feat(diffusion): add Qwen-Image support with true_cfg_scale parameter

**Problem:**
Qwen-Image models from Alibaba use a different parameter name (true_cfg_scale)
instead of the standard guidance_scale used by other diffusion models. Users
could not properly configure Qwen-Image models.

**Solution:**
- Added true_cfg_scale parameter to generate_image_diffusers.py
- Detects Qwen/QwenImage pipeline types and uses appropriate parameter
- Updated Swift PythonDiffusersService to pass trueCfgScale
- Backward compatible: falls back to guidance_scale if true_cfg_scale not specified
- Python script already had QwenImagePipeline support via dynamic detection

**Changes:**
1. scripts/generate_image_diffusers.py:
   - Added true_cfg_scale parameter to generateImage() function
   - Added model type detection (is_qwen) to choose correct parameter
   - Build guidance_kwargs dict with either true_cfg_scale or guidance_scale
   - Updated all pipeline calls to use **guidance_kwargs
   - Added --true-cfg-scale command-line argument

2. Sources/StableDiffusionIntegration/PythonDiffusersService.swift:
   - Added trueCfgScale: Float? parameter to generateImage()
   - Added logging for trueCfgScale value
   - Pass --true-cfg-scale to Python script when provided

**Testing:**
✅ Build: PASS
✅ Python syntax: PASS
✅ Backward compatible: existing models continue to work

**Notes:**
- Qwen-Image recommended settings: 50 steps, true_cfg_scale 4.0
- Models already supported via HuggingFace browser in preferences
- UI can search/download Qwen/Qwen-Image model directly

Author: fewtarius

Sources/StableDiffusionIntegration/PythonDiffusersService.swift

@@ -176,7 +176,8 @@ public class PythonDiffusersService {
     ///   - modelName: Name of model in staging directory
     ///   - scheduler: Scheduler to use
     ///   - steps: Number of inference steps
-    ///   - guidanceScale: Guidance scale
+    ///   - guidanceScale: Guidance scale (standard SD models)
+    ///   - trueCfgScale: True CFG scale (Qwen-Image and similar models, overrides guidanceScale)
     ///   - width: Image width
     ///   - height: Image height
     ///   - seed: Random seed (nil for random)
@@ -195,6 +196,7 @@ public class PythonDiffusersService {
         scheduler: PythonScheduler = .dpmppSDEKarras,
         steps: Int = 25,
         guidanceScale: Float = 7.5,
+        trueCfgScale: Float? = nil,
         width: Int = 512,
         height: Int = 512,
         seed: Int? = nil,
@@ -225,6 +227,7 @@ public class PythonDiffusersService {
             "scheduler": .string(scheduler.rawValue),
             "steps": .stringConvertible(steps),
             "guidance": .stringConvertible(guidanceScale),
+            "trueCfgScale": trueCfgScale != nil ? .stringConvertible(trueCfgScale!) : .string("nil"),
             "size": .string("\(width)×\(height)"),
             "mode": .string(inputImage != nil ? "img2img" : "txt2img")
         ])
@@ -243,6 +246,11 @@ public class PythonDiffusersService {
             "--num-images", String(imageCount)
         ]
 
+        /// Add true-cfg-scale if provided (for Qwen-Image and similar models)
+        if let trueCfg = trueCfgScale {
+            args.append(contentsOf: ["--true-cfg-scale", String(trueCfg)])
+        }
+
         if let negPrompt = negativePrompt, !negPrompt.isEmpty {
             args.append(contentsOf: ["-n", negPrompt])
         }

scripts/generate_image_diffusers.py

@@ -455,6 +455,7 @@ def generate_image(
     scheduler: str = "dpm++_sde_karras",
     steps: int = 25,
     guidance_scale: float = 7.5,
+    true_cfg_scale: Optional[float] = None,
     width: int = 512,
     height: int = 512,
     seed: Optional[int] = None,
@@ -475,7 +476,8 @@ def generate_image(
         negative_prompt: Negative prompt
         scheduler: Scheduler name (dpm++_sde_karras, euler, etc.)
         steps: Number of inference steps
-        guidance_scale: Guidance scale
+        guidance_scale: Guidance scale (standard SD models)
+        true_cfg_scale: True CFG scale (Qwen-Image and similar models, overrides guidance_scale)
         width: Image width (text-to-image only)
         height: Image height (text-to-image only)
         seed: Random seed (None for random)
@@ -862,6 +864,23 @@ def generate_image(
                 prompt_embeds, negative_embeds = compel_result
                 print("Using Compel embeddings for SD 1.x")
 
+    # Determine which guidance parameter to use
+    # Qwen-Image uses true_cfg_scale, standard SD uses guidance_scale
+    is_qwen = model_type.lower() in ["qwenimage", "qwen"]
+    
+    # Use true_cfg_scale if provided, otherwise use guidance_scale for both
+    effective_cfg_scale = true_cfg_scale if true_cfg_scale is not None else guidance_scale
+    
+    # Build kwargs for pipeline call
+    guidance_kwargs = {}
+    if is_qwen:
+        # Qwen-Image uses true_cfg_scale parameter
+        guidance_kwargs['true_cfg_scale'] = effective_cfg_scale
+        print(f"  True CFG: {effective_cfg_scale} (Qwen-Image mode)")
+    else:
+        # Standard SD models use guidance_scale
+        guidance_kwargs['guidance_scale'] = guidance_scale
+    
     # Generate image
     print(f"Generating {num_images} image(s)...")
     print(f"  Mode: {'Image-to-Image' if is_img2img else 'Text-to-Image'}")
@@ -873,7 +892,8 @@ def generate_image(
     else:
         print(f"  Resolution: {width}×{height}")
     print(f"  Steps: {steps}")
-    print(f"  Guidance: {guidance_scale}")
+    if not is_qwen:
+        print(f"  Guidance: {guidance_scale}")
     if use_compel and prompt_embeds is not None:
         print(f"  Compel: ENABLED (prompt weighting active)")
 
@@ -886,9 +906,9 @@ def generate_image(
                     image=init_img,
                     strength=strength,
                     num_inference_steps=steps,
-                    guidance_scale=guidance_scale,
                     num_images_per_prompt=num_images,
                     generator=generator,
+                    **guidance_kwargs,
                     **prompt_embeds
                 )
             elif prompt_embeds is not None:
@@ -899,9 +919,9 @@ def generate_image(
                     prompt_embeds=prompt_embeds,
                     negative_prompt_embeds=negative_embeds,
                     num_inference_steps=steps,
-                    guidance_scale=guidance_scale,
                     num_images_per_prompt=num_images,
-                    generator=generator
+                    generator=generator,
+                    **guidance_kwargs
                 )
             else:
                 # Standard prompts
@@ -911,21 +931,21 @@ def generate_image(
                     strength=strength,
                     negative_prompt=negative_prompt if negative_prompt else None,
                     num_inference_steps=steps,
-                    guidance_scale=guidance_scale,
                     num_images_per_prompt=num_images,
-                    generator=generator
+                    generator=generator,
+                    **guidance_kwargs
                 )
         else:
             # Text-to-Image generation
             if isinstance(prompt_embeds, dict):
                 # SDXL with Compel
                 result = pipe(
                     num_inference_steps=steps,
-                    guidance_scale=guidance_scale,
                     width=width,
                     height=height,
                     num_images_per_prompt=num_images,
                     generator=generator,
+                    **guidance_kwargs,
                     **prompt_embeds
                 )
             elif prompt_embeds is not None:
@@ -934,23 +954,23 @@ def generate_image(
                     prompt_embeds=prompt_embeds,
                     negative_prompt_embeds=negative_embeds,
                     num_inference_steps=steps,
-                    guidance_scale=guidance_scale,
                     width=width,
                     height=height,
                     num_images_per_prompt=num_images,
-                    generator=generator
+                    generator=generator,
+                    **guidance_kwargs
                 )
             else:
                 # Standard prompts
                 result = pipe(
                     prompt=prompt,
                     negative_prompt=negative_prompt if negative_prompt else None,
                     num_inference_steps=steps,
-                    guidance_scale=guidance_scale,
                     width=width,
                     height=height,
                     num_images_per_prompt=num_images,
-                    generator=generator
+                    generator=generator,
+                    **guidance_kwargs
                 )
 
     images = result.images
@@ -1043,6 +1063,7 @@ def main():
     )
     parser.add_argument('--steps', type=int, default=25, help='Number of inference steps (default: 25)')
     parser.add_argument('--guidance', type=float, default=7.5, help='Guidance scale (default: 7.5)')
+    parser.add_argument('--true-cfg-scale', type=float, default=None, help='True CFG scale for Qwen-Image (overrides --guidance)')
     parser.add_argument('--width', type=int, default=512, help='Image width (default: 512)')
     parser.add_argument('--height', type=int, default=512, help='Image height (default: 512)')
     parser.add_argument('--seed', type=int, default=None, help='Random seed (default: random)')
@@ -1064,6 +1085,7 @@ def main():
             scheduler=args.scheduler,
             steps=args.steps,
             guidance_scale=args.guidance,
+            true_cfg_scale=getattr(args, 'true_cfg_scale', None),
             width=args.width,
             height=args.height,
             seed=args.seed,