pottokao/Qwen-Image-2.1-PE-T2I-Heretic-GGUF

Qwen-Image-2.1-PE-T2I — Heretic — GGUF

Not affiliated with, or endorsed by, Alibaba / Qwen. Community derivative of `Qwen/Qwen-Image-2.1-PE-T2I`, redistributed under the Qwen Research License (copy included as LICENSE, per §3.a). Non-commercial use only; commercial use needs a separate licence from Qwen.

Q4_K_M build of `pottokao/Qwen-Image-2.1-PE-T2I-Heretic` — the Qwen-Image-2.1 prompt rewriter with refusal behaviour removed (3/100 refusals @ KL 0.036, from 98/100 — best of a 1,600-trial search; dominates the earlier 7/100 @ 0.039, tag v1-trial384).

`system_prompt.txt` is included and the model is useless without it. That 10 KB document is the output contract — structure, register, length, and the JSON shape. Load it as the system message.

Made with a simple node

The sample / showcase images in this repo were generated with the technique behind [ComfyUI-QwenImage-PhotoStyles](https://github.com/pottokao-dotcom/ComfyUI-QwenImage-PhotoStyles) — a small ComfyUI node (17 photographic styles; one short prompt becomes a full styled prompt via the PE-T2I rewriter). If you're interested in how the images are made, the node and its recipe are there.

What this is for

A short request in any language goes in; one JSON line comes out with a detailed English image prompt and a recommended aspect ratio:

{"rewritten_prompt": "...", "wh_ratio": "3:2"}

It is an LLM, not a text encoder. In ComfyUI it belongs in an LLM/GGUF node that rewrites your prompt before it reaches the DiT — it does not load into CLIPLoader. For the Qwen-Image-2.1 text encoder, see `…-Text-Encoder-Heretic`.

Available quantizations

All load the same way (llama-server -m <file>); pick by available VRAM.

One deviation from stock Q4KM

token_embd is kept at Q6_K instead of Q4K. Everything else is standard Q4KM (llama.cpp already promotes `output` to Q6K on its own).

The reason: every official NVIDIA NVFP4 checkpoint for this model family leaves the token embedding untouched — across Qwen3-8B, Qwen3-32B, Qwen3-30B-A3B, Qwen3.5-122B-A10B and Qwen3.5-397B-A17B, and also in the one official mixed-precision recipe (Qwen3.6-27B), which quantizes lm_head to 4-bit but still does not touch the embedding. Stock Q4_K_M does compress it. Raising it back costs 0.25 GB and cannot be worse than leaving it at Q4_K.

Run it

llama-server -m pe_t2i_heretic-Q4_K_M.gguf --host 0.0.0.0 --port 8080 \
             -ngl 99 -c 16384 --jinja -a pe-t2i

Then POST to /v1/chat/completions with system_prompt.txt as the system message. Loads in a few seconds; fits on an 8 GB card.

Conversion

python convert_hf_to_gguf.py <bf16 model> --outfile pe_bf16.gguf --outtype bf16 --no-mtp
llama-quantize --token-embedding-type q6_K --output-tensor-type q6_K \
               pe_bf16.gguf pe_t2i_heretic-Q4_K_M.gguf Q4_K_M

`--no-mtp` is required. This model's config.json carries text_config.mtp_num_hidden_layers with the value null. Without the flag the converter does int += None and dies with a TypeError. The key existing with a null value is the trap — .get(key, 0) returns None, not the default.

What was checked, and what wasn't

Checked — it still does its job. Four requests (three Chinese, one English) through llama-server: 4/4 produced strictly valid JSON with exactly the rewritten_prompt / wh_ratio keys, 2053–3351 characters. The bf16 source produces 2450–3320 on the same style of input, so the length discipline survives quantization.

Not checked — a clean perplexity comparison. I ran one and it returned self-contradictory numbers (the unquantized bf16 GGUF scored worse than its own quantizations, which is not physically sensible), so I am not reporting it. Take the functional check above as the evidence, and nothing more.

Tried and not shipped — an imatrix build. I generated an importance matrix from in-domain data and quantized with it. It passed the same functional check, but the calibration corpus was badly designed: the 10 KB system prompt appears in every sample, so ~88% of the 341 KB corpus was one repeated document, which would dominate the importance statistics. Rather than ship a build I can't justify, it's left out. A properly diversified corpus would be the way to revisit this.

🖼️ Showcase

<table> <tr><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/01.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/01.jpg" width="100%"></a><br>_{Warm Documentary}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/02.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/02.jpg" width="100%"></a><br>_{Candid B&W Street}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/04.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/04.jpg" width="100%"></a><br>_{Vivid Everyday Color}</td></tr> <tr><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/05.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/05.jpg" width="100%"></a><br>_{Black Fury}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/06.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/06.jpg" width="100%"></a><br>_{Geometry of Light}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/07.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/07.jpg" width="100%"></a><br>_{Frozen Film Still}</td></tr> <tr><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/08.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/08.jpg" width="100%"></a><br>_{Deadpan Large-Format}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/09.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/09.jpg" width="100%"></a><br>_{White Leap}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/10.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/10.jpg" width="100%"></a><br>_{Corner Elegance}</td></tr> <tr><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/11.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/11.jpg" width="100%"></a><br>_{Raw B&W Fashion}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/12.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/12.jpg" width="100%"></a><br>_{Color Behind Glass}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/13.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/13.jpg" width="100%"></a><br>_{High-Contrast Reportage}</td></tr> <tr><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/14.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/14.jpg" width="100%"></a><br>_{Lyrical Landscape}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/15.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/15.jpg" width="100%"></a><br>_{Hyper Chic East}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/16.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/16.jpg" width="100%"></a><br>_{Street in the Mirror}</td></tr> <tr><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/18.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/18.jpg" width="100%"></a><br>_{Plain Portrait}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/19.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/19.jpg" width="100%"></a><br>_{Melancholy B&W}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/20.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/20.jpg" width="100%"></a><br>_{Fairytale Giants}</td></tr> <tr><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/21.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/21.jpg" width="100%"></a><br>_{Cold Power}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/22.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/22.jpg" width="100%"></a><br>_{Cold Power}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/23.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/23.jpg" width="100%"></a><br>_{Hyper Chic East}</td></tr> <tr><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/24.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/24.jpg" width="100%"></a><br>_{Fairytale Giants}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/25.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/25.jpg" width="100%"></a><br>_{Hyper Chic East}</td><td width="33%"><a href="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/blob/main/showcaseassets/26.jpg"><img src="https://huggingface.co/pottokao/Qwen-Image-2.1-DiT-GGUF/resolve/main/showcaseassets/26.jpg" width="100%"></a><br>_{White Leap}</td></tr> </table>

This is what Q4_K_M across the board looks like — every image above was made with:

_{AI-generated. Style homage only — not affiliated with or endorsed by the photographers.}

Showcase

Outputs from the full pipeline this rewriter drives (rewriter → Qwen-Image-2.1 text encoder → DiT) — the image engine behind a "photography masters" system that originally ran on DiffusionGemma and has since been moved onto this pipeline. Each was generated from a one-line prompt; the rewriter expanded it into the scene, palette and framing below. One seed per cell, no retouching.

<table> <tr> <td width="50%"><img src="showcaseassets/blackfurynightdog.png" width="100%"></td> <td width="50%"><img src="showcaseassets/blackfurystreetdog.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/deadpanportraitpair.png" width="100%"></td> <td width="50%"><img src="showcaseassets/geometryoflightstairs.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/colorbehindglassrain.png" width="100%"></td> <td width="50%"><img src="showcaseassets/surrealbwstreet.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/softeverydaylightchild.png" width="100%"></td> <td width="50%"><img src="showcaseassets/softeverydaylightharbour.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/warmdocumentaryportrait.png" width="100%"></td> <td width="50%"><img src="showcaseassets/palelightglass.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/vivideverydaydiner.png" width="100%"></td> <td width="50%"><img src="showcaseassets/vividgardengoldfish.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/boldminimalflower.png" width="100%"></td> <td width="50%"><img src="showcaseassets/softcolorslip.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/privatediaryflowers.png" width="100%"></td> <td width="50%"><img src="showcaseassets/privatediaryportrait.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/privatediarystreet.png" width="100%"></td> <td width="50%"><img src="showcaseassets/lyricallandscape.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/nocturnalstage.png" width="100%"></td> <td width="50%"><img src="showcaseassets/deadpanindustrialtower.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/minimalseascape.png" width="100%"></td> <td width="50%"><img src="showcaseassets/largeformatcontainers.png" width="100%"></td> </tr> <tr> <td width="50%"><img src="showcaseassets/largeformatretail.png" width="100%"></td> </tr> </table>

_{AI-generated. Style homage only — not affiliated with or endorsed by the photographers.}

Other builds

Results, the Pareto front of other operating points, and the measurement caveats are in the base repo.