Reduce Ollama analysis image context

ltx_director.py

@@ -482,6 +482,30 @@ def _analysis_text_is_usable(text: str) -> bool:
     return True
 
 
+def _compress_analysis_image_b64(b64_payload: str, max_dim: int = 768, quality: int = 82) -> str:
+    """Shrink analysis images so multimodal providers do not burn their full context on pixels."""
+    try:
+        raw = base64.b64decode(_normalise_b64_payload(b64_payload))
+        with Image.open(_io.BytesIO(raw)) as img:
+            img = img.convert("RGB")
+            w, h = img.size
+            if max(w, h) > max_dim:
+                scale = max_dim / float(max(w, h))
+                img = img.resize((max(1, int(round(w * scale))), max(1, int(round(h * scale)))), Image.LANCZOS)
+            out = _io.BytesIO()
+            img.save(out, format="JPEG", quality=quality, optimize=True)
+            return base64.b64encode(out.getvalue()).decode("ascii")
+    except Exception:
+        return _normalise_b64_payload(b64_payload)
+
+
+def _prepare_analysis_images(cleaned_b64_list: list[str], max_dim: int = 768, quality: int = 82) -> list[str]:
+    return [
+        _compress_analysis_image_b64(b64, max_dim=max_dim, quality=quality)
+        for b64 in cleaned_b64_list
+    ]
+
+
 def _resolve_provider(data):
     provider = (data.get("provider") or "ollama").lower()
     defs = _PROVIDER_DEFAULTS.get(provider, _PROVIDER_DEFAULTS["ollama"])
@@ -513,6 +537,7 @@ async def analyze_character_endpoint(request):
             cleaned_b64_list.append(b64)
         if not cleaned_b64_list:
             return web.json_response({"status": "error", "message": "No valid base64 images decoded."})
+        analysis_images = _prepare_analysis_images(cleaned_b64_list, max_dim=768, quality=82)
 
         if provider in ("lmstudio", "custom") and not model_name:
             return web.json_response({
@@ -526,10 +551,11 @@ async def analyze_character_endpoint(request):
         try:
             async with aiohttp.ClientSession() as session:
                 if provider == "ollama":
+                    generated_text = ""
                     payload = {
                         "model": model_name,
                         "prompt": _ANALYZE_PROMPT,
-                        "images": cleaned_b64_list,
+                        "images": analysis_images,
                         "stream": False,
                         "keep_alive": 0,
                         "options": {
@@ -538,6 +564,17 @@ async def analyze_character_endpoint(request):
                         },
                     }
                     async with session.post(f"{base_url}/api/generate", json=payload, timeout=300) as response:
+                        if response.status != 200:
+                            err_txt = await response.text()
+                            if response.status == 400 and "exceeds the available context size" in err_txt:
+                                analysis_images = _prepare_analysis_images(cleaned_b64_list, max_dim=512, quality=70)
+                            else:
+                                return web.json_response({"status": "error", "message": f"Ollama HTTP {response.status}: {err_txt}"})
+                        else:
+                            resp_json = await response.json()
+                            generated_text = _extract_ollama_generated_text(resp_json)
+                    if not generated_text:
+                        async with session.post(f"{base_url}/api/generate", json={**payload, "images": analysis_images}, timeout=300) as response:
                             if response.status != 200:
                                 err_txt = await response.text()
                                 return web.json_response({"status": "error", "message": f"Ollama HTTP {response.status}: {err_txt}"})
@@ -552,7 +589,7 @@ async def analyze_character_endpoint(request):
                             "messages": [{
                                 "role": "user",
                                 "content": _ANALYZE_PROMPT,
-                                "images": cleaned_b64_list,
+                                "images": analysis_images,
                             }],
                             "stream": False,
                             "keep_alive": 0,
@@ -565,10 +602,31 @@ async def analyze_character_endpoint(request):
                             if response.status == 200:
                                 resp_json = await response.json()
                                 generated_text = _extract_ollama_generated_text(resp_json)
+                            else:
+                                err_txt = await response.text()
+                                if response.status == 400 and "exceeds the available context size" in err_txt:
+                                    smaller_images = _prepare_analysis_images(cleaned_b64_list, max_dim=384, quality=60)
+                                    retry_payload = {
+                                        **chat_payload,
+                                        "messages": [{
+                                            "role": "user",
+                                            "content": _ANALYZE_PROMPT,
+                                            "images": smaller_images,
+                                        }],
+                                    }
+                                    async with session.post(f"{base_url}/api/chat", json=retry_payload, timeout=300) as retry_response:
+                                        if retry_response.status == 200:
+                                            resp_json = await retry_response.json()
+                                            generated_text = _extract_ollama_generated_text(resp_json)
+                                        else:
+                                            err_txt = await retry_response.text()
+                                            return web.json_response({"status": "error", "message": f"Ollama HTTP {retry_response.status}: {err_txt}"})
+                                else:
+                                    return web.json_response({"status": "error", "message": f"Ollama HTTP {response.status}: {err_txt}"})
                 else:
                     # OpenAI-compatible vision chat (LM Studio / Custom).
                     content = [{"type": "text", "text": _ANALYZE_PROMPT}]
-                    for b64 in cleaned_b64_list:
+                    for b64 in analysis_images:
                         content.append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64}"}})
                     payload = {
                         "model": model_name,