Local Models Overview
OpenTryOn provides adapters for local inference models that run directly on your hardware. Unlike the cloud API adapters in tryon.api, these models require local GPU resources but offer several advantages:
- No API costs: Run unlimited inferences without per-request charges
- Privacy: Your data never leaves your machine
- Customization: Fine-tune models for your specific use case
- Offline capability: Work without internet connectivity
Adding a new local or API model? Start with Model Integration Guidelines.
Available Models
| Model | Type | VRAM Required | Speed | Use Case |
|---|---|---|---|---|
| FLUX.2-dev Turbo | Image Generation | 12GB+ | 6x faster | Fast text-to-image, image-to-image |
| FLUX 3 Image Klein | Image Generation / Edit | RTX 3090+ class (see BFL) | ~16-step distilled | Open FLUX.3 Klein; API twin flux3-image-klein |
| Clad body (local) | Body measurements (measure) | CPU (numpy paths) | Fast | ISO 8559-1 on Anny/MHR params; photo sizing via Clad API |
| Kimi-VL | Image/Video Understanding | 24GB+ | - | Open-weight counterpart to the Kimi K2.6/K2.7 Code APIs |
| Qwen3.8 | Image/Video Understanding | ~50GB+ (bf16) | - | Open Qwen3.8-27B: multimodal understand + thinking; counterpart to DashScope Max |
| Hy4 preview | Text LLM (vLLM/SGLang) | Datacenter (FP8 ~770GB+, TP=8) | - | Open tencent/Hy4-preview; OpenTryOn calls localhost OpenAI API — not in-process |
| Qwen-Image | Image Generation / Edit / VTON | ~40GB+ (bf16; offload default) | - | Open Qwen-Image-2512 T2I + Edit-2511 I2I; counterpart to DashScope qwen-image |
| LTX-2.5 | Video Generation | 16GB+ (24GB+ preferred) | Distilled few-step | Local T2V / I2V with synced audio |
| Cosmos 3 Nano | Video Generation | Large (16B, BF16 only; offload optional) | Heavy | Local T2V / I2V; self-hosted twin of NIM cosmos3 |
| MiniMax H3 | Video Generation | 80GB+ preferred (offload; ~75GB host RAM if int8) | Heavy | Local T2V / I2V with stereo audio (768p base) |
| Wan 2.2 | Video Generation | ~12GB+ (TI2V-5B) | Moderate | Local T2V / I2V open weights (Wan 3.0 is API-only) |
| Leffa | Virtual Try-On | 12GB+ recommended | ~6s on A100 | Dedicated local VTON (CVPR 2025; MIT code) |
| FASHN VTON v1.5 | Virtual Try-On | Consumer GPU (VRAM not published) | Moderate | Maskless pixel-space VTON, Apache-2.0 code + weights; model-worn or flat-lay garments |
| CatVTON | Virtual Try-On | <8GB @ 1024×768 | Moderate | Concatenation VTON (ICLR 2025; CC BY-NC-SA) |
| DeepSeek-VL2 | Image/Video Understanding | Single-GPU (tiny) to 40GB+ (small/base) | - | Open DeepSeek-VL2; no first-party hosted API exists. Needs DeepSeek's own deepseek_vl2 package (not on PyPI) |
| LightOnOCR-3 | Document OCR (parse) | 0.8B ~2 GB BF16; transformers>=5.2 | Fast | Page images and PDFs to markdown, optional labelled blocks; size charts, tech sheets. Apache-2.0, no hosted API |
| DeepSeek-OCR | Document/Image OCR | Requires flash-attn + CUDA | - | "Contexts Optical Compression"; care labels, size tags, SKU sheets. No first-party hosted API exists |
| Ternary Bonsai 2 27B | Image/Text Understanding | 5.9-8.5GB on disk (ternary-quantized) | Fast (CPU/Metal-friendly) | OpenAI-compatible client for PrismML's own llama.cpp/MLX server — no torch needed on the OpenTryOn side |
| Liquid d1 | Decision model (decide) | d1-3B ~3B / d1-omni-600M ~0.6B | Fast (1 forward pass) | Calibrated yes/no, choice, score answers about text, images (and speech for omni); needs transformers>=5.15 |
| VideoJev | Video decision model (decide) | 4B / 35B-A3B MoE | Fast (direct scoring) | Temporal grounding + multiple-choice video QA; local only (no hosted API); Apache-2.0 |
| EmbeddingGemma 2 | Multimodal embeddings (embed) | 270M-740M by modality | Fast | One 768-d space for text, code, image, video, audio; Apache-2.0 tag |
| FLUX.2-klein-4B try-on | Virtual Try-On | ~13GB (offload) | Fast (4 steps) | Apache-2.0 base + LoRA (commercially clean); person + top + bottom, clothing-agnostic person input; v1 community LoRA |
| Try-on safety classifier | Moderation (classify) | ~0.9GB (SigLIP2-so400m) | Fast | Gate garments before try-on: ok / revealing / lingerie / adult; Apache-2.0 |
| Fashion Florence | Garment tagging (understand) | 0.77B (CPU or GPU) | Fast | Florence-2 LoRA → JSON category / colour / material / style tags; MIT |
| Extract-Outfit | Garment extraction (edit) | Qwen-Image-Edit-2511 class (~40GB+ bf16, offload) | Moderate | Worn outfit photo → flat studio mockup; Apache-2.0 LoRA |
| Fashion embeddings | Fashion retrieval (embed) | ~0.2-0.4B (consumer GPU / CPU) | Fast | Marqo-FashionSigLIP, MODA, ZooClaw-FashionSigLIP2: catalogue search by text or photo |
| pplx-embed-v2-late | Multi-vector retrieval (embed) | 0.6B (340M active) / 9B (7.4B active) | Moderate | ColBERT-style MaxSim over text, images and pages; MIT; no hosted API |
| Limite 1B - Violetto | Text Math Reasoning (out of fashion/media scope) | ~2.1GB | Fast | Text-only 1B math model; named-model-only, not a fashion/media fit — see the page for why it's here |
Requirements
Most local models require:
- CUDA-capable GPU (NVIDIA recommended)
- PyTorch 2.1+ with CUDA support
- diffusers >= 0.29.0
- transformers
Exception: Ternary Bonsai 2 27B and Hy4 preview
are OpenAI-compatible HTTP clients for a server you run yourself (locally or
on your own cluster) — they need no opentryon[local] extra or GPU on the
machine running OpenTryOn itself.
- accelerate
VRAM Considerations
Local models are memory-intensive. FLUX.2-dev Turbo supports automatic model selection based on available VRAM:
| Available VRAM | Model Selection |
|---|---|
| ≥64GB | Full precision model |
| ≥48GB | 8-bit quantized |
| ≥38GB | 4-bit quantized |
| <38GB | 4-bit quantized (with warnings) |
Quick Start
from tryon.models import Flux2TurboAdapter
# Initialize (auto-selects model based on VRAM)
adapter = Flux2TurboAdapter()
# Generate image
images = adapter.generate_text_to_image(
prompt="A fashion model wearing an elegant dress",
width=1024,
height=1024
)
images[0].save("output.png")
Installation
Install the required dependencies:
pip install diffusers>=0.29.0 transformers accelerate torch
# For quantized models (lower VRAM requirements)
pip install bitsandbytes
Memory Optimization Tips
-
Enable CPU Offloading: For GPUs with limited VRAM
adapter = Flux2TurboAdapter(enable_cpu_offload=True) -
Use Attention Slicing: Reduces peak memory at slight speed cost
adapter = Flux2TurboAdapter(enable_attention_slicing=True) -
Lower Resolution: Start with smaller images for testing
images = adapter.generate_text_to_image(prompt="...", width=512, height=512) -
Clear CUDA Cache: Between generations
import torch
torch.cuda.empty_cache()