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

README.md

@@ -22,12 +22,13 @@ Scrollsmith is a Node.js tool for generating Dungeons & Dragons one-page dungeon
 
 - Node.js 22+
 - Ollama server running and accessible
-- Nextcloud (optional) for PDF uploads
+- Gitea Releases (optional) for PDF uploads
 - `.env` file with:
 
 ```env
 OLLAMA_API_URL=http://localhost:3000/api/chat/completions
 OLLAMA_API_KEY=your_api_key_here
+COMFYUI_URL=http://192.168.1.124:8188
 ````
 
 ---

compressPng.js

@@ -1,27 +0,0 @@
-import fs from 'fs/promises';
-import UPNG from 'upng-js';
-
-const countUniqueColors = (data) => {
-  const uniqueColors = new Set();
-  for (let i = 0; i < data.length; i += 4) {
-    uniqueColors.add(`${data[i]},${data[i + 1]},${data[i + 2]},${data[i + 3]}`);
-    if (uniqueColors.size > 256) {
-      return uniqueColors.size;
-    }
-  }
-  return uniqueColors.size;
-};
-
-export async function compressPng (filePath) {
-  const buffer = await fs.readFile(filePath);
-  const arrayBuffer = buffer.buffer.slice(buffer.byteOffset, buffer.byteOffset + buffer.byteLength);
-  const img = UPNG.decode(arrayBuffer);
-  const rgba = new Uint8Array(img.data);
-  const frameData = new Uint8Array(img.width * img.height * 4);
-  frameData.set(rgba.subarray(0, frameData.length));
-  const cnum = countUniqueColors(rgba) <= 256 ? 256 : 0;
-  const optimizedArrayBuffer = UPNG.encode([frameData.buffer], img.width, img.height, cnum, img.depth);
-  const optimizedBuffer = Buffer.from(optimizedArrayBuffer);
-  await fs.writeFile(filePath, optimizedBuffer);
-  return filePath;
-};

imageGenerator.js

@@ -2,29 +2,27 @@ import path from "path";
 import { mkdir, writeFile } from "fs/promises";
 import { fileURLToPath } from "url";
 import { callOllama } from "./ollamaClient.js";
-import { compressPng } from "./compressPng.js";
 
 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:
@@ -53,7 +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,
@@ -158,7 +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...");
@@ -182,9 +181,6 @@ async function generateImageViaComfyUI(prompt, filename) {
 
     console.log("Downloading image...");
     const filepath = await downloadImage(comfyFilename, filename);
-
-    console.log("Compressing PNG...");
-    await compressPng(filepath);
     return filepath;
 
   } catch (err) {

package-lock.json

@@ -10,8 +10,7 @@
       "license": "SEE LICENSE IN README.md",
       "dependencies": {
         "dotenv": "^17.2.1",
-        "puppeteer": "^24.17.1",
-        "upng-js": "^2.1.0"
+        "puppeteer": "^24.17.1"
       },
       "devDependencies": {
         "@eslint/js": "^9.34.0",
@@ -1476,12 +1475,6 @@
         "node": ">= 14"
       }
     },
-    "node_modules/pako": {
-      "version": "1.0.11",
-      "resolved": "https://registry.npmjs.org/pako/-/pako-1.0.11.tgz",
-      "integrity": "sha512-4hLB8Py4zZce5s4yd9XzopqwVv/yGNhV1Bl8NTmCq1763HeK2+EwVTv+leGeL13Dnh2wfbqowVPXCIO0z4taYw==",
-      "license": "(MIT AND Zlib)"
-    },
     "node_modules/parent-module": {
       "version": "1.0.1",
       "resolved": "https://registry.npmjs.org/parent-module/-/parent-module-1.0.1.tgz",
@@ -1878,15 +1871,6 @@
       "license": "MIT",
       "optional": true
     },
-    "node_modules/upng-js": {
-      "version": "2.1.0",
-      "resolved": "https://registry.npmjs.org/upng-js/-/upng-js-2.1.0.tgz",
-      "integrity": "sha512-d3xzZzpMP64YkjP5pr8gNyvBt7dLk/uGI67EctzDuVp4lCZyVMo0aJO6l/VDlgbInJYDY6cnClLoBp29eKWI6g==",
-      "license": "MIT",
-      "dependencies": {
-        "pako": "^1.0.5"
-      }
-    },
     "node_modules/uri-js": {
       "version": "4.4.1",
       "resolved": "https://registry.npmjs.org/uri-js/-/uri-js-4.4.1.tgz",

package.json

@@ -13,8 +13,7 @@
   "description": "",
   "dependencies": {
     "dotenv": "^17.2.1",
-    "puppeteer": "^24.17.1",
-    "upng-js": "^2.1.0"
+    "puppeteer": "^24.17.1"
   },
   "devDependencies": {
     "@eslint/js": "^9.34.0",