dphn/Dolphin3.0-Mistral-24B is the ungated mirror of the Dolphin 3.0 Mistral 24B — exactly what you asked for. It's ~48GB fp16, which needs GPU+CPU split (device_map="auto" with 32GB on GPU, ~16GB in RAM). Let me kick off the download and update the service in parallel.

tour-comfy/edit_api.py

@@ -2337,10 +2337,11 @@ def _extract_frame_at(video_path: str, t: float) -> Image.Image:
 
 
 class SceneryRequest(BaseModel):
-    model_filename: str           # person image in output_dir
-    scene_bytes: str | None = None  # base64-encoded PNG/JPEG of the reference scene
+    model_filename: str           # person image in output_dir → image2 (Picture 2)
+    scene_bytes: str | None = None  # base64-encoded PNG/JPEG of the reference scene → image1
     scene_video: str | None = None  # wireframe video name to extract frame from
     scene_time: float = 0.0       # timestamp (seconds) to extract from video
+    extra_filename: str | None = None  # optional extra reference in output_dir → image3 (Picture 3)
     prompt: str | None = None     # override; auto-built if None
     seed: int = -1
 
@@ -2361,18 +2362,19 @@ def _make_side_by_side(img1: Image.Image, img2: Image.Image,
 
 
 def _scenery_worker(job_id: str, model_filename: str, scene_pil: Image.Image,
-                    prompt: str, seed: int):
+                    prompt: str, seed: int, extra_pils: list | None = None):
     output_dir = _load_output_dir()
     try:
         model_path = os.path.join(output_dir, model_filename)
         model_pil  = Image.open(model_path).convert("RGB")
 
-        # image1=scene (→ Picture 1, output sized to scene), image2=person (→ Picture 2)
-        # The node prepends "Picture 1: <img> Picture 2: <img>" to the prompt so the
-        # model can reason about both images by name.
+        # image1=scene (→ Picture 1, output sized to scene), image2=person (→ Picture 2),
+        # image3=optional extra ref (→ Picture 3). The node prepends "Picture 1: <img>
+        # Picture 2: <img> ..." to the prompt so the model can reason about each by name.
+        extra_images = [model_pil] + list(extra_pils or [])
         png_bytes = _run_pipeline(
             scene_pil.convert("RGB"), prompt, seed, MAX_AREA,
-            extra_images=[model_pil],
+            extra_images=extra_images,
         )
         ts         = time.strftime("%Y%m%d_%H%M%S")
         base_name  = naming.get_base_name(model_filename)
@@ -2396,6 +2398,10 @@ def _scenery_worker(job_id: str, model_filename: str, scene_pil: Image.Image,
             print(f"[scenery] DB error: {db_err}")
         jobs[job_id]["status"] = "done"
         jobs[job_id]["output"] = out_name
+        # Regenerate the static JSON so the frontend's polling surfaces the new
+        # image immediately (matching _batch_worker / _multi_ref_worker). Without
+        # this the output only appeared after a server restart.
+        _invalidate_static()
     except Exception as e:
         print(f"[scenery] error: {e}")
         jobs[job_id]["status"] = "error"
@@ -2427,19 +2433,28 @@ def generate_scenery(req: SceneryRequest):
     else:
         raise HTTPException(400, "Provide scene_bytes or scene_video")
 
+    # Optional image3 (Picture 3) — extra reference from output_dir
+    extra_pils = []
+    if req.extra_filename:
+        extra_path = os.path.join(output_dir, req.extra_filename)
+        if not os.path.exists(extra_path):
+            raise HTTPException(404, f"Extra reference not found: {req.extra_filename}")
+        extra_pils.append(Image.open(extra_path).convert("RGB"))
+
     prompt = req.prompt or (
         "Place the person from Picture 2 naturally inside the environment shown in Picture 1. "
         "Keep the person's face, body proportions, clothing and pose exactly as in Picture 2. "
         "Use the location, lighting and atmosphere from Picture 1 as the background. "
         "Match the color temperature and shadows so it looks like one photograph taken on location. "
-        "Output a single photorealistic image. High quality, detailed."
+        + ("Incorporate the reference from Picture 3 as well. " if extra_pils else "")
+        + "Output a single photorealistic image. High quality, detailed."
     )
 
     job_id = uuid.uuid4().hex[:8]
     jobs[job_id] = {"status": "running", "type": "scenery", "total": 1, "done": 0, "failed": 0}
     threading.Thread(
         target=_scenery_worker,
-        args=(job_id, req.model_filename, scene_pil, prompt, req.seed),
+        args=(job_id, req.model_filename, scene_pil, prompt, req.seed, extra_pils),
         daemon=True,
     ).start()
     return {"job_id": job_id, "model": req.model_filename}