{"passport":{"unfragile":{"@version":"1.0","version":"2026-05","artifact":{"id":"github-cloneofsimo--lora","slug":"cloneofsimo--lora","name":"lora","type":"model","url":"https://arxiv.org/abs/2106.09685","page_url":"https://unfragile.ai/cloneofsimo--lora","categories":["image-generation"],"tags":["diffusion","dreambooth","fine-tuning","lora","stable-diffusion"],"pricing":{"model":"open_source","free":true,"starting_price":null},"status":"inactive","verified":false},"capabilities":[{"id":"github-cloneofsimo--lora__cap_0","uri":"capability://code.generation.editing.low.rank.weight.decomposition.for.diffusion.model.fine.tuning","name":"low-rank weight decomposition for diffusion model fine-tuning","description":"Decomposes model weight updates into low-rank matrix products (W' = W + ΔW where ΔW = A×B^T) using trainable matrices A and B with rank d << min(n,m), reducing trainable parameters by 10-100× compared to full fine-tuning. Implements LoraInjectedLinear and LoraInjectedConv2d layer classes that wrap original weights and apply low-rank updates during forward passes without modifying base model weights.","intents":["Fine-tune Stable Diffusion on custom concepts with minimal VRAM and storage","Train specialized image generation models in 2-3× less time than full fine-tuning","Create portable 1-6MB model adapters instead of full multi-gigabyte checkpoints","Combine multiple fine-tuned LoRA models for compositional generation"],"best_for":["Individual researchers and hobbyists with limited GPU memory (8-16GB)","Teams building custom image generation services with cost constraints","Practitioners needing rapid iteration on model specialization"],"limitations":["Rank parameter d must be manually tuned; too low loses expressiveness, too high negates efficiency gains","Cannot capture structural changes to model behavior—only fine-grained weight adjustments","Requires careful initialization of A and B matrices to avoid training instability","Performance gains diminish when combining >3-4 LoRA adapters simultaneously due to cumulative latency"],"requires":["PyTorch 1.9+","Hugging Face diffusers library","GPU with minimum 8GB VRAM for training (6GB for inference)","Original Stable Diffusion base model weights"],"input_types":["Pre-trained diffusion model (safetensors or CKPT format)","Training dataset (image-text pairs or concept images)"],"output_types":["LoRA weight matrices (1-6MB files)","Merged model weights (full checkpoint)"],"categories":["code-generation-editing","model-training","parameter-efficient-fine-tuning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_1","uri":"capability://image.visual.dreambooth.training.with.prior.preservation.regularization","name":"dreambooth training with prior-preservation regularization","description":"Implements subject-specific fine-tuning by training on a small set of target images (3-5) while using class-prior images to prevent overfitting and catastrophic forgetting. The training loop alternates between updating the model on target images and regularizing with class images, using a weighted loss that balances concept learning against generalization. Integrates with LoRA to make this process memory-efficient.","intents":["Fine-tune a model to generate a specific person, object, or style from just 3-5 example images","Prevent the model from forgetting how to generate the broader class (e.g., 'dog') while learning a specific instance","Train subject-specific adapters without full model fine-tuning overhead","Create reusable identity tokens that work across different prompts and contexts"],"best_for":["Content creators personalizing image generation for specific subjects","E-commerce platforms generating product variations with consistent branding","Portrait and headshot generation services"],"limitations":["Requires manual collection of 3-5 high-quality target images; poor image quality degrades results","Prior-preservation class images must be semantically similar but visually distinct; mismatched classes cause mode collapse","Training is sensitive to hyperparameters (learning rate, prior loss weight); requires experimentation per subject","Cannot learn complex multi-subject compositions or fine-grained attribute combinations"],"requires":["PyTorch 1.9+","3-5 target images of the subject (512×512 or higher)","100-200 class-prior images (can be auto-generated or sourced)","8GB+ GPU VRAM","Stable Diffusion base model"],"input_types":["Target image set (JPEG/PNG)","Class-prior image set (JPEG/PNG)","Text prompt template with [V] placeholder for learned token"],"output_types":["LoRA adapter weights","Trained token embeddings"],"categories":["image-visual","model-training","fine-tuning-specialized"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_10","uri":"capability://data.processing.analysis.lora.model.composition.and.interpolation","name":"lora model composition and interpolation","description":"Enables combining multiple trained LoRA adapters by stacking their low-rank updates (ΔW_total = α₁·ΔW₁ + α₂·ΔW₂ + ...) with learnable or fixed weights. Supports linear interpolation between LoRA models in weight space, enabling smooth transitions between different concepts or styles. Implements composition without retraining by directly manipulating weight matrices.","intents":["Combine style and subject LoRA adapters for compositional generation (e.g., 'person in style')","Blend between multiple fine-tuned models to create intermediate concepts","Create dynamic LoRA weights that vary across generation steps","Explore the latent space of LoRA adaptations via interpolation"],"best_for":["Creative practitioners exploring compositional generation","Researchers studying model interpolation and concept blending","Production systems requiring flexible concept composition"],"limitations":["Composing >3-4 LoRA adapters causes cumulative latency (~50-100ms per adapter) and potential numerical instability","Interpolation assumes LoRA models are trained on compatible base models; mismatched bases produce artifacts","No built-in conflict detection when composing adapters targeting overlapping layers; manual tuning of blend weights required","Composition weights must be manually tuned; no automatic optimization for semantic coherence"],"requires":["PyTorch 1.9+","Multiple trained LoRA checkpoints","Base model compatible with all LoRA adapters"],"input_types":["Multiple LoRA checkpoints","Composition weights (α₁, α₂, ...)","Interpolation parameter (0-1 for blending)"],"output_types":["Composed LoRA weights","Interpolated LoRA weights","Generated images from composed adapters"],"categories":["data-processing-analysis","model-training","model-composition"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_11","uri":"capability://image.visual.inference.with.multi.lora.application.and.dynamic.weight.scheduling","name":"inference with multi-lora application and dynamic weight scheduling","description":"Enables applying multiple LoRA adapters during inference with per-step or per-layer weight scheduling. Supports dynamic adjustment of LoRA influence across diffusion timesteps, allowing different concepts to dominate at different denoising stages. Implements efficient inference by caching composed weights and avoiding redundant computation.","intents":["Generate images with multiple LoRA adapters applied simultaneously","Control concept influence across diffusion steps (e.g., strong subject early, strong style late)","Create dynamic generation effects by varying LoRA weights during inference","Optimize inference latency by caching composed weights"],"best_for":["Creative applications requiring compositional control","Production systems generating images with multiple concepts","Researchers studying diffusion dynamics and concept interaction"],"limitations":["Multi-LoRA inference is slower than single-adapter (cumulative latency ~50-100ms per adapter)","Dynamic weight scheduling requires manual tuning; no automatic optimization for semantic coherence","Caching composed weights increases memory usage; not suitable for memory-constrained environments","Numerical instability can occur with >4-5 adapters or extreme weight values"],"requires":["PyTorch 1.9+","Multiple trained LoRA checkpoints","Base diffusion model","6GB+ GPU VRAM"],"input_types":["Text prompt","Multiple LoRA checkpoints","Composition weights (per-step or per-layer)","Diffusion parameters (steps, guidance scale, etc.)"],"output_types":["Generated image","Intermediate latents (optional)"],"categories":["image-visual","automation-workflow","inference-optimization"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_12","uri":"capability://data.processing.analysis.batch.preprocessing.and.dataset.preparation.utilities","name":"batch preprocessing and dataset preparation utilities","description":"Provides CLI tool lora_ppim for automated preprocessing of training datasets including image resizing, cropping, augmentation, and caption generation. Handles batch operations on image directories, validates image quality, and generates metadata files required for training. Supports multiple preprocessing strategies (center crop, random crop, aspect-ratio preservation).","intents":["Prepare raw image datasets for training without manual preprocessing","Standardize image dimensions and formats across datasets","Generate captions or metadata for training images","Validate dataset quality before training"],"best_for":["Practitioners with large raw image collections","Teams standardizing dataset preparation workflows","Researchers managing multiple datasets with varying formats"],"limitations":["Automated preprocessing may introduce artifacts (e.g., center crop loses important details)","Caption generation requires external models (BLIP, CLIP); quality varies significantly","Batch operations are I/O bound; slow on network storage or spinning disks","No built-in validation for semantic consistency; garbage-in-garbage-out if source images are poor quality"],"requires":["Python 3.7+","PIL/Pillow for image processing","Optional: BLIP or CLIP for caption generation","Disk space for preprocessed images (typically 2-3× source size)"],"input_types":["Raw image directory","Configuration file specifying preprocessing steps"],"output_types":["Preprocessed image directory","Metadata file (JSON/CSV with captions and image info)"],"categories":["data-processing-analysis","automation-workflow","dataset-preparation"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_2","uri":"capability://image.visual.textual.inversion.token.embedding.learning","name":"textual inversion token embedding learning","description":"Learns new token embeddings in the CLIP text encoder's vocabulary space by optimizing a learnable embedding vector [V] that captures a concept's visual characteristics. During training, the model freezes all diffusion weights and only updates the embedding vector via backpropagation through the text encoder and UNet, allowing the model to bind arbitrary concepts to new tokens without modifying model weights.","intents":["Create reusable text tokens that represent specific visual concepts or styles","Learn embeddings for rare concepts from a small image set without full model fine-tuning","Enable prompt-based control over concept injection across different base models","Compose multiple learned concepts in a single prompt"],"best_for":["Style transfer and artistic concept learning","Rare object or subject representation with minimal training data","Researchers studying concept representation in diffusion models"],"limitations":["Embedding learning is slower than LoRA (requires full backprop through text encoder) and produces larger files (50-100KB per token)","Learned embeddings are tied to the specific CLIP model; transferring to different text encoders requires retraining","Cannot learn structural or compositional concepts—only visual style/appearance","Embedding quality degrades with dataset size <10 images; requires careful prompt engineering"],"requires":["PyTorch 1.9+","CLIP text encoder (from diffusers)","10-50 training images of the concept","6GB+ GPU VRAM","Stable Diffusion base model"],"input_types":["Concept image set (JPEG/PNG, 512×512)","Initialization embedding (random or from pretrained)"],"output_types":["Learned embedding vector (PyTorch tensor, 50-100KB)","Text token identifier (e.g., 'sks')"],"categories":["image-visual","model-training","embedding-learning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_3","uri":"capability://image.visual.pivotal.tuning.inversion.pti.hybrid.fine.tuning","name":"pivotal tuning inversion (pti) hybrid fine-tuning","description":"Combines DreamBooth and Textual Inversion by jointly optimizing both LoRA weights and learned token embeddings during training. The method alternates between updating LoRA parameters on target images and refining the learned embedding, allowing the model to capture both structural adaptations (via LoRA) and semantic concept binding (via embeddings) simultaneously.","intents":["Learn both structural and semantic aspects of a concept in a single training pass","Achieve higher quality subject-specific generation than DreamBooth or Textual Inversion alone","Create more robust concept representations that generalize across diverse prompts","Combine the benefits of parameter-efficient fine-tuning with embedding-based concept binding"],"best_for":["High-fidelity subject-specific generation where quality is critical","Practitioners with moderate computational budgets (16GB+ VRAM)","Use cases requiring both visual fidelity and semantic flexibility"],"limitations":["Training is 1.5-2× slower than DreamBooth alone due to dual optimization","Requires careful balancing of LoRA and embedding learning rates; suboptimal tuning causes mode collapse","Produces larger artifacts (LoRA + embedding combined, ~6-7MB)","Hyperparameter sensitivity is higher than single-method approaches"],"requires":["PyTorch 1.9+","16GB+ GPU VRAM","3-5 target images","100-200 class-prior images","Stable Diffusion base model"],"input_types":["Target image set (JPEG/PNG)","Class-prior image set (JPEG/PNG)","Text prompt template"],"output_types":["LoRA adapter weights","Learned embedding vector","Combined checkpoint"],"categories":["image-visual","model-training","hybrid-fine-tuning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_4","uri":"capability://data.processing.analysis.lora.weight.extraction.and.model.merging","name":"lora weight extraction and model merging","description":"Extracts trained LoRA matrices (A and B) from fine-tuned models via extract_lora_ups_down function, enabling separation of adaptation weights from base model. Supports merging LoRA weights back into the original model (collapse_lora) to create standalone checkpoints, or composing multiple LoRA adapters by stacking their low-rank updates. Handles both safetensors and CKPT formats.","intents":["Extract trained LoRA weights for distribution or storage without base model","Merge LoRA adapters into base models to create standalone inference checkpoints","Combine multiple LoRA adapters for compositional generation (e.g., style + subject)","Convert between model formats (CKPT ↔ safetensors) while preserving LoRA weights"],"best_for":["Model distribution platforms (Hugging Face Hub, Civitai) sharing adapters","Production inference pipelines requiring merged checkpoints","Researchers studying model composition and adapter interaction"],"limitations":["Merging is irreversible—once collapsed, LoRA weights cannot be extracted again","Composing >3-4 LoRA adapters causes cumulative latency (~50-100ms per adapter) and potential numerical instability","Format conversion (CKPT ↔ safetensors) requires full model loading; memory-intensive for large models","No built-in conflict detection when merging adapters targeting overlapping layers"],"requires":["PyTorch 1.9+","Trained LoRA checkpoint or base model with injected LoRA layers","safetensors library (optional, for format support)","8GB+ RAM for model loading"],"input_types":["LoRA checkpoint (PyTorch .pt or safetensors format)","Base model checkpoint (CKPT or safetensors)","Multiple LoRA adapters (for composition)"],"output_types":["Extracted LoRA weight matrices (A, B tensors)","Merged checkpoint (standalone model)","Composed adapter stack"],"categories":["data-processing-analysis","model-training","model-composition"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_5","uri":"capability://image.visual.inpainting.specific.fine.tuning.with.mask.conditioning","name":"inpainting-specific fine-tuning with mask conditioning","description":"Extends LoRA training to inpainting tasks by conditioning the diffusion model on binary masks that specify regions to regenerate. During training, the model receives concatenated input (image + mask) and learns to denoise only masked regions while preserving unmasked content. Implements mask-aware loss weighting to focus gradient updates on inpainted areas.","intents":["Fine-tune models for object removal, replacement, or region-specific editing","Learn inpainting-specific adaptations for particular object categories or styles","Create specialized inpainting models for domain-specific use cases (e.g., furniture, faces)","Improve inpainting quality on custom datasets with minimal training overhead"],"best_for":["Content creation tools requiring domain-specific inpainting","E-commerce platforms automating product image editing","Researchers studying conditional diffusion and mask-aware generation"],"limitations":["Mask quality directly impacts training stability; noisy or ambiguous masks degrade results","Model may learn to ignore masks if loss weighting is not carefully tuned","Inpainting quality is sensitive to mask-to-image ratio; extreme aspect ratios cause artifacts","Requires paired image-mask datasets; automatic mask generation often introduces bias"],"requires":["PyTorch 1.9+","Training dataset with paired images and binary masks","8GB+ GPU VRAM","Stable Diffusion inpainting model variant"],"input_types":["Image set (JPEG/PNG, 512×512)","Binary mask set (same dimensions as images)","Text prompts describing desired inpainted content"],"output_types":["LoRA adapter for inpainting","Inpainted image outputs"],"categories":["image-visual","model-training","conditional-generation"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_6","uri":"capability://image.visual.face.specific.conditioning.and.identity.preservation","name":"face-specific conditioning and identity preservation","description":"Implements face-aware fine-tuning by integrating face detection and embedding extraction (via face recognition models) to preserve identity consistency across generated images. During training, the model receives face embeddings as additional conditioning signals, and the loss function includes face similarity terms to ensure generated faces maintain identity characteristics of the training set.","intents":["Generate consistent portraits of a specific person across diverse poses and expressions","Fine-tune models for face-specific applications (headshots, avatars, ID photos)","Preserve facial identity while varying background, clothing, or other attributes","Create identity-aware adapters for personalized portrait generation"],"best_for":["Portrait and headshot generation services","Avatar creation platforms requiring consistent identity","Researchers studying identity preservation in generative models"],"limitations":["Requires face detection and embedding extraction; fails on non-frontal or occluded faces","Face embedding models introduce their own biases; identity preservation quality depends on embedding model quality","Training is slower due to per-image face embedding computation","Cannot preserve identity across extreme pose changes or with multiple faces in frame"],"requires":["PyTorch 1.9+","Face detection model (e.g., RetinaFace, MTCNN)","Face embedding model (e.g., ArcFace, FaceNet)","Training dataset with clear, frontal face images","16GB+ GPU VRAM"],"input_types":["Face image set (JPEG/PNG, 512×512, frontal-facing preferred)","Text prompts for generation"],"output_types":["Face-conditioned LoRA adapter","Generated portrait images with preserved identity"],"categories":["image-visual","model-training","biometric-conditioning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_7","uri":"capability://code.generation.editing.xformers.memory.efficient.attention.integration","name":"xformers memory-efficient attention integration","description":"Integrates XFormers library's optimized attention implementations (flash attention, memory-efficient attention) into the diffusion model's forward pass to reduce peak memory usage and accelerate training. Automatically replaces standard PyTorch attention with XFormers kernels when available, reducing attention complexity from O(n²) to O(n) in sequence length.","intents":["Train LoRA adapters on GPUs with limited VRAM (6-8GB) that would otherwise run out of memory","Accelerate training by 20-40% through optimized attention kernels","Enable higher batch sizes or longer context without additional hardware","Reduce inference latency for real-time image generation applications"],"best_for":["Practitioners with consumer-grade GPUs (RTX 3060, RTX 4060)","Cost-sensitive production deployments requiring inference optimization","Research teams maximizing throughput on limited GPU budgets"],"limitations":["XFormers requires CUDA 11.6+ and specific GPU architectures (Ampere, Turing); not available on CPU or older GPUs","Memory savings are model-dependent; benefits diminish for small batch sizes or short sequences","XFormers kernels may produce slightly different numerical results due to different precision handling","Installation can be complex; precompiled wheels not available for all CUDA versions"],"requires":["PyTorch 1.12+","CUDA 11.6+","XFormers library (pip install xformers)","Compatible GPU (Ampere, Turing, or newer NVIDIA architecture)"],"input_types":["Diffusion model with standard attention layers"],"output_types":["Model with XFormers attention kernels enabled"],"categories":["code-generation-editing","automation-workflow","performance-optimization"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_8","uri":"capability://data.processing.analysis.ckpt.and.safetensors.format.conversion.and.compatibility","name":"ckpt and safetensors format conversion and compatibility","description":"Provides bidirectional conversion between CKPT (PyTorch pickle format) and safetensors (safe tensor format) while preserving LoRA weights and metadata. Handles format detection, weight mapping, and safe serialization to prevent data corruption. Supports loading models in either format and saving to preferred format without manual intervention.","intents":["Convert legacy CKPT models to safetensors for safer distribution and faster loading","Load models from community repositories (Civitai, Hugging Face) regardless of format","Ensure compatibility across different tools and frameworks that prefer different formats","Safely archive models without pickle security risks"],"best_for":["Model distribution platforms and community repositories","Production systems requiring format standardization","Security-conscious teams avoiding pickle deserialization"],"limitations":["Conversion requires full model loading into memory; infeasible for models >50GB","Some custom metadata may be lost during conversion if not explicitly mapped","Safetensors format does not support all PyTorch data types (e.g., complex numbers); conversion may fail for exotic models","Conversion is one-way for some metadata; cannot perfectly reconstruct original CKPT from safetensors"],"requires":["PyTorch 1.9+","safetensors library","8GB+ RAM for model loading","Source model file (CKPT or safetensors)"],"input_types":["CKPT checkpoint file","Safetensors checkpoint file"],"output_types":["Converted checkpoint in target format","Metadata mapping (optional)"],"categories":["data-processing-analysis","automation-workflow","format-conversion"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"github-cloneofsimo--lora__cap_9","uri":"capability://automation.workflow.command.line.training.orchestration.and.hyperparameter.management","name":"command-line training orchestration and hyperparameter management","description":"Provides CLI tools (lora_pti, lora_add, lora_distill, lora_ppim) that abstract training complexity into simple command-line interfaces with YAML/JSON configuration files. Handles argument parsing, configuration validation, checkpoint management, and training loop orchestration without requiring code modification. Supports distributed training setup and experiment tracking.","intents":["Train LoRA adapters without writing Python code via CLI commands","Manage hyperparameters and training configurations via config files","Reproduce experiments by sharing configuration files","Orchestrate multi-stage training pipelines (preprocessing → training → merging)"],"best_for":["Non-technical users and content creators","Teams standardizing training workflows across practitioners","Researchers managing large numbers of experiments with varying hyperparameters"],"limitations":["CLI abstractions hide implementation details; debugging training failures requires understanding underlying code","Configuration files can become complex for advanced use cases; limited validation of invalid hyperparameter combinations","CLI tools are less flexible than programmatic API; custom loss functions or training loops require code modification","Error messages from CLI may be cryptic; requires familiarity with underlying libraries"],"requires":["Python 3.7+","PyTorch 1.9+","lora package installed (pip install lora)","YAML or JSON configuration file"],"input_types":["Configuration file (YAML/JSON)","Training dataset (images)","Base model checkpoint"],"output_types":["Trained LoRA checkpoint","Training logs and metrics","Merged model (optional)"],"categories":["automation-workflow","tool-use-integration","experiment-management"],"confidence":0.5,"matches":0,"success_rate":0}],"trust":{"score":31,"verified":false,"data_access_risk":"low","permissions":["PyTorch 1.9+","Hugging Face diffusers library","GPU with minimum 8GB VRAM for training (6GB for inference)","Original Stable Diffusion base model weights","3-5 target images of the subject (512×512 or higher)","100-200 class-prior images (can be auto-generated or sourced)","8GB+ GPU VRAM","Stable Diffusion base model","Multiple trained LoRA checkpoints","Base model compatible with all LoRA adapters"],"failure_modes":["Rank parameter d must be manually tuned; too low loses expressiveness, too high negates efficiency gains","Cannot capture structural changes to model behavior—only fine-grained weight adjustments","Requires careful initialization of A and B matrices to avoid training instability","Performance gains diminish when combining >3-4 LoRA adapters simultaneously due to cumulative latency","Requires manual collection of 3-5 high-quality target images; poor image quality degrades results","Prior-preservation class images must be semantically similar but visually distinct; mismatched classes cause mode collapse","Training is sensitive to hyperparameters (learning rate, prior loss weight); requires experimentation per subject","Cannot learn complex multi-subject compositions or fine-grained attribute combinations","Composing >3-4 LoRA adapters causes cumulative latency (~50-100ms per adapter) and potential numerical instability","Interpolation assumes LoRA models are trained on compatible base models; mismatched bases produce artifacts","builder identity is not verified yet","no observed match outcomes yet"],"rank_breakdown":{"adoption":0.3344711490835181,"quality":0.25,"ecosystem":0.55,"match_graph":0.25,"freshness":0.27,"weights":{"adoption":0.35,"quality":0.2,"ecosystem":0.1,"match_graph":0.3,"freshness":0.05}},"observed_outcomes":{"matches":0,"success_rate":0,"avg_confidence":0,"top_intents":[],"last_matched_at":null},"maintenance":{"status":"inactive","updated_at":"2026-05-06T15:12:23.810Z","last_scraped_at":"2026-05-03T13:58:42.319Z","last_commit":"2024-03-22T03:48:10Z"},"community":{"stars":7536,"forks":495,"weekly_downloads":null,"model_downloads":null,"model_likes":null}},"distribution":{"claim_url":"https://unfragile.ai/submit?claim=cloneofsimo--lora","compare_url":"https://unfragile.ai/compare?artifact=cloneofsimo--lora"}},"signature":"nU7BCwDzIN+c7Vv2Ps2YR9f9HYfi9rBRmWG9k+BUvYLDrt6Yd/FvW/hDd3yb4GrpzD+ql+8ZwJgcrH1C95t8AA==","signedAt":"2026-06-20T18:41:04.864Z","signedBy":"unfragile.ai","version":1},"_links":{"self":"https://unfragile.ai/api/v1/passport/cloneofsimo--lora","artifact":"https://unfragile.ai/cloneofsimo--lora","verify":"https://unfragile.ai/api/v1/verify?slug=cloneofsimo--lora","publicKey":"https://unfragile.ai/api/v1/trust-passport-public-key","spec":"https://unfragile.ai/trust","schema":"https://unfragile.ai/schema.json","docs":"https://unfragile.ai/docs"}}