> ## Documentation Index
> Fetch the complete documentation index at: https://comfyart.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Models

> Which model files each ComfyArt feature needs and where they go.

Every canvas action runs one of the API workflows in `workflows/` (or a small graph built in code) through ComfyUI's `/prompt`. You only need the models for the features you use. The model picker in the generator shows which image and video models are installed.

Folders below are relative to `ComfyUI/models/`.

## Image generation and editing

The default image model. It powers generation, quick edit, edit text, expand, inpaint, style transfer, comment edits and creative upscale.

| File | Folder |
| - | - |
| `qwen_image_2.1_int8_convrot.safetensors` | `diffusion_models` |
| `qwen3vl_8b_int8_convrot.safetensors` | `text_encoders` |
| `qwen_image_2.1_vae_bf16.safetensors` | `vae` |

### Other image models (optional)

| Model | Files |
| - | - |
| Z-Image Turbo | `z_image_turbo_bf16` (`diffusion_models`), `qwen_3_4b` (`text_encoders`), `ae` (`vae`) |
| Flux.2 Klein 9B | `flux-2-klein-9b-nvfp4` (`diffusion_models`), `qwen_3_8b_fp8mixed` (`text_encoders`), `flux2-vae` (`vae`) |
| Qwen Image 2512 | `qwen_image_2512_fp8_e4m3fn` (`diffusion_models`), `qwen_2.5_vl_7b_fp8_scaled` (`text_encoders`), `qwen_image_vae` (`vae`) |
| Ideogram 4 | `ideogram4_fp8_scaled` and `ideogram4_unconditional_fp8_scaled` (`diffusion_models`), `qwen3vl_8b_fp8_scaled` (`text_encoders`), `flux2-vae` (`vae`) |

<Note>
  Qwen Image 2512 is loaded with the **ComfyArt Precise UNET Loader** node. ComfyUI's `--fast fp8_matrix_mult` turns unscaled fp8 checkpoints like this one into noise; the loader skips that path for the model it loads. Ideogram 4 expects a structured caption, which the default [Agent](/guides/agent) writes for you.
</Note>

## Image tools

| Feature | Files | Folder |
| - | - | - |
| Upscale, video upscale | `4x-UltraSharp.pth` | `upscale_models` |
| Remove background | `birefnet.safetensors` | `background_removal` |
| Eraser, move object, split layers background fill | `big-lama.pt`, needs [comfyui-inpaint-nodes](https://github.com/Acly/comfyui-inpaint-nodes) | `inpaint` |
| Smart select (inpaint, move object, custom split) | `sam3.1_multiplex_fp16.safetensors` | `checkpoints` |
| Multi-angle | `qwen_image_edit_2511_fp8mixed` (`diffusion_models`), `qwen-image-edit-2511-multiple-angles-lora` and `Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16` (`loras`), `qwen_2.5_vl_7b_fp8_scaled` (`text_encoders`), `qwen_image_vae` (`vae`) | |
| Split layers | `qwen_image_layered_bf16` (`diffusion_models`), `qwen_image_layered_vae` (`vae`), `qwen_2.5_vl_7b_fp8_scaled` (`text_encoders`) | |

## Video

Animate image, video generation and video extend use MiniMax H3:

| File | Folder |
| - | - |
| `minimax_h3_fl2va_pruned_int8_convrot.safetensors` | `diffusion_models` |
| `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` | `loras` |
| `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors` | `text_encoders` |
| `minimax_h3_video_vae_fp16.safetensors`, `minimax_h3_audio_vae_fp32.safetensors` | `vae` |

Video trim and the video editor use core nodes only.

## Speech and text recognition

`gemma4_e4b_it_fp8_scaled.safetensors` in `text_encoders` powers voice input and, by default, text recognition for Edit Text and Extract Text. Text recognition can use the Agent instead; see [Agent settings](/guides/agent#settings).

## Reference modes

Line art, depth, pose and the other structure modes use the [comfyui\_controlnet\_aux](https://github.com/Fannovel16/comfyui_controlnet_aux) preprocessors. MLSD, HED, normal and segmentation download their annotator weights on first use. With Qwen Image 2512, structure modes also use `Qwen-Image-2512-Fun-Controlnet-Union-2602.safetensors` in `controlnet`.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.