Fix SDXL Diffusers Format Support & Add Manual Model Type Override

[fewtarius]

Summary

This PR fixes loading of SDXL models in diffusers directory format and adds a new --model-type CLI parameter for manual model version control.

Problem

When loading SDXL models in diffusers directory format (e.g., duchaiten-pony-real-v20-sdxl), the text encoder tensors failed to load with "unknown tensor" or "tensor not in model file" errors. This occurred because:

1. Missing prefix conversions: The te. and te.1. prefixes used by diffusers format weren't in the name conversion list
2. Incorrect tensor prefixes: Text encoders were loaded with prefixes that didn't match what the model graph expects
3. Version detection issues: Later components (VAE) could override the already-detected SDXL version

Solution

1. Name Conversion Support (name_conversion.cpp)

Added te. and te.1. prefixes to the cond_stage_model conversion list, enabling proper tensor name translation from diffusers format to checkpoint format.

2. Correct Tensor Prefixes (model.cpp)

• Changed text_encoder prefix from te. to cond_stage_model.transformer.
• Changed text_encoder_2 prefix from te.1. to cond_stage_model.1.transformer.
• Added early return in get_sd_version() when SDXL is detected to prevent VAE from overriding version

3. Version Caching (stable-diffusion.cpp)

Added version caching to preserve SDXL detection across component loading phases.

4. Manual Model Type Override (--model-type parameter)

New CLI parameter allowing users to manually specify the model version:

--model-type sdxl          # Force SDXL version
--model-type sd1           # Force SD 1.x version
--model-type flux          # Force FLUX version

Supported values: sd1, sd2, sdxl, sdxl_inpaint, sdxl_pix2pix, flux, sd3, svd

Use cases:

• Auto-detection fails or is ambiguous
• Testing model behavior with different version settings
• Working with modified/custom models

Files Changed

File	Changes
src/name_conversion.cpp	+2 lines: Added te. and te.1. prefix mappings
src/model.cpp	+15/-3 lines: Updated diffusers text encoder prefixes, improved SDXL detection
src/stable-diffusion.cpp	+18/-4 lines: Added version caching and manual override logic
include/stable-diffusion.h	+1 line: Added version_override field to sd_ctx_params_t
examples/common/common.hpp	+40 lines: Added --model-type CLI parameter and string-to-enum conversion

Testing

• ✅ SDXL diffusers models now load successfully
• ✅ SD 1.5 models continue to work (regression tested)
• ✅ All text encoder tensors are found and loaded correctly
• ✅ Manual override with --model-type sdxl works
• ✅ Auto-detection still works when --model-type is not specified

Commits

1. fix(diffusers): use correct tensor name prefixes for SDXL text encoders - Core fix for tensor prefix matching
2. feat(cli): add --model-type parameter for manual version override - User-facing control over model detection
3. fix(diffusers): add support for diffusers SDXL text encoder prefixes - Name conversion support for diffusers format

Compatibility

• Backward compatible: All existing workflows continue to work unchanged
• No breaking changes: Auto-detection remains the default behavior
• Additive only: New functionality doesn't modify existing behavior

Related Issues

Fixes loading of diffusers-format SDXL models that previously failed with tensor loading errors.