/off switch, preview, name, strength pill (drag to scrub, click to type, ‹ › ±0.05, Shift for ±0.01, Alt-click resets), drag handle to reorder.
Picker — search by name, trigger word or base model; folder chips, Recent, LoRA Manager favorites; image/video previews; ↑ ↓ Enter, Shift adds several.
ⓘ panel — Civitai page (.com / .red), creator, example images with prompts, file facts (LoRA/LoKr, rank, layers, training steps). Click trigger words to send them to the trigger_words output.
CLIP — optional input; the T toggle gives each LoRA its own CLIP strength.
Audio-video models (LTX-2.3, MiniMax H3) — LoRAs with audio layers get V× / A× multipliers: audio keys get STR × A×, everything else STR × V×.
Status after a run — ✓ patched, ◐ partly matched, ⚠ matched nothing (LoRA for another base model), ✖ missing, ↪ found in another loras subfolder (click to save the new path).
Outputs — MODEL, CLIP, trigger_words, lora_tags (<lora:name:strength>).
⚡ keeps LoRA files in RAM between runs (capped at 15% of RAM); compact rows and six themes are in the header.

Info comes from LoRA Manager / Civitai Helper / A1111 sidecars when present, otherwise the file is hashed once (off the UI thread) and looked up on Civitai; results are cached in lora_loader/.cache/. Missing LoRAs stop the queue with their names, and DaSiWa stack_data JSON can be pasted into the stack widget.

Nodes

AlchemistKrea2 / AlchemistLoraCaption / AlchemistH3 / AlchemistPrompt / AlchemistDatasetCaptioner
PromptRotate / PromptRotatePick
CrypsAdvancedLoraLoader (✦ ADVANCED LORA LOADER)
The Model Hub is a UI panel (no graph nodes)

Existing workflows that used the standalone packs keep the same class names.

License

MIT. See LICENSE.