TensorSort - The Model Organizer

"Because filenames lie, but tensors don't."

Who is this for?

Beginners:**
"Error loading LoRA" and no idea why?
→ TensorSort shows you what each file is.

Collectors:**
100+ files, duplicates somewhere, chaos?
→ TensorSort brings order.

Power Users:**
400+ models, grown organically, wildly named?
→ TensorSort gives you clean structure.

Your download folder looks like this?:

downloads/
├── aDetailedPurposeV2_v14e.safetensors        ← Checkpoint? LoRA? For SDXL?
├── flux_lora_style_v3_FINAL_REAL.safetensors  ← Really Flux? Or SDXL?
├── some_model_q8_0.gguf                       ← Is that Q8? Or Q4?
├── controlnet_thing_fp16.safetensors          ← ControlNet? T2I-Adapter?
├── pony_or_sdxl_idk_v2.safetensors            ← ???
└── ... 73 more

You don't know:

• What each file is (Checkpoint vs LoRA vs VAE vs ControlNet)

• Which model it was trained for (Flux vs SDXL vs SD1.5 vs Pony)

• Where it belongs in ComfyUI (checkpoints? unet? loras? controlnet?)

• If you already have it (Duplicates with different names?)

The usual "solutions"

**Sort manually?**
→ With 50+ files? Good luck.

**ComfyUI-Manager?**
→ Installs **Nodes**, not model organization.

**Model-Manager Extensions?**
→ GUI to **view** files. You still need to know what's what.

**Trust the filename?**
→ LOL. Filenames lie. Constantly.

TensorSort does it differently

TensorSort doesn't guess . It READS.

Every .safetensors and .gguf file has a header with metadata. TensorSort opens the file binary and reads this data directly:

┌─────────────────────────────────────────────────────────────────┐
│  FILE: random_lora_thing_v2.safetensors                         │
│─────────────────────────────────────────────────────────────────│
│                                                                 │
│  TensorSort reads Bytes 0-8:  Header-Size = 14,847 bytes        │
│  TensorSort parses Header:    JSON with 847 Tensor-Keys         │
│                                                                 │
│  FOUND KEYS:                                                    │
│  ├── lora_unet_double_blocks.0.img_attn.proj.lora_down.weight   │
│  ├── lora_unet_double_blocks.0.img_attn.proj.lora_up.weight     │
│  └── ... (844 more)                                             │
│                                                                 │
│  ANALYSIS:                                                      │
│  ✓ Keys with "lora_" prefix     → FILE TYPE: LoRA               │
│  ✓ Keys with "double_blocks"    → BASE MODEL: Flux              │
│  ✓ Tensor dtype: bfloat16       → PRECISION: BF16               │
│─────────────────────────────────────────────────────────────────│
│  RESULT:                                                        │
│  Name:   FluxD_LoRA-BF16_General_Thing_v2.safetensors           │
│  Folder: loras/                                                 │
└─────────────────────────────────────────────────────────────────┘

The filename said: "random_lora_thing"

The tensor keys say: "Flux LoRA in BF16"

Who's right? The keys.

This is where it gets really interesting.

GGUF files often have WRONG quantization in the filename . The uploader writes "Q8_0" but the file is actually Q4_K_M.

┌─────────────────────────────────────────────────────────────────┐
│  FILE: flux1-schnell-Q8_0-whatever.gguf                         │
│─────────────────────────────────────────────────────────────────│
│                                                                 │
│  TensorSort reads GGUF-Header:                                  │
│  ├── Magic Bytes: "GGUF"                                        │
│  ├── Version: 3                                                 │
│  ├── Tensor Count: 394                                          │
│  └── Metadata KV Count: 27                                      │
│                                                                 │
│  TensorSort reads Metadata:                                     │
│  └── general.file_type = 15                                     │
│                                                                 │
│  Mapping: 15 → Q4_K_M                                           │
│                                                                 │
│─────────────────────────────────────────────────────────────────│
│  FILENAME SAYS:    Q8_0                                         │
│  FILE SAYS:        Q4_K_M                                       │
│                                                                 │
│  The filename LIED.                                             │
│─────────────────────────────────────────────────────────────────│
│  RESULT:                                                        │
│  Name:   FluxS_GGUF-Q4_K_M_Whatever_v1.gguf                     │
│  Folder: unet/                                                  │
└─────────────────────────────────────────────────────────────────┘

What TensorSort Detects

File Types (from Tensor-Key-Patterns)

Type            | Detection Pattern                    | Target Folder
----------------|--------------------------------------|---------------
Checkpoint      | model.diffusion_model.*              | checkpoints/
Flux UNET       | double_blocks.*, single_blocks.*     | unet/
LoRA            | lora_up.weight, lora_down.weight     | loras/
LyCORIS         | hada_w1, lokr_w1                     | loras/LyCORIS/
VAE             | decoder.conv_in, encoder.conv_out    | vae/
ControlNet      | control_model.*                      | controlnet/
CLIP            | text_model.encoder.*                 | clip/
Embedding       | emb_params                           | embeddings/

Base Models (from Architecture-Patterns)

| Model | Detection Pattern |
|-------|-------------------|
| Flux Dev/Schnell | `double_blocks`, `single_blocks` |
| SDXL | `conditioner.embedders`, `model.diffusion_model` with SDXL-Shapes |
| Pony | SDXL-Architecture + Metadata-Check (ss_base_model, sd_merge_models) |
| SD 1.5 | `cond_stage_model.transformer` |
| Z-Image | `context_refiner`, `noise_refiner` |
| WAN Video | `diffusion_model.blocks.*.cross_attn` (without Flux blocks) |
| Qwen-Image-Edit | `model.diffusion_model.txt_in` (without Flux blocks) |
| Lotus-Depth | `class_embedding` + `down_blocks.*.attentions` |

Before → After

Your Download Folder (BEFORE)

downloads/
├── aDetailedPurposeV2_v14e.safetensors
├── flux-dev-fp16-some-finetune.safetensors
├── ponyDiffusionV6XL_v6StartWithThisOne.safetensors
├── some_lora_nsfw_v3_FINAL.safetensors
├── flux1-schnell-Q8.gguf
└── ... (41 more)

Your ComfyUI models/ Folder (AFTER)

models/
├── checkpoints/
│   ├── SDXL_Full-FP16_Realism_DetailedPurpose_v14.safetensors
│   └── Pony_Full-FP16_General_PonyDiffusion_v6.safetensors
│
├── unet/
│   ├── FluxD_Full-FP16_General_SomeFinetune_v1.safetensors
│   └── FluxS_GGUF-Q5_K_M_Schnell_v1.gguf  ← Was NOT Q8!
│
├── loras/
│   └── Flux_NSFW_SomeLora_Trig-nude_v3.safetensors
│
├── controlnet/
│   └── SDXL_CN-Canny_FP16_v2.safetensors
│
└── upscale_models/
    └── 4x_UltraSharp.pth

At a glance you can see: [/BOLD] What it is • Which model • Which precision • Name and version

The 15 Modules

| # | Module | Detects | Target Folder |
|---|--------|---------|---------------|
| 1 | Base Models | Checkpoints, UNET, GGUF, WAN, Lotus, Qwen-Edit | checkpoints/, unet/, diffusion_models/ |
| 2 | VAE | VAE, TAESD, VAE-Approx | vae/, vae_approx/ |
| 3 | Text Encoders | CLIP, T5, UMT5, Qwen3, BERT | clip/, text_encoders/ |
| 4 | LoRAs | LoRA, LoHa, LoKr, LoCon, WAN-LoRA | loras/, loras/LyCORIS/ |
| 5 | ControlNet | ControlNet, T2I-Adapter | controlnet/, t2i_adapter/ |
| 6 | Upscalers | ESRGAN, RealESRGAN, SwinIR | upscale_models/ |
| 7 | Embeddings | Textual Inversion | embeddings/ |
| 8 | PhotoMaker | PhotoMaker v1/v2 | photomaker/ |
| 9 | InsightFace | Face Analysis Models | insightface/ |
| 10 | IP-Adapter | SD/SDXL/Flux variants | ipadapter/, xlabs/ |
| 11 | AnimateDiff | Motion Modules | animatediff_models/ |
| 12 | SAM | Segment Anything | sams/ |
| 13 | GroundingDINO | Object Detection | grounding-dino/ |
| 14 | YOLO | Ultralytics Models | ultralytics/ |
| 15 | VLM & LLM | Vision/Language Models | VLM/, LLM/ |

The Two Modes

Mode A: Installation 

1. Put files in downloads/
2. Start TensorSort → Mode A
3. PREVIEW: See what TensorSort detected
4. Confirm
5. Files get sorted + renamed

Mode B: Cleanup 

1. Start TensorSort → Mode B
2. Scans ALL files in models/
3. Finds: Wrong folders, wrong names, duplicates
4. Shows what would be changed
5. You confirm → Chaos becomes order

Bonus Features

### Recursive Scanning

Already organized your downloads into subfolders? No problem!

### Duplicate Detection (SHA256)

TensorSort calculates a SHA256 hash for each file. Identical files are detected - **no matter what they're named**.

### Cross-Folder Rescue

File in the wrong folder? TensorSort finds it and moves it automatically:

```
checkpoints/some_lora.safetensors  ← LoRA in checkpoint folder!

→ TensorSort Mode B detects this
→ Moves to loras/Flux_LoRA-BF16_Some_v1.safetensors

Trigger-Word Extraction (LoRAs)

TensorSort reads `ss_tag_frequency` from LoRA metadata and adds the most frequent tag as trigger word in the name:

```
Before:  some_anatomy_lora_v3.safetensors
After:   Flux_Anatomy_SomeLora_Trig:nude_v3.safetensors
                                    ^^^^^^^^
                                    Extracted from metadata!

Quick Start

### Requirements
- ComfyUI installed (Standard OR Portable)
- That's it!

**ComfyUI Portable users:** No system Python needed! TensorSort automatically finds the `python_embeded` from your ComfyUI folder.

**First run without config?** TensorSort asks once for the ComfyUI path - completely without Python, pure batch logic.

On first start: Enter ComfyUI path → Config is created → Done!

ComfyUI caches filenames in the browser. After TensorSort changes:

**Ctrl + Shift + R** (Windows/Linux)
**Cmd + Shift + R** (Mac)

Download & Support

Author: K0DA Parallax Studio

Email: [email protected]

GitHub: https://github.com/K0DA-PARALLAXStudio

Patreon: https://patreon.com/K0DAParallaxStudio

Instagram: https://www.instagram.com/k0da_parallax_studio/

Buy Me Coffee: https://www.buymeacoffee.com/K0DA_Parallax_Studio

*"Because filenames lie, but tensors don't."*