jaahas/tarmomatic

All of the necessary models to run tarmomatic local edition with ComfyUI. You also need ComfyUI-GGUF custom nodes. Place all of the models in their respective folders in the ComfyUI models-folder.

Installation

1. Go to your ComfyUI directory
2. Download with HF CLI (or git):

curl -LsSf https://hf.co/cli/install.sh | bash
hf download jaahas/tarmomatic --local-dir models

All Models List

Model Filename	Required ComfyUI Folder	Used In
flux1-schnell-fp8.safetensors	models/checkpoints	Flux
flux1-schnell-Q4_K_M.gguf	models/unet	Flux
qwen_2.5_vl_7b_fp8_scaled.safetensors	models/text_encoders	Qwen Image Edit
qwen_image_vae.safetensors	models/vae	Qwen Image Edit
Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors	models/loras	Qwen Image Edit
Qwen-Image-Edit-2509-Q4_K_M.gguf	models/unet	Qwen Image Edit
t5xxl_fp8_e4m3fn_scaled.safetensors	models/text_encoders	LTX Video
ltxv-2b-0.9.8-distilled-fp8.safetensors	models/checkpoints	LTX Video
umt5_xxl_fp8_e4m3fn_scaled.safetensors	models/text_encoders	Wan Video
wan2.2_vae.safetensors	models/vae	Wan Video
Wan2.2-TI2V-5B-Q5_K_M.gguf	models/unet	Wan Video

---

Benchmarks (RTX 5090, cold start)

Model	Speed
Flux Schnell Q4_K_M (1024x1024)	28s
Qwen Image Edit 2509 Q4_K_M Lightning (1 image, 1024x1024)	112s
Wan 2.2 TI2V 5B Q5_K_M (10s, 720p)	460s
Wan 2.2 TI2V 5B Q5_K_M (10s, 720p, optimised)	148s
LTXV 2b 0.9.8 distilled fp8 (10s, 512p)	47s
TBA	---
Wan 2.2 I2V A14B Q5_K_M Lightning (10s, 720p)	1074s
Wan 2.2 I2V A14B Q5_K_M Lightning (10s, 480p)	296s
Eigen Banana Qwen Image Edit 2509 Q4_K_M (1 image, 1024x1024)	151s

---

Flux Models

Used for general image generation (Workflow: flux_schnell-GGUF.json).

• flux1-schnell-fp8.safetensors
  • Folder: models/checkpoints
  • Note: This provides the CLIP (Text Encoder) and VAE for the workflow.
• flux1-schnell-Q4_K_M.gguf
  • Folder: models/unet
  • Note: This provides the actual diffusion model (UNet) in a compressed (quantized) format for better performance.

Qwen Image Models

Used for image editing and synthesis (Workflows: image_qwen_image_edit_2509-GGUF-*.json).

• qwen_2.5_vl_7b_fp8_scaled.safetensors
  • Folder: models/text_encoders
• qwen_image_vae.safetensors
  • Folder: models/vae
• Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors
  • Folder: models/loras
• Qwen-Image-Edit-2509-Q4_K_M.gguf
  • Folder: models/unet

LTX Models

Used for image-to-video generation (Workflow: ltxv_image_to_video.json).

• t5xxl_fp8_e4m3fn_scaled.safetensors
  • Folder: models/text_encoders
• ltxv-2b-0.9.8-distilled-fp8.safetensors
  • Folder: models/checkpoints

Wan Models

Used for video generation (Workflow: video_wan2_2_5B_ti2v-GGUF.json).

• umt5_xxl_fp8_e4m3fn_scaled.safetensors
  • Folder: models/text_encoders
• wan2.2_vae.safetensors
  • Folder: models/vae
• Wan2.2-TI2V-5B-Q5_K_M.gguf
  • Folder: models/unet

---

FAQ

Why does Flux need both a GGUF and a Checkpoint?

The workflow uses a "hybrid" loading strategy:

1. Checkpoint (flux1-schnell-fp8.safetensors): Loads the CLIP (text understanding) and VAE (image decoding) components.
2. GGUF (flux1-schnell-Q4_K_M.gguf): Loads the UNet (image generation core). This setup allows you to use a highly compressed, fast UNet (GGUF) while still getting the necessary support components from the standard checkpoint.