Skip to content

LM_STUDIO_SETUP.md

Latest commit

 

History

History
383 lines (298 loc) · 8.66 KB

LM_STUDIO_SETUP.md

File metadata and controls

383 lines (298 loc) · 8.66 KB

LM Studio Configuration Guide for Prompt Expansion

Quick Setup (5 Minutes)

1. Load a Good Model

Recommended models for prompt expansion:

Best (Most capable):

  • Meta-Llama-3.1-8B-Instruct - Great instruction following
  • Qwen2.5-7B-Instruct - Excellent at detailed descriptions
  • Mistral-7B-Instruct-v0.3 - Good balance of speed and quality

Good (Faster):

  • Meta-Llama-3-8B-Instruct - Solid, widely compatible
  • Phi-3-medium-4k-instruct - Good for shorter prompts

Avoid:

  • Base models (non-instruct versions)
  • Very small models (< 7B parameters)
  • Chat-tuned models optimized for conversation rather than instruction following

2. Configure Server Settings

In LM Studio, go to Local Server tab:

A. Context Length

Setting: Context Length / Max Context Recommended: 8192 or higher Why: Our prompts can be long (especially cinematic tier)

Low context (2048) = May cut off instructions ❌
Good context (8192) = Handles all tiers comfortably ✅
High context (16384+) = Overkill but works ✅

B. GPU Layers

Setting: GPU Offload / GPU Layers Recommended: As many as your GPU can handle Why: Faster generation = better experience

CPU only (0 layers) = Very slow but works
Partial GPU (20-30 layers) = Good balance
Full GPU (all layers) = Fastest ✅

C. Server Configuration

Setting: Server Port Default: 1234 Our node expects: http://localhost:1234/v1

✅ Keep default unless you have a conflict

Temperature Settings

Temperature is CRITICAL for instruction following!

In ComfyUI Node:

You control temperature in the node interface (0.1 - 2.0)

Recommended Temperatures:

Use Case Temperature Why
Maximum detail 0.3 - 0.5 LLM follows instructions precisely
Balanced 0.6 - 0.8 Good creativity + instruction following
Creative/Random 0.9 - 1.2 More unexpected choices
Wild experiments 1.3 - 2.0 Very unpredictable

For prompt expansion, start with 0.6-0.7

What Temperature Does:

Low (0.3):

  • Follows word count requirements better
  • More predictable output
  • Less creative descriptions
  • Better for specific cinematography terms

Medium (0.7):

  • Good balance
  • Still follows instructions
  • More varied language
  • Recommended starting point

High (1.2+):

  • Very creative
  • May ignore word counts
  • May not use exact technical terms
  • Good for "random" preset

Advanced Settings (Optional)

In LM Studio Local Server:

1. Repeat Penalty

Setting: Repeat Penalty / Repetition Penalty Recommended: 1.1 to 1.15 Why: Prevents repetitive descriptions

Too low (1.0) = May repeat same adjectives
Good (1.1) = Varied language ✅
Too high (1.3+) = May avoid important terms

2. Top P (Nucleus Sampling)

Setting: Top P Recommended: 0.9 to 0.95 Why: Controls randomness

Lower (0.8) = More focused
Default (0.95) = Good balance ✅
Higher (0.99) = More random

3. Top K

Setting: Top K Recommended: 40 to 50 Why: Limits vocabulary choices

Lower (20) = More predictable
Default (40-50) = Good ✅
Higher (100) = More varied

System Prompt (DO NOT SET IN LM STUDIO)

Important: Our ComfyUI node sends the system prompt automatically!

❌ Do NOT set a system prompt in LM Studio's "System Prompt" field

Why?

  • Our node sends detailed instructions already
  • LM Studio's system prompt would conflict
  • Could confuse the model

Leave LM Studio's system prompt EMPTY or use the default.

Model-Specific Tips

For Llama 3 / 3.1:

Context: 8192+
Temperature: 0.6-0.7
Repeat Penalty: 1.1
Works great with all presets ✅

For Qwen:

Context: 8192+
Temperature: 0.5-0.7
Repeat Penalty: 1.05-1.1
Excellent at descriptive detail ✅
May need lower temperature for instruction following

For Mistral:

Context: 8192+
Temperature: 0.7-0.8
Repeat Penalty: 1.1
Good balance of speed and quality ✅

For Phi-3:

Context: 4096+ (has lower context)
Temperature: 0.6
Repeat Penalty: 1.1
Faster but may struggle with cinematic tier
Best for basic/enhanced tiers

Troubleshooting

Problem: Output Too Short

Solutions:

  1. Lower temperature (try 0.4-0.5)
  2. Use better model (Llama 3.1 > Llama 3 > smaller models)
  3. Increase context length to 8192+
  4. Try different model altogether

Test:

Input: "cat playing piano"
Tier: advanced
Temperature: 0.5
Expected: 400-600 words minimum

Problem: LLM Not Following Instructions

Solutions:

  1. Lower temperature to 0.5 or less
  2. Use instruct-tuned model (must have "Instruct" in name)
  3. Check you're not setting conflicting system prompt in LM Studio
  4. Try different model

Test:

Input: "robot in city"
Preset: random
Expected: Output mentions "robot" and "city"
If not: Model not following instructions

Problem: Output Ignores Aesthetic Controls (Advanced Node)

Solutions:

  1. Lower temperature (0.4-0.6)
  2. Use Llama 3.1 or Qwen (better instruction following)
  3. Try "advanced" or "cinematic" tier
  4. Check model is instruct-tuned

Test:

Input: "spaceship flying"
Shot Size: wide shot
Lighting: edge lighting
Expected: Output must mention "wide shot" and "edge lighting"

Problem: Repeating Same Phrases

Solutions:

  1. Increase repeat penalty to 1.15-1.2
  2. Increase temperature slightly (0.7-0.8)
  3. Generate multiple variations

Problem: Wrong Format / Echoing Input

Solutions:

  1. Lower temperature (0.5)
  2. Use newer model (Llama 3.1, Qwen 2.5)
  3. Check node is using latest version with better parsing
  4. Try different model

Performance Settings

For Speed:

✅ Use quantized models (Q4_K_M or Q5_K_M)
✅ Full GPU offload
✅ Lower context (4096-8192)
✅ Batch size: 512
⚠️ Avoid F16/F32 models (slower)

For Quality:

✅ Use Q6_K or higher quantization
✅ Higher context (16384)
✅ Lower temperature (0.5-0.6)
✅ Better models (Llama 3.1, Qwen 2.5)

For Batch Generation:

✅ Use Q4_K_M quantization (faster)
✅ Temperature 0.7-0.8 (variety)
✅ Generate 3 variations per run
✅ Use wildcards for more variety

Optimal Configuration (Recommended Starting Point)

Model: Meta-Llama-3.1-8B-Instruct-Q5_K_M
Context Length: 8192
GPU Layers: Maximum available
Temperature: 0.7 (set in ComfyUI node)
Repeat Penalty: 1.1
Top P: 0.95
Top K: 40
System Prompt: EMPTY (let ComfyUI node handle it)

In ComfyUI Node:

llm_backend: lm_studio
model_name: llama3.1
api_endpoint: http://localhost:1234/v1
temperature: 0.7

Testing Your Setup

Quick Test Script:

  1. Start LM Studio with recommended settings
  2. In ComfyUI, use this test:
Input: "cat playing piano"
Preset: cinematic
Tier: advanced
Temperature: 0.6
Expected: 400-600 words, mentions cat, piano, professional cinematography terms

  1. Check output length:

    • < 200 words = Model not following instructions
    • 200-400 words = Okay, but try lower temperature
    • 400+ words = Good! ✅

  2. Check quality:

    • Mentions subject (cat, piano)? ✅
    • Uses cinematography terms? ✅
    • Flows naturally? ✅
    • Detailed descriptions? ✅

Model Download Tips

Where to Get Models:

In LM Studio:

  1. Click Search icon (🔍)
  2. Search for "Llama-3.1-8B-Instruct"
  3. Download Q5_K_M or Q4_K_M version
  4. Wait for download
  5. Load model in Local Server tab

Recommended downloads:

Meta-Llama-3.1-8B-Instruct-Q5_K_M.gguf (6GB)
or
Meta-Llama-3-8B-Instruct-Q5_K_M.gguf (6GB)

For faster generation (if GPU limited):

Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf (4.5GB)

Summary Checklist

Before using the prompt expansion node:

  • Good model loaded (Llama 3.1 or Qwen 2.5)
  • Context length: 8192+
  • GPU layers: Maximum
  • Server running on port 1234
  • System prompt: EMPTY in LM Studio
  • Temperature: 0.6-0.7 (in ComfyUI node)

Then test:

  • Generate a prompt
  • Check it's 400+ words (for advanced tier)
  • Verify it mentions your input concept
  • Confirm cinematography terms are used

Getting Help

If things still aren't working:

  1. Check LM Studio console for errors

  2. Check ComfyUI console for errors

  3. Try the absolute simplest test:

    • Input: "cat"
    • Tier: basic
    • Temperature: 0.5
    • Should get 150+ words about a cat

  4. Model issues?

    • Try different model
    • Llama 3.1 is most reliable
    • Qwen 2.5 is best for detail

  5. Still problems?

    • Check model is "Instruct" version
    • Verify server is running
    • Test with curl or another client

Quick Start: Load Llama 3.1, set context to 8192, leave system prompt empty, use temperature 0.7 in the node. That's it!