---
language:
- en
- ru
- zh
base_model:
- ideogram-ai/ideogram-4-fp8
- ideogram-ai/ideogram-4-nf4
tags:
- gguf
- text-to-image
- image-generation
- comfyui
- stable-diffusion.cpp
pipeline_tag: text-to-image
license: other
license_name: ideogram-non-commercial-model-agreement
license_link: https://huggingface.co/ideogram-ai/ideogram-4-fp8/blob/main/LICENSE.md
---

# Ideogram4 GGUF quantized files

```tree
.
├── diffusion/
│   ├── cond/
│   │   ├── ideogram4_Q4_0.gguf
│   │   ├── ideogram4_Q4_1.gguf
│   │   ├── ideogram4-Q4_K.gguf
│   │   ├── ideogram4-Q5_0.gguf
│   │   ├── ideogram4_Q5_1.gguf
│   │   ├── ideogram4_Q5_K.gguf
│   │   ├── ideogram4-Q6_K.gguf
│   │   └── ideogram4-Q8_0.gguf
│   └── uncond/
│       ├── ideogram4_unconditional_Q4_0.gguf
│       ├── ideogram4_unconditional_Q4_1.gguf
│       ├── ideogram4_unconditional_Q4_K.gguf
│       ├── ideogram4_unconditional_Q5_0.gguf
│       ├── ideogram4_unconditional_Q5_1.gguf
│       ├── ideogram4_unconditional_Q5_K.gguf
│       ├── ideogram4_unconditional_Q6_K.gguf
│       └── ideogram4_unconditional-Q8_0.gguf
├── text_encoder/
│   ├── Qwen3-VL-8B-Q4_0.gguf
│   ├── Qwen3-VL-8B-Q4_1.gguf
│   ├── Qwen3-VL-8B-Q4_K_S.gguf
│   ├── Qwen3-VL-8B-Q4_K_M.gguf
│   ├── Qwen3-VL-8B-Q5_K_S.gguf
│   ├── Qwen3-VL-8B-Q5_K_M.gguf
│   ├── Qwen3-VL-8B-Q6_K.gguf
│   └── Qwen3-VL-8B-Q8_0.gguf
└── vae/
│   ├── flux2-vae.safetensors
│   └── flux2-hdr-vae.safetensors
└── lora/
    ├── realism_engine_v3.safetensors
    ├── big_boobs.safetensors
    ├── cum.safetensors
    ├── innie_vulva_x.safetensors
    ├── vintage_beauties_womans.safetensors
    ├── missionary_sex.safetensors
    ├── 80s_anime.safetensors
    ├── penis.safetensors
    └── penix.safetensors
```

### Model Selection & Quantization Guide

To balance generation quality, memory usage, and inference speed, we recommend the following quantization choices for each component:

#### 1. Conditional Diffusion Model (`diffusion/cond/`)
* **Recommended:** `Q6_K` or `Q8_0`
* Since this model handles the main conditional generation pass, keeping a higher quantization level is key to preserving detail and prompt adherence.

#### 2. Unconditional Diffusion Model (`diffusion/uncond/`)
* **Recommended:** `Q4_K` or `Q5_K`
* **Note:** Using `Q6_K` or `Q8_0` for the unconditional model is generally unnecessary (overkill) and may slow down generation without providing a noticeable improvement in quality.

#### 3. Text Encoder (`text_encoder/`)
* **Recommended:** `Q5_K_M` or `Q4_K_M`
* These medium-sized "K-measure" quants offer a good trade-off, retaining the text encoder's comprehension capabilities while fitting within reasonable memory limits.

---

### General Recommendations for Quantization Types

If you are optimizing for inference speed or trying to fit a specific model entirely into VRAM/RAM, keep these rules of thumb in mind:

* **Prefer `_K` variants over `_0` and `_1`:** When choosing between `Q4` or `Q5` options, always prefer the `_K` variants (e.g., `Q4_K_M`, `Q5_K_M`, or standard `_K`). 
* **Avoid `_0` and `_1` if possible:** The older `_0` and `_1` quants (like `Q4_0` or `Q4_1`) perform worse in terms of quality loss. While they are marginally smaller, the minor size reduction rarely justifies the drop in generation quality compared to `_K` equivalents.