improve image gen prompting. increase cfg from 1->3 to make prompt follow more aggressively

imageGenerator.js

@@ -7,23 +7,22 @@ const __dirname = path.dirname(fileURLToPath(import.meta.url));
 const COMFYUI_URL = process.env.COMFYUI_URL || "http://localhost:8188";
 
 // Drawing style prefix
-const STYLE_PREFIX = `a high-contrast, black and white pen and ink drawing, hand-drawn sketch aesthetic, very low detail, extremely minimal, visible loose linework, expressive simple hatching for shadows, quick conceptual sketch, subtle color accent`;
+const STYLE_PREFIX = `clean line art, minimalist sketch, concept art sketch, black and white line drawing, lots of white space, sparse shading, simple hatching, very low detail, subtle color accent`;
 
 // 1. Generate engineered visual prompt
 async function generateVisualPrompt(flavor) {
   const rawPrompt = await callOllama(
-    `You are a prompt engineer specializing in visual prompts for AI image generation (e.g., Stable Diffusion, ComfyUI). Given a piece of fantasy flavor text, your job is to extract and translate the visual elements into a highly descriptive image prompt.
+    `You are a prompt engineer specializing in visual prompts for AI image generation. Your goal is to translate fantasy flavor text into a sparse, minimalist scene description.
 
-Your output should be structured like a list of visual tags, not a story or paragraph. Focus on describing the environment, mood, architecture, lighting, materials, and color. Avoid abstract, emotional, and visual language. Be literal, specific, and visual only.
+Your output must be a simple list of visual tags describing only the most essential elements of the scene. Focus on the core subject and mood.
 
 Rules:
-- Do NOT repeat phrases or wording from the input.
-- Only include things that could be seen in a single, still image.
-- Use visual keywords, rich adjectives, and clear scene descriptors.
-- Keep the prompt concise, 40-80 words.
-- Maintain simplicity in descriptions. The image must be a minimal, hand drawn sketch aesthetic with low detail.
-- Avoid characters or creatures unless clearly described.
-- Avoid referencing rendering style, color technique, camera effects, or drawing medium — focus only on the visual content of the scene.
+- Describe a sparse scene with a single focal point or area.
+- Use only 3-5 key descriptive phrases or tags.
+- The entire output should be very short, 20-50 words maximum.
+- Do NOT repeat wording from the input.
+- Focus only on visual content, not style, medium, or camera effects.
+- Avoid describing fine details; focus on large forms and the overall impression.
 - Do NOT include phrases like “an image of” or “a scene showing”.
 
 Input:
@@ -52,7 +51,7 @@ function buildComfyWorkflow(promptText, negativeText = "") {
       "inputs": {
         "seed": Math.floor(Math.random() * 100000),
         "steps": 4,
-        "cfg": 1,
+        "cfg": 3,
         "sampler_name": "euler",
         "scheduler": "simple",
         "denoise": 1,
@@ -157,7 +156,8 @@ async function downloadImage(filename, localFilename) {
 
 // 4c. Submit prompt and handle full image pipeline
 async function generateImageViaComfyUI(prompt, filename) {
-  const workflow = buildComfyWorkflow(prompt, "photorealism, hyper-realistic, high detail, 3D render, CGI, ray tracing, glossy, polished, smooth shading, realistic lighting, digital painting, anime, cartoon, pixelated, noisy, cluttered, blurry, sharp focus, symmetrical, perfect perspective, detailed textures, high-resolution textures, high contrast lighting, lens flare, bokeh, camera artifacts, text, logo, signature, watermark, overexposed, underexposed, glowing edges");
+  const negativePrompt = `heavy shading, deep blacks, cross-hatching, dark, gritty, shadow-filled, chiaroscuro, scratchy lines, photorealism, hyper-realistic, high detail, 3D render, CGI, polished, smooth shading, detailed textures, noisy, cluttered, blurry, text, logo, signature, watermark`;
+  const workflow = buildComfyWorkflow(prompt, negativePrompt);
 
   try {
     console.log("Submitting prompt to ComfyUI...");