Quality fix: majicMIX model, Euler a, CLIP skip 2, 512x768 default, stronger negatives (#8)

- Quality fix: majicMIX model, Euler a, CLIP skip 2, 512x768 default, stronger negatives (903bdd98b974777129df83dab9891a00b1634498)


Co-authored-by: pormungtailaw <pormungtai@users.noreply.huggingface.co>

app.py

@@ -109,8 +109,7 @@ def _apply_model(models, model_id):
     """Shared: given a selected model id, return the dependent UI updates."""
     cfg = pm.get_model(models, model_id)
     if not cfg:
-        return (gr.update(), gr.update(), gr.update(), gr.update(), gr.update(),
-                "ไม่พบโมเดล / model not found")
+        return (*[gr.update() for _ in range(7)], "ไม่พบโมเดล / model not found")
     choices = modes_for(models, model_id)
     return (
         gr.update(choices=choices, value=choices[0][1]),           # mode radio
@@ -118,6 +117,8 @@ def _apply_model(models, model_id):
         gr.update(value=cfg.get("negative_prompt", "")),           # negative
         gr.update(value=cfg.get("default_steps", 28)),             # steps
         gr.update(value=cfg.get("default_guidance", 6.0)),         # guidance
+        gr.update(value=cfg.get("default_width", 768)),            # width
+        gr.update(value=cfg.get("default_height", 768)),           # height
         f"**เลือก / Selected:** {cfg['label']}",                   # selected label
     )
 
@@ -208,7 +209,7 @@ with gr.Blocks(css=CSS, theme=gr.themes.Soft(primary_hue="blue"),
                     steps = gr.Slider(1, 50, value=28, step=1, label="Steps")
                     guidance = gr.Slider(0.0, 15.0, value=6.5, step=0.1, label="Guidance (CFG)")
                 with gr.Row():
-                    width = gr.Slider(384, 1280, value=768, step=64, label="Width")
+                    width = gr.Slider(384, 1280, value=512, step=64, label="Width")
                     height = gr.Slider(384, 1280, value=768, step=64, label="Height")
                 with gr.Row():
                     seed = gr.Number(value=-1, label="Seed (-1 = random)", precision=0)
@@ -218,7 +219,7 @@ with gr.Blocks(css=CSS, theme=gr.themes.Soft(primary_hue="blue"),
     model_gallery.select(
         on_gallery_select, inputs=[ids_state],
         outputs=[selected_id, mode_radio, prompt, negative_prompt,
-                 steps, guidance, selected_md],
+                 steps, guidance, width, height, selected_md],
     )
     reload_btn.click(
         reload_registry,

models.json

@@ -1,7 +1,25 @@
 {
-  "_comment": "Editable model registry. Add/remove entries freely, then restart the Space (or click 🔄 Reload models). Each entry may include a `preview` image URL shown in the picker. See README_TH.md for the field guide.",
+  "_comment": "Editable model registry. Add/remove entries freely, then restart (or click 🔄 Reload models). `clip_skip` uses A1111 convention (2 = skip one layer, common for SD1.5 community models). `default_width`/`default_height` set the canvas when the model is picked. See README_TH.md.",
 
   "models": [
+    {
+      "id": "majicmix-realistic-v7",
+      "label": "majicMIX Realistic V7 (SD1.5)",
+      "base": "sd15",
+      "type": "checkpoint",
+      "repo_id": "digiplay/majicMIX_realistic_v7",
+      "single_file_url": null,
+      "preview": "previews/majicmix-realistic-v7.webp",
+      "trigger": "",
+      "clip_skip": 2,
+      "default_width": 512,
+      "default_height": 768,
+      "recommended_prompt": "RAW photo, portrait of a beautiful woman, detailed skin, soft cinematic light, 50mm, depth of field, ultra detailed",
+      "negative_prompt": "(worst quality, low quality, normal quality:1.4), lowres, bad anatomy, bad hands, extra digits, fewer digits, deformed, disfigured, mutated, extra limbs, duplicate, cropped, watermark, text, signature, jpeg artifacts",
+      "default_steps": 28,
+      "default_guidance": 7.0,
+      "enabled": true
+    },
     {
       "id": "sd15-realistic-base",
       "label": "SD1.5 · Realistic Base",
@@ -11,8 +29,11 @@
       "single_file_url": null,
       "preview": "previews/sd15-realistic-base.webp",
       "trigger": "",
+      "clip_skip": 1,
+      "default_width": 512,
+      "default_height": 768,
       "recommended_prompt": "RAW photo, a beautiful woman, detailed skin, soft lighting, 50mm, depth of field",
-      "negative_prompt": "(worst quality, low quality:1.4), deformed, extra fingers, watermark, text",
+      "negative_prompt": "(worst quality, low quality:1.4), lowres, bad anatomy, bad hands, extra digits, deformed, duplicate, watermark, text",
       "default_steps": 28,
       "default_guidance": 6.5,
       "enabled": true
@@ -29,8 +50,11 @@
       "lora_scale": 0.8,
       "preview": "https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/031e4408-3098-4f4a-9a05-d499cedf513a/width=450/747945.jpeg",
       "trigger": "",
+      "clip_skip": 2,
+      "default_width": 512,
+      "default_height": 768,
       "recommended_prompt": "RAW photo, asian girl, pretty face, natural skin texture, cinematic light",
-      "negative_prompt": "(worst quality, low quality:1.4), deformed, watermark, text",
+      "negative_prompt": "(worst quality, low quality:1.4), lowres, bad anatomy, bad hands, deformed, duplicate, watermark, text",
       "default_steps": 28,
       "default_guidance": 6.5,
       "enabled": true
@@ -44,8 +68,11 @@
       "single_file_url": "https://civitai.com/api/download/models/127016",
       "preview": "https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9c67da8b-5e57-413c-a5ca-135f5ed9af18/width=450/1743066.jpeg",
       "trigger": "",
+      "clip_skip": 2,
+      "default_width": 512,
+      "default_height": 768,
       "recommended_prompt": "RAW photo, beautiful asian woman, elegant dress, soft cinematic light, ultra detailed",
-      "negative_prompt": "(worst quality, low quality:1.4), deformed, watermark, text",
+      "negative_prompt": "(worst quality, low quality:1.4), lowres, bad anatomy, bad hands, deformed, duplicate, watermark, text",
       "default_steps": 28,
       "default_guidance": 6.0,
       "enabled": true
@@ -59,6 +86,9 @@
       "single_file_url": null,
       "preview": "previews/sdxl-base.webp",
       "trigger": "",
+      "clip_skip": 1,
+      "default_width": 1024,
+      "default_height": 1024,
       "recommended_prompt": "cinematic photo of a beautiful woman, 35mm, highly detailed, soft natural light",
       "negative_prompt": "lowres, bad anatomy, worst quality, watermark, text",
       "default_steps": 30,
@@ -74,6 +104,9 @@
       "single_file_url": null,
       "preview": "previews/flux-schnell.webp",
       "trigger": "",
+      "clip_skip": 1,
+      "default_width": 1024,
+      "default_height": 1024,
       "recommended_prompt": "a photorealistic portrait of a young woman, studio light, sharp focus, ultra detailed",
       "negative_prompt": "",
       "default_steps": 4,
@@ -87,8 +120,11 @@
       "type": "checkpoint",
       "repo_id": "black-forest-labs/FLUX.1-dev",
       "single_file_url": null,
-      "preview": "https://placehold.co/400x520/1e293b/93c5fd/png?text=FLUX.1%20Dev",
+      "preview": "previews/flux-schnell.webp",
       "trigger": "",
+      "clip_skip": 1,
+      "default_width": 1024,
+      "default_height": 1024,
       "recommended_prompt": "a photorealistic portrait of a young woman, golden hour, 85mm, bokeh, ultra detailed",
       "negative_prompt": "",
       "default_steps": 25,

pipeline_manager.py

@@ -165,6 +165,12 @@ def _build_base_pipeline(cfg):
         except Exception as e:  # noqa
             print(f"[lora] fuse skipped: {e}")
 
+    # SD1.5 / SDXL community checkpoints are tuned for the Euler Ancestral sampler;
+    # it matches the look people get in A1111 / ComfyUI far better than the default.
+    if base in ("sd15", "sdxl"):
+        from diffusers import EulerAncestralDiscreteScheduler
+        pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)
+
     pipe.set_progress_bar_config(disable=True)
     pipe._cs_adapter = None  # track loaded IP-Adapter / FaceID state
     return pipe
@@ -284,6 +290,11 @@ def run_generation(cfg, mode, prompt, negative_prompt, ref_image,
     else:
         call["guidance_scale"] = float(guidance)
         call["negative_prompt"] = negative_prompt or None
+        # CLIP skip: registry stores the A1111 value (2 = skip one layer); diffusers
+        # counts skipped layers, so A1111 "2" == diffusers clip_skip=1.
+        a1111_cs = int(cfg.get("clip_skip", 1))
+        if a1111_cs > 1:
+            call["clip_skip"] = a1111_cs - 1
 
     # ----- mode wiring -----
     if mode == "txt2img":