Before this fix, using Wan 2.2 I2V models like 'wan2_2_i2v_a14b' would
incorrectly try to load 'Wan-AI/Wan2.1-T2V-14B-Diffusers' as the base
model because the fallback logic only matched 'wan' generically.

Now the base_model_fallbacks dictionary includes specific entries for:
- Wan 2.2 I2V models: wan2_2_i2v, wan2.2_i2v
- Wan 2.2 T2V models: wan2_2, wan2.2
- Wan 2.1 I2V models: wan2_1_i2v, wan2.1_i2v
- Wan 2.1 T2V models: wan2_1, wan2.1
- Generic Wan fallback: wan

The more specific keys are checked first, so model IDs like
'wan2_2_i2v_a14b' will correctly match 'wan2_2_i2v' and use
'Wan-AI/Wan2.2-I2V-14B-Diffusers' as the base model.

- Add get_pipeline_for_task() function that determines pipeline
  based on model ID AND task type (t2v, i2v, t2i, i2i, v2v)
- Pipeline class is now ALWAYS detected at runtime, not from config
- Remove old dynamic switching code that's now redundant
- Update check_model.py to show runtime detection instead of fixing config
- Update check_pipelines.py to show V2V pipelines

- Add WanImageToVideoPipeline and WanVideoToVideoPipeline to PIPELINE_CLASS_MAP
- Add CogVideoXImageToVideoPipeline and CogVideoXVideoToVideoPipeline
- Add AnimateDiffVideoToVideoPipeline
- Add StableDiffusionImg2ImgPipeline for SD 1.5
- Add dynamic pipeline switching logic for Wan, LTX, CogVideoX, AnimateDiff
- The pipeline class is now selected at runtime based on task mode
- Fix detect_pipeline_class to correctly identify Wan models
- Remove duplicate LTX handling code

Adjust VRAM checking to allow models up to 10% less than available VRAM, or full VRAM with offload strategy

- When loading LTX-Video base model in I2V mode, use LTXImageToVideoPipeline
- When loading LTX-Video base model in T2V mode, use LTXPipeline
- Update PipelineClass after loading base pipeline to match the actual class used
- This fixes the 'LTXPipeline.__call__() got an unexpected keyword argument image' error

- Add LTXImageToVideoPipeline to PIPELINE_CLASS_MAP as I2V type
- Change LTXPipeline to T2V type (it doesn't support image input)
- Auto-switch from LTXPipeline to LTXImageToVideoPipeline when in I2V mode
- Update I2V check to only verify pipeline class name (removed supports_i2v flag check)
- This allows LTX-Video models to work in I2V mode with image input

- LTXPipeline is T2V-only, not I2V-capable
- Check both supports_i2v flag AND pipeline class name
- I2V-capable pipelines: StableVideoDiffusionPipeline, I2VGenXLPipeline, LTXImageToVideoPipeline, WanPipeline
- Prevents passing 'image' argument to T2V-only pipelines like LTXPipeline

- In T2I mode, only show image models in the main model select
- In other modes (T2V, I2V, etc.), show video models
- Repopulate model select when switching modes
- Image model select for I2V mode always shows image models

- Remove orphaned 'except Exception as e:' block that had no matching try
- Fix indentation for wan scheduler configuration
- Change 'I2V model loaded' to generic 'Model loaded' message

- Set pipeline_loaded_successfully=True when component loading succeeds
- Fix indentation for LoRA and offloading code blocks
- Define 'off' variable inside the correct scope
- This fixes loading models like Muinez/ltxvideo-2b-nsfw which are
  fine-tuned transformer weights without a full pipeline

- sentencepiece>=0.2.0 required for LTX-Video tokenizer
- protobuf>=5.27.0 required by sentencepiece
- These are needed to parse the spiece.model tokenizer file

- Move T2I detection before time estimation
- Add has_t2i parameter to estimate_total_time
- T2I models now show image_generation time instead of video_generation
- Add Lumina pipelines to T2I model detection
- Suppress 'Loaded models' message when --json flag is used

- Move JSON output to happen before table printing
- Add better error handling in webapp.py for debugging
- This fixes the web interface model list not showing

- Can remove by numeric ID (from --model-list) or by name
- Also supports removing by HuggingFace model ID
- Updates models.json after removal

- Add Lumina pipeline classes to PIPELINE_CLASS_MAP
- Fix Lumina model detection in detect_pipeline_class
- Remove non-existent LuminaVideoPipeline references
- Update Alpha-VLLM/Lumina-Next-SFT to use correct pipeline class
- Lumina-Next-SFT is a T2I model, not T2V (~20-30GB VRAM)

- Add --json argument to output model list in JSON format
- Include model capabilities, disabled status, and fail count in JSON output
- This fixes the web interface model list not showing models

- Detect tokenizer parsing errors and provide helpful cache clearing instructions
- Add retry logic for corrupted cache files
- Improve error messages for component-only model loading

- The correct class name in diffusers is LTXPipeline, not LTXVideoPipeline
- Updated PIPELINE_CLASS_MAP and detect_pipeline_class
- Updated all references throughout the codebase
- This fixes loading models like Muinez/ltxvideo-2b-nsfw

- When loading fine-tuned component models (like LTXVideoTransformer3DModel),
  use the correct pipeline class for the base model instead of the configured
  PipelineClass which may be wrong
- Add proper pipeline class detection for LTX, Wan, SVD, CogVideo, Mochi
- This fixes loading models like Muinez/ltxvideo-2b-nsfw which have
  config.json only (no model_index.json)

- Add character profile management (create, list, show, delete)
- Add IP-Adapter and InstantID support for character consistency
- Fix model loading for models with config.json only (no model_index.json)
- Add component-only model detection (fine-tuned weights)
- Update MCP server with character consistency tools
- Update SKILL.md and README.md documentation
- Add memory management for dubbing/translation
- Add chunked processing for Whisper transcription
- Add character persistency options to web interface

Users can now upload their own image for I2V mode instead of
only being able to generate one. The image upload box is now
visible in I2V mode, allowing users to either:
- Upload an existing image to animate
- Or let the system generate an image first

This provides more flexibility for the I2V workflow.

Style Detection (detect_generation_type):
- Detects 9 artistic styles: anime, photorealistic, digital_art, cgi, cartoon, fantasy, traditional, scifi, horror
- Extracts style keywords from prompts for matching
- Returns style info in generation type dict

Style Matching (select_best_model):
- Matches LoRA adapters to requested style (+60 bonus for style match)
- Matches base models to requested style (+50 bonus for style match)
- Checks model name, ID, and tags for style indicators
- Examples:
  - 'anime girl' → selects anime-optimized models/LoRAs
  - 'photorealistic portrait' → selects realism models
  - 'cyberpunk city' → selects sci-fi models/LoRAs

This allows --auto mode to intelligently select models based on
the artistic style requested in the prompt.