Local LLM Server Manager

Appearance

Image Generation ​

The Image Studio provides a local environment to generate high-resolution images. You can generate images through Stable Diffusion WebUI Forge or ComfyUI.

The studio supports FLUX.1, SDXL, and Stable Diffusion 1.5 base models.

Supported Model Families ​

Select the model family that matches your GPU memory and artistic goals:

Model FamilyNative ResolutionMinimum VRAMRecommended Use Case
FLUX.1 (dev / schnell)1024 × 102412 GBBest text rendering, complex prompts, and realistic anatomy.
SDXL (Stable Diffusion XL)1024 × 10248 GBPhotorealism, cinematic lighting, and custom community styles.
Stable Diffusion 1.5512 × 5124 GBHigh-speed rendering and lightweight game assets.

NOTE

You can download base models and checkpoints directly through the Model Management tab from Hugging Face or CivitAI.

Generation Procedure ​

Follow these steps to generate an image using the Fluid Studio Canvas and Creative Prompt Dock:

Step 1: Check Backend Engine Status ​

  1. Check the SD Forge card in the Telemetry Header or Ribbon.
  2. Ensure the status dot is green (Online).
  3. If stopped, click ▷ Start on the card to launch the engine on port 7860.

Step 2: Select Image Modality ​

  1. Open the Studio workspace from the left Activity Rail.
  2. In the top modality selector bar, click 🎨 Image.
  3. The Center Stage canvas activates the interactive Image Viewport with pan/zoom preview controls.

Step 3: Choose a Style Preset or Starter Prompt ​

Select a preset from the Studio Preset Bar or click a starter prompt chip directly within the canvas:

  • Standard Square (1024 × 1024): Default setting for general art and character portraits.
  • Landscape Wallpaper (1344 × 768): Wide aspect ratio for environments and desktop backgrounds.
  • Portrait Photo (768 × 1152): Vertical aspect ratio for full-body human figures and posters.
  • Classic SD 1.5 (512 × 512): Lightweight resolution for legacy models and rapid drafting.

Step 4: Enter Prompt & Fine-Tune in Dock ​

  1. Prompt Textarea: Describe your subject, lighting, and composition in the autosizing prompt dock.
  2. Parameters Flyout (⚙️ Settings): Click the settings button to slide open fine-tuning sliders:
    • Steps: 20 to 30 steps for SDXL, 4 to 8 for FLUX schnell.
    • CFG Scale: 5.0 to 7.0 for SDXL, 1.0 for FLUX.
    • Denoise: Adjust denoising strength (default: 0.75).
    • Seed: Numeric seed (-1 for randomized variations).
    • Aspect Ratio: Click 1:1, 16:9, 9:16, or 4:3 pills.
  3. Reference Attachment (📎): Optionally attach an image for image-to-image or style conditioning.

Step 5: Click Generate ↵ ​

  1. Click the prominent Generate ↵ action button in the dock.
  2. Forge WebUI processes the prompt via GPU diffusion sampling.
  3. Inspect the completed image in the interactive Center Stage canvas.

IMPORTANT

The application automatically saves rendered images to outputs/images/ with embedded prompt and parameter metadata.

Applying LoRAs (Low-Rank Adaptations) ​

LoRAs are compact model files (typically 50 MB to 200 MB). A LoRA modifies an existing checkpoint into a specific character, clothing style, or visual aesthetic.

How to Download LoRAs ​

  1. Open the Stable Diffusion (CivitAI) tab.
  2. Set the Type dropdown filter to LoRA.
  3. Sort results by Highest Rated or Most Downloaded.
  4. Type your desired aesthetic into the search bar (for example, Pixel Art XL or Detail Tweaker).
  5. Click ⬇ Download to Forge. The system stores the file in your models directory.
Style CategorySearch TermRecommended Base ModelEffect
Pixel ArtPixel Art XLSDXLGenerates authentic 16-bit sprites and backgrounds.
Retro 3DPS1 GraphicsSDXL / SD 1.5Recreates jagged low-polygon PlayStation 1 aesthetics.
Anime & CartoonCel ShadedSDXL / Pony V6Applies clean outlines and flat anime color fills.
Micro-DetailDetail Tweaker XLSDXL / FLUXSharpens skin pores, fabric weaves, and rim lighting.

Method 1: Using LoRAs in Prompts (Forge & WebUI) ​

Add the trigger tag directly into your positive prompt box:

text
<lora:pixel_art_xl:0.8> a cyberpunk street market at night, 16-bit retro game asset
  • Specify the filename of the LoRA between the colons.
  • Add a weight value at the end (typically 0.6 to 1.0).
  • Lower the weight if the style distorts facial anatomy or colors.

Method 2: Using LoRAs in ComfyUI ​

When using ComfyUI as your studio backend:

  1. Add a Load LoRA node to your ComfyUI canvas.
  2. Connect the MODEL output from your Checkpoint Loader into the model input of the LoRA node.
  3. Connect the CLIP output into the clip input of the LoRA node.
  4. Select your downloaded LoRA filename inside the node dropdown menu.
  5. Set strength_model and strength_clip to 0.8.
  6. Connect the modified outputs into your positive prompt and KSampler nodes.

VRAM Management and Hardware Fit ​

High-resolution diffusion models require dedicated GPU VRAM. Review these memory guidelines:

  • Automated LLM Offload: The VRAM Orchestrator automatically unloads Ollama models before diffusion starts.
  • FLUX Models: FLUX.1-dev requires at least 12 GB VRAM in FP8 or NF4 quantization.
  • Resolution Limits: Do not set resolutions above 1024 × 1024 on GPUs with 8 GB VRAM. Use high-resolution fix or latent upscalers instead.

WARNING

Generating at resolutions above 1536 × 1536 without tiling causes Out-Of-Memory errors on 16 GB GPUs. Use the standard presets to maintain safe VRAM bounds.

Released under the MIT License.

Copyright © Spelech