cleanup

Signed-off-by: vladmandic <mandic00@live.com>

CHANGELOG.md

@@ -1,12 +1,12 @@
 # Change Log for SD.Next
 
-## Update for 2025-12-09
+## Update for 2025-12-10
 
 ### TBD
 
 Merge commit: `f903a36d9`
 
-### Highlights for 2025-12-09
+### Highlights for 2025-12-10
 
 New native [kanvas](https://vladmandic.github.io/sdnext-docs/Kanvas/) module for image manipulation that fully replaces *img2img*, *inpaint* and *outpaint* controls, massive update to **Captioning/VQA** models and features  
 New generation of **Flux.2** large image model, new **Z-Image** model that is creating a lot of buzz, new **Kandinsky 5 Lite** image model and new **Photoroom PRX** model  
@@ -18,8 +18,8 @@ Plus a lot of internal improvements and fixes
 
 [ReadMe](https://github.com/vladmandic/automatic/blob/master/README.md) | [ChangeLog](https://github.com/vladmandic/automatic/blob/master/CHANGELOG.md) | [Docs](https://vladmandic.github.io/sdnext-docs/) | [WiKi](https://github.com/vladmandic/automatic/wiki) | [Discord](https://discord.com/invite/sd-next-federal-batch-inspectors-1101998836328697867) | [Sponsor](https://github.com/sponsors/vladmandic)  
 
-### Details for 2025-12-09
-  
+### Details for 2025-12-10
+
 - **Models**
   - [Black Forest Labs FLUX.2 Dev](https://bfl.ai/blog/flux-2) and prequantized variation [SDNQ-SVD-Uint4](https://huggingface.co/Disty0/FLUX.2-dev-SDNQ-uint4-svd-r32)  
     **FLUX.2-Dev** is a brand new model from BFL and uses large 32B DiT together with Mistral 24B as text encoder  
@@ -50,23 +50,23 @@ Plus a lot of internal improvements and fixes
   kanvas is a full replacement for *img2img, inpaint and outpaint* controls  
   see [docs](https://vladmandic.github.io/sdnext-docs/Kanvas/) for details  
   *experimental*: report any feedback in master [issue](https://github.com/vladmandic/sdnext/issues/4358)  
-- **Captioning** and **VQA: visual query analyzer**  
+- **Captioning** and **VQA: Visual Question & Answer**  
   massive update to both features and supported models, thanks @CalamitousFelicitousness  
-  models:
+  models:  
   - additional `mooondream-2` features  
   - support for `moondream-3-preview`  
   - support for `qwen3-vl` with thinking  
   - additional `gemma-3-vl` finetunes  
   - support for `XiaomiMiMo`  
-  features:  
+  ui:  
   - ability to annotate actual image, not just generate captions/answers  
     e.g. actualy mark detected regions/points  
+  features:  
   - ui indicator of model capabilities  
   - support for *prefill* style of prompting/answering  
   - support for *reasoning* mode for supported models  
     with option to output answer-only or reasoning-process   
   - additional debug logging  
-
 - **Other Features**
   - **wildcards**: allow recursive inline wildcards using curly braces syntax  
   - **sdnq**: simplify pre-quantization saved config  
@@ -99,6 +99,7 @@ Plus a lot of internal improvements and fixes
   - python: start work on future-proofing for modern python versions, thanks @awsr  
   - nunchaku: update to `1.0.2`  
   - lint: add rules for run-on-windows  
+  - gallery: setting to enable/disable client-side caching, thanks @awsr  
   - gallery: faster thumbnail generation, thanks @awsr  
   - gallery: purge old thumbnails, thanks @awsr  
 - **Docs**

html/locale_en.json

@@ -38,7 +38,7 @@
   {"id":"","label":"VLM: Batch Caption","localized":"Batch Caption","reload":"","hint":"Process multiple images in a batch using VLM."},
   {"id":"","label":"CLiP: Advanced Options","localized":"Advanced Options","reload":"","hint":"Advanced configuration options for CLiP interrogation."},
   {"id":"","label":"CLiP: Batch Interrogate","localized":"Batch Interrogate","reload":"","hint":"Process multiple images in a batch using CLiP."},
-  {"id":"","label":"Task","localized":"","reload":"","hint":"Changes which task the model will perform. Regular text prompts can be used when the default option <b>Use Prompt</b> is selected.<br>When other options are selected, see the hint text inside an empty <b>Prompt</b> field for guidance."},
+  {"id":"","label":"Task","localized":"","reload":"","hint":"Changes which task the model will perform. Regular text prompts can be used when the task is set to <b>Use Prompt</b>.<br>When other options are selected, see the hint text inside an empty <b>Prompt</b> field for guidance."},
   {"id":"","label":"Prefill text","localized":"","reload":"","hint":"Pre-fills the start of the model's response to guide its output format or content by forcing it to continue the prefill text.<br>Prefill is filtered out and does not appear in the final response.<br><br>Leave empty to let the model generate its own response from scratch."},
   {"id":"","label":"Start","localized":"","reload":"","hint":"Start"},
   {"id":"","label":"End","localized":"","reload":"","hint":"End"},

modules/interrogate/vqa.py

@@ -359,7 +359,6 @@ class VQA:
         self.processor = None
         self.model = None
         self.loaded: str = None
-        self.quant_args = None
         self.last_annotated_image = None
         self.last_detection_data = None
 
@@ -386,7 +385,6 @@ class VQA:
             shared.log.error(f'VQA load: unknown model="{model_name}"')
             return
 
-        self.quant_args = model_quant.create_config(module='LLM')
         shared.log.debug(f'VQA load: pre-loading model="{model_name}" repo="{repo}"')
 
         # Dispatch to appropriate loader (same logic as interrogate)
@@ -448,6 +446,7 @@ class VQA:
         """Load FastVLM model and tokenizer."""
         if self.model is None or self.loaded != repo:
             shared.log.debug(f'Interrogate load: vlm="{repo}"')
+            quant_args = model_quant.create_config(module='LLM')
             self.model = None
             self.processor = transformers.AutoTokenizer.from_pretrained(repo, trust_remote_code=True, cache_dir=shared.opts.hfcache_dir)
             self.model = transformers.AutoModelForCausalLM.from_pretrained(
@@ -455,7 +454,7 @@ class VQA:
                 torch_dtype=devices.dtype,
                 trust_remote_code=True,
                 cache_dir=shared.opts.hfcache_dir,
-                **self.quant_args,
+                **quant_args,
             )
             self.loaded = repo
             devices.torch_gc()
@@ -502,11 +501,12 @@ class VQA:
                 cls_name = transformers.Qwen2VLForConditionalGeneration
             else:
                 cls_name = transformers.AutoModelForCausalLM
+            quant_args = model_quant.create_config(module='LLM')
             self.model = cls_name.from_pretrained(
                 repo,
                 torch_dtype=devices.dtype,
                 cache_dir=shared.opts.hfcache_dir,
-                **self.quant_args,
+                **quant_args,
             )
             self.processor = transformers.AutoProcessor.from_pretrained(repo, max_pixels=1024*1024, cache_dir=shared.opts.hfcache_dir)
             if 'LLM' in shared.opts.cuda_compile:
@@ -638,11 +638,12 @@ class VQA:
                 cls = transformers.Gemma3nForConditionalGeneration  # pylint: disable=no-member
             else:
                 cls = transformers.Gemma3ForConditionalGeneration
+            quant_args = model_quant.create_config(module='LLM')
             self.model = cls.from_pretrained(
                 repo,
                 torch_dtype=devices.dtype,
                 cache_dir=shared.opts.hfcache_dir,
-                **self.quant_args,
+                **quant_args,
             )
             if 'LLM' in shared.opts.cuda_compile:
                 self.model = sd_models_compile.compile_torch(self.model)
@@ -839,11 +840,12 @@ class VQA:
         if self.model is None or self.loaded != repo:
             shared.log.debug(f'Interrogate load: vlm="{repo}"')
             self.model = None
+            quant_args = model_quant.create_config(module='LLM')
             self.model = transformers.AutoModelForVision2Seq.from_pretrained(
                 repo,
                 cache_dir=shared.opts.hfcache_dir,
                 torch_dtype=devices.dtype,
-                **self.quant_args,
+                **quant_args,
             )
             self.processor = transformers.AutoProcessor.from_pretrained(repo, max_pixels=1024*1024, cache_dir=shared.opts.hfcache_dir)
             if 'LLM' in shared.opts.cuda_compile:
@@ -1150,13 +1152,14 @@ class VQA:
             shared.log.debug(f'Interrogate load: vlm="{repo_name}" revision="{effective_revision}" path="{shared.opts.hfcache_dir}"')
             transformers.dynamic_module_utils.get_imports = get_imports
             self.model = None
+            quant_args = model_quant.create_config(module='LLM')
             self.model = transformers.Florence2ForConditionalGeneration.from_pretrained(
                 repo_name,
                 dtype=torch.bfloat16,
                 revision=effective_revision,
                 torch_dtype=devices.dtype,
                 cache_dir=shared.opts.hfcache_dir,
-                **self.quant_args,
+                **quant_args,
             )
             self.processor = transformers.AutoProcessor.from_pretrained(repo_name, max_pixels=1024*1024, trust_remote_code=True, revision=effective_revision, cache_dir=shared.opts.hfcache_dir)
             transformers.dynamic_module_utils.get_imports = _get_imports
@@ -1230,7 +1233,6 @@ class VQA:
         self.last_detection_data = None
         jobid = shared.state.begin('Interrogate LLM')
         t0 = time.time()
-        self.quant_args = model_quant.create_config(module='LLM')
         model_name = model_name or shared.opts.interrogate_vlm_model
         prefill = vlm_prefill if prefill is None else prefill  # Use provided prefill when specified
         if isinstance(image, list):

modules/shared.py

@@ -502,7 +502,7 @@ options_templates.update(options_section(('saving-images', "Image Options"), {
     "font_color": OptionInfo("#FFFFFF", "Font color", gr.ColorPicker, {}),
 
     "image_sep_browser": OptionInfo("<h2>Image Gallery</h2>", "", gr.HTML),
-    "browser_cache": OptionInfo(True, "Use image gallery cache [STRONGLY recommended]"),
+    "browser_cache": OptionInfo(True, "Use image gallery cache"),
     "browser_folders": OptionInfo("", "Additional image browser folders"),
     "browser_fixed_width": OptionInfo(False, "Use fixed width thumbnails"),
     "viewer_show_metadata": OptionInfo(True, "Show metadata in full screen image browser"),
@@ -643,7 +643,7 @@ options_templates.update(options_section(('postprocessing', "Postprocessing"), {
 }))
 
 options_templates.update(options_section(('interrogate', "Interrogate"), {
-    "interrogate_default_type": OptionInfo("OpenCLiP", "Default type", gr.Radio, {"choices": ["OpenCLiP", "VLM", "DeepBooru"]}),
+    "interrogate_default_type": OptionInfo("VLM", "Default caption type", gr.Radio, {"choices": ["OpenCLiP", "VLM", "DeepBooru"]}),
     "interrogate_offload": OptionInfo(True, "Offload models "),
     "interrogate_score": OptionInfo(False, "Include scores in results when available"),
 

modules/ui_caption.py

@@ -3,6 +3,8 @@ from modules import shared, ui_common, generation_parameters_copypaste
 from modules.interrogate import openclip
 
 
+default_task = "Short Caption"
+
 def vlm_caption_wrapper(question, system_prompt, prompt, image, model_name, prefill, thinking_mode):
     """Wrapper for vqa.interrogate that handles annotated image display."""
     from modules.interrogate import vqa
@@ -17,7 +19,7 @@ def update_vlm_prompts_for_model(model_name):
     """Update the task dropdown choices based on selected model."""
     from modules.interrogate import vqa
     prompts = vqa.get_prompts_for_model(model_name)
-    return gr.update(choices=prompts, value=prompts[0] if prompts else "Use Prompt")
+    return gr.update(choices=prompts, value=prompts[0] if prompts else default_task)
 
 
 def update_vlm_prompt_placeholder(question):
@@ -68,9 +70,9 @@ def create_ui():
                     with gr.Row():
                         vlm_system = gr.Textbox(label="System Prompt", value=vqa.vlm_system, lines=1, elem_id='vlm_system')
                     with gr.Row():
-                        vlm_question = gr.Dropdown(label="Task", allow_custom_value=False, choices=initial_prompts, value=initial_prompts[0] if initial_prompts else "Use Prompt", elem_id='vlm_question')
+                        vlm_question = gr.Dropdown(label="Task", allow_custom_value=False, choices=initial_prompts, value=default_task, elem_id='vlm_question')
                     with gr.Row():
-                        vlm_prompt = gr.Textbox(label="Prompt", placeholder=vqa.get_prompt_placeholder(initial_prompts[0] if initial_prompts else "Use Prompt"), lines=2, elem_id='vlm_prompt')
+                        vlm_prompt = gr.Textbox(label="Prompt", placeholder=vqa.get_prompt_placeholder(initial_prompts[0]), lines=2, elem_id='vlm_prompt')
                     with gr.Row(elem_id='interrogate_buttons_query'):
                         vlm_model = gr.Dropdown(list(vqa.vlm_models), value=current_vlm_model, label='VLM Model', elem_id='vlm_model')
                     with gr.Row():