Add Top-K and Top-P sampling parameters to match VQA capabilities:
- Top-K: 0-100 (default 0 = disabled)
- Top-P: 0.0-1.0 (default 0 = disabled)
Parameters are only passed to the model when > 0, matching VQA behavior.
Update default settings for both Prompt Enhance and VQA:
- max_tokens: 512 (prompt enhance)
- do_sample: True
- temperature: 0.8
These defaults provide better output quality with more creative
variation while maintaining coherent results.
Add fallback to ui_control_helpers.input_source when Gradio image
component is empty, which occurs when Kanvas is active (replaces
gr.Image with gr.HTML).
Also standardize debug logging:
- Add debug at enhance() entry, chat template, tokenization
- Change shared.log.trace() calls to use debug_log() for consistency
- All debug output controlled by SD_LLM_DEBUG environment variable
Clarify purpose of empty fields with placeholder text:
- System prompt: explains built-in instructions used when empty
- Prefix/suffix: describe where text is added to result
- Output: indicates where enhanced prompt appears
Improve formatting and clarity of prompt enhance tooltips:
- Add spacing with <br><br> for better readability
- Clarify that prefix/suffix bypass LLM rewriting
Add prefill text support and proper thinking tag handling:
- Prefill text input to pre-fill start of model response
- Keep Prefill toggle to retain/remove prefill from output
- Keep Thinking Trace toggle to format <think> tags as readable output
- Proper <think> tag cleanup: strip or format as "Reasoning:/Answer:"
- Support for prefill + thinking combination via keep_think_block_open()
Chat template uses continue_final_message=True when prefill provided,
allowing model to continue from the prefill text naturally.
Add visual capability indicators and user control for image input:
- Vision symbol (eye icon) for VL-capable models in dropdown
- Reasoning symbol (lightbulb) for thinking-capable models
- "Use vision" checkbox to control image input for enhancement
- Toggle dims and unchecks when non-VL model selected
- Vision auto-enables when switching to VL model
Also:
- Rename "Do sample" to "Use samplers" for consistency with VQA
- Add tooltips/hints for all prompt enhance UI elements
- Add CSS styling for dimmed vision toggle appearance
Disty0
vladmandic
awsr
CalamitousFelicitousness