OrakulStorm/Zero-Disk-VRAM-Streaming-Engine

Orakul Studio - Zero-Disk VRAM Streaming Engine

Important Note on RAM/VRAM & Internet Requirements

• Identical Memory Footprint: Memory consumption (System RAM & VRAM) is 100% identical whether you load a model locally or stream it via Zero-Disk.
• Zero Disk Consumption: You save 0 bytes to your local drive. No temp files, no local caches, and zero SSD/HDD write cycles or wear.
• Saves Slow Drives: If you are using a traditional HDD or have a full SSD, streaming over a fast connection completely bypasses local disk bottlenecks.
• Speed is Key: The only requirement is a fast and stable internet connection. If your system can already run Flux 2, you can stream models and LoRAs of any size (even 50GB–100GB+) seamlessly without freeing up a single gigabyte of storage!

Viking HF Multi-LoRA Loader (Zero-Disk VRAM/RAM Streaming Engine)

An official, free companion API node for my experimental and high-rank models hosted on Civitai.

Many of my experimental LoRA models are 20–40 GB in size. If you have a fast internet connection but don’t want to fill up terabytes of your SSD space just for testing, this node is your solution. It allows you to stream large model weights directly from my Hugging Face archive into your PC’s RAM/VRAM, completely bypassing the need to save files to your local disk.

Key Advantages

• Zero-Disk Architecture: Models are downloaded directly into RAM. You can test dozens of massive LoRA models without them occupying a single megabyte of space on your storage drive.
• Identical RAM Usage: The node consumes exactly the same amount of RAM as it would if you were loading a model already saved on your local disk. No extra resources are wasted.
• Smart RAM Cache: With the keep_in_ram toggle, you control the model's lifecycle:
• ON: The model is loaded into RAM once and remains there, allowing for near-instant subsequent generations.
• OFF: The model is downloaded, used to generate an image, and then immediately cleared from memory, freeing up system resources entirely.
• 4 Slots for Multi-Generation: The node features 4 independent slots, each with its own strength settings (str_model, str_clip). You can blend different models on the fly—for example, by combining a character-specific LoRA with a style-based LoRA (such as one for tattoos) in a single generation. ---

How It Works (For Users of My Models)

The node is hardcoded to point to my repository (OrakulStorm/Zero-Disk-VRAM-Streaming-Engine). This ensures everything stays configured correctly, sparing you from having to enter redundant information. The interface follows a simple "copy-paste" workflow.

Steps to follow:

1. Go to my Hugging Face repository and open the Files and versions tab.
2. Copy the path to the desired .safetensors file.

• Example of a copied path: FLUX2-RANK2048-VanGogh/r2048f2dvangog.safetensors

3. Paste this path into the filename_1 field (or any available slot in the node).
4. Ensure the lora_1_switch toggle is set to ON.
5. Click Queue Prompt — that’s it! The node will automatically fetch the required model from my repository directly into memory and generate the result.

(You do not need to change the repo_id field in the ComfyUI interface; it serves merely as a visual indicator of which repository is being accessed.)

---

👨‍💻 For Advanced Users (Using Other Repositories)

The node’s code is fully open-source. If you are familiar with Python and would like to use this "Zero-Disk" streaming engine with other models (such as official Flux weights, Diffusers, etc.), you can easily detach the node from my repository. How ​​to change the base repository:

1. Open the __init__.py file (located in the custom_nodes/VikingHF node folder) using any text editor (e.g., Notepad++).
2. Locate the class VikingHFMultiLoraLoader: definition and scroll down to the def INPUT_TYPES(s): block.
3. Within the slot settings, find the following line: "repo_id_1": ("STRING", {"default": "OrakulStorm/Zero-Disk-VRAM-Streaming-Engine"}),
4. Replace "OrakulStorm/Zero-Disk-VRAM-Streaming-Engine" with the ID of your desired Hugging Face repository.
5. Save the file and restart ComfyUI.

You can now provide paths to .safetensors files from your target repository. If the repository is private, remember to enter your access token into the hf_token field within the node's interface.