Calculate: base_vram + 2GB + 50%
This ensures a 14B model estimated at 18GB will require ~29GB instead of 22.5GB.

Instead of adding a fixed 2GB overhead, now calculates 25% more
VRAM for base models that will have fine-tuned weights/tensors
or LoRA adapters loaded on top.

The user confirmed that Wan2.2-I2V models should use Wan2.2-T2V
as the base model, not the I2V variant.

The issue was that model IDs from HuggingFace use dots (wan2.2-i2v-a14b)
while user config names use underscores (wan2_2_i2v_a14b).

Now we normalize the model ID by replacing underscores with dots before
matching against the base_model_fallbacks dictionary.

The issue was that model IDs like 'wan2_2_i2v_a14b' would match
'wan2_2' (T2V) before 'wan2_2_i2v' (I2V) because 'wan2_2' comes
first in the dictionary.

Now the dictionary is ordered with more specific keys first:
- wan2_2_i2v_a14b (most specific)
- wan2.2_i2v_a14b
- wan2_2_i2v
- wan2.2_i2v
- wan2_2
- wan2.2
etc.

This ensures longer/more specific keys are checked before shorter ones.

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