pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF — new model trending #30 on Hugging Face
Community GGUF, FP8, and BF16 builds of the Qwen-Image-2.1 text encoder are trending on Hugging Face.
A community Hugging Face repo is trending with GGUF, FP8, and BF16 builds of a Heretic Qwen3-VL 8B text encoder for Qwen-Image-2.1. The Q4_K_M GGUF is 5.0 GB and needs a 1.2 GB mmproj; FP8 and BF16 safetensors are 9.3 GB and 17.5 GB. The author says a ComfyUI-GGUF add-on fixes vision-tower loading and unknown-architecture errors so the text encoder and DiT can both run as GGUF. Suggested settings are 25 steps, using CFG 1.0 or a mid-run switch to CFG 3.0 for readable text.
- Trending #30: Heretic Qwen3-VL 8B text-encoder weights for Qwen-Image-2.1.
- GGUF Q4_K_M is 5.0 GB plus a 1.2 GB mmproj; FP8 is 9.3 GB and BF16 is 17.5 GB.
- A ComfyUI add-on loads the Qwen3-VL vision tower so GGUF text-encoder and DiT files run.
- Recommended sampling is 25 steps, with CFG split from 1.0 to 3.0 when rendering text.
Full article1,499 words · extracted from huggingface.co · click to collapse
# Qwen-Image-2.1 Text Encoder (Heretic) — GGUF · FP8 · bf16
> [!IMPORTANT]
> **Got `got input of size [1, 512, 12288]` with the GGUF? It's fixed.**
> The GGUF files were fine all along — ComfyUI-GGUF just didn't load the Qwen3-VL vision tower.
> Install the add-on node **[ComfyUI-GGUF-Qwen3VL-TE](https://github.com/pottokao-dotcom/ComfyUI-GGUF-Qwen3VL-TE)**,
> make sure the `mmproj` file sits next to the GGUF, restart ComfyUI. No need to re-download anything.
> The same patch also fixes `Unknown model architecture!` for Qwen-Image-2.1 **DiT** GGUFs — so the whole
> pipeline can run on GGUF.
## 📥 Which file should I use?
| Your setup | File(s) | Size | Loader |
|---|---|---:|---|
| **Smallest** (GGUF) | `qwen3vl_8b_heretic-Q4_K_M.gguf` **+** `mmproj-qwen3vl_8b_heretic-f16.gguf` | 5.0 + 1.2 GB | `CLIPLoaderGGUF` + [add-on node](https://github.com/pottokao-dotcom/ComfyUI-GGUF-Qwen3VL-TE) |
| **NVIDIA GPU**, no extra nodes | `qwen3vl_8b_fp8_heretic.safetensors` | 9.3 GB | stock `CLIPLoader` |
| **Mac / non-CUDA**, no extra nodes | `qwen3vl_8b_bf16_heretic.safetensors` | 17.5 GB | stock `CLIPLoader` |
All three are in **this repo**. Every loader uses **type `qwen_image`** and feeds `TextEncodeQwenImage21`.
## ⚙️ Recommended sampler settings
Qwen-Image-2.1 is a full (non-distilled) model. **Use 25 steps** — more steps don't make text sharper
and start to burn the highlights.
- **No text in the image:** plain `KSampler`, **cfg 1.0** (fastest, softest look; the negative prompt is ignored at cfg 1).
- **Text in the image (signs, labels, posters):** switch cfg **partway through** — the first ½–⅔ of the steps
at **cfg 1** lock in composition and materials, the rest at **cfg 3** with a negative prompt redraws the
lettering. You get cfg 1's look *and* cfg 3's readable text, in one pass with one seed.
In ComfyUI that's two chained `KSamplerAdvanced` nodes (same model, seed and steps):
| | `add_noise` | `steps` | `cfg` | `start_at_step` | `end_at_step` | `return_with_leftover_noise` |
|---|---|---:|---:|---:|---:|---|
| 1st | enable | 25 | 1.0 | 0 | **12–17** | enable |
| 2nd | disable | 25 | 3.0 | **12–17** | 25 | disable |
Split later (17) keeps more of the cfg 1 look; split earlier (12) gives crisper text.
Example negative prompt: `oversaturated, overexposed, gibberish text`.
## 🍱 All-GGUF Qwen-Image-2.1
| Part | Where | Loader |