Apex Pipeline
Four Apex nodes for Krea 2 generation, LoRA control, metadata-aware saves and a resizable image preview.
Nodes (7)
Apex Pipeline V2
Apex Generator V2 keeps the usual Krea 2 image workflow in four Apex nodes. The controls, saved output and large preview stay on one clear row, so you can focus on the image instead of wiring a long chain of small utility nodes.
The four nodes
- Apex Model Matrix loads the diffusion model, text encoder and VAE. It uses the installed LC123 optimizer when available.
- Apex Generation Nexus holds the prompt, canvas, seed, sampler and optional LoRAs. Search local LoRA names, then choose a result from the dropdown. The stack supports up to 64 entries, with independent model and CLIP strengths, reordering, duplicate and remove controls, and a clear per-LoRA ON/OFF switch. Changes are applied on the next queued run. Strength has a −5 to +5 slider, 0.05 step buttons and a number field for any finite value.
- Apex Archive saves the image and embeds workflow, generation settings and resource hashes through LC123. It includes the run timer and one adjustable Last save preview.
- Apex Image Preview shows the generated image on its own. Drag its lower-right corner to resize it.
The Find triggers button checks selected LoRA metadata on your computer first. If that has no trigger words, it can look up the file's SHA-256 against Civitai's public model index. The LoRA file is not uploaded. Results appear as buttons you can click to copy. Civitai may not have a match for every file.
The workflow uses ComfyUI's normal model, image and widget data. The Apex surfaces only change how the controls look and are arranged. The sampler, model loading, decode and save paths remain in the node execution graph.
Missing-node pack warning
The V2.0.7 starter writes the Comfy Registry package identity (apex-pipeline) onto every Apex node. This lets ComfyUI group missing Apex nodes under the correct pack instead of showing Unknown pack. If you already downloaded an older workflow copy, download the refreshed JSON below. If ComfyUI does not offer Registry installation yet, install Apex Pipeline from this repository using the steps above; Comfy Registry may still be reviewing the newest version.
Install
-
Install ComfyUI_LC123_nodes with ComfyUI Manager, or follow its instructions at the LC123 GitHub page. Apex uses LC123 for Speed Boost and metadata-aware image saving; the model optimizer uses LC123 when available and has a native fallback.
-
Install Apex Pipeline with ComfyUI Manager, or use a terminal in your active
ComfyUI/custom_nodesfolder:git clone https://github.com/ApexLightSpeed/ComfyUI_ApexPipeline.gitIf
ComfyUI_ApexPipelineis already there, open that folder in a terminal and rungit pullinstead of cloning a second copy. -
Restart ComfyUI. If you added or updated model files, restart once more so the pickers see them.
-
Open
workflows/Apex Generator V2.jsonin ComfyUI.
For ComfyUI Desktop, use the custom_nodes folder belonging to the installation you actually run. Avoid keeping duplicate Apex Pipeline folders active at the same time.
Model files
Use model files compatible with Comfy-Org's Krea 2 release, and follow that repository's model license. The standard Turbo setup uses:
| File | ComfyUI folder |
| --- | --- |
| krea2_turbo_fp8_scaled.safetensors | models/diffusion_models/ |
| qwen3vl_4b_fp8_scaled.safetensors | models/text_encoders/ |
| qwen_image_vae.safetensors | models/vae/ |
You can pick other compatible files in Apex Model Matrix. The supplied workflow starts with a blank prompt and no LoRAs, so it will not ask you to download someone else's LoRA collection.
Model, text encoder and VAE dropdowns list files installed in ComfyUI, including files inside subfolders. Use the search field to filter names or paths across folders, or choose a folder in the dropdown to browse it. The LoRA search works the same way: it filters the Add LoRA and stack-row dropdowns across all local LoRA folders. These controls do not offer remote downloads or recommended files.
The starter workflow does not save model filenames or LoRA selections. Choose files from your own ComfyUI model folders after opening it.
First run
- In Apex Model Matrix, choose your Krea 2 model, text encoder and VAE.
- Type a prompt in Apex Generation Nexus. The starter uses 896 × 1344, batch 1, a random seed, 8 steps, CFG 1, Euler/simple and LC Speed Boost.
- Add LoRAs only if you need them. Adjust model and CLIP strengths independently; use Find triggers for an optional trigger-word lookup.
- Choose the filename, folder, image format and metadata options in Apex Archive. PNGs automatically include the Apex workflow resource reference. The separate Embed workflow checkbox remains visible and on by default to embed the ComfyUI graph.
- Queue the prompt. The Archive reports the save and shows a small preview; Apex Image Preview shows the same image larger.
Notes
- ComfyUI_LC123_nodes is required. Apex calls its installed optimizer and save nodes; their source code is not copied into this repository.
- Pixaroma is not required by the V2 graph.
- The Apex preview is separate from saving: Archive writes the file to ComfyUI's normal
output/Apexfolder, and Image Preview displays the generated image. Firefox uploaded from that shared folder during validation; Opera GX showed a local file-access warning. - Find triggers runs only when you press its button. Its public lookup sends a file hash and LoRA name to Civitai, never the LoRA weights.
- If something fails, include your ComfyUI version and the complete console traceback in a GitHub issue. Do not include API keys or access tokens.
License
Apex Pipeline is MIT licensed; see LICENSE. LC123 and the model files retain their own licenses. This repository does not redistribute their code or model weights.
Publishing a Registry version
The Registry package is apex-pipeline under publisher apexlightspeed. To publish, create a Registry personal access token, add it to this GitHub repository as the Actions secret REGISTRY_ACCESS_TOKEN, then run Actions → Publish to Comfy Registry. The action reads the version from pyproject.toml; published versions cannot be overwritten, so increase that version for each release.