---

license: other

license_name: upstream-model-licenses

license_link: https://huggingface.co/t8star/Flux3-PE-GGUF-Comfy/blob/main/LICENSES.md

language:

- en

- zh

base_model: black-forest-labs/flux-3-image-klein-pup

base_model_relation: quantized

library_name: gguf

pipeline_tag: image-text-to-text

tags:

- gguf

- comfyui

- prompt-enhancement

- flux3

- qwen3.5

- heretic

---

FLUX 3 PE GGUF · T8

English | 简体中文

Local Prompt Enhancement for ComfyUI using Q4_K_M versions of official FLUX 3 Image Klein PUP and Heretic PUP. Supports text, layout constraints and up to 10 reference images. Output is a detailed prompt for a text encoder; image generation uses separate diffusion, text-encoder and VAE weights.

ComfyUI node · v0.2.3 release · Workflows

Downloads

Choose one variant and download its model, projector and manifest together. Each Q4 model is 5.63 GB; each F16 projector is 0.92 GB. Sizes use decimal GB.

| Variant | Q4_K_M model | F16 projector | Manifest | Destination in node folder |

|---|---|---|---|---|

| Official PUP | Download | Download | Download | models/gguf/ |

| Heretic PUP | Download | Download | Download | models/gguf/heretic/ |

Keep the manifests in their respective directories. SHA256SUMS.txt covers all four GGUF files. Q4_0 and BF16 are not distributed in this repository.

With the Hugging Face CLI installed, run from the node folder to download Heretic:

hf download t8star/Flux3-PE-GGUF-Comfy --include "heretic/*" --local-dir models/gguf

ComfyUI

1. Install the node, version 0.2.2+ for Heretic; v0.2.3 is the model-distribution and documentation update.
2. Install llama-cpp-python >=0.3.36, <0.4 in the same Python that runs ComfyUI. CPU wheels · CUDA 12.4 wheels · CUDA 13.2 wheels. Match your Python, OS, architecture and hardware runtime.
3. Restart ComfyUI. Select the model in FLUX 3 PE Model (GGUF) and choose auto or llama-cpp-python.
4. Drag a workflow JSON into the canvas. For Heretic, start with text enhancement or PE → FLUX 3. Reference workflows use the two images in examples/reference_images/; copy them to ComfyUI's input folder with the FLUX3_PE_ filename prefix.

The wheel backend runs in Python without a server executable. GGUF files are platform-independent; acceleration depends on the chosen wheel and compatible runtime libraries. Keep keep_loaded=false to release PE memory before image generation. The legacy server runtime remains optional.

Provenance and validation

| Variant | Source | Pinned revision |

|---|---|---|

| Official PUP | Black Forest Labs | 8b069751eb5dcab23829d9c5eccb470eb8b8a4ed |

| Heretic PUP | tuanlai on ModelScope | 80efba65e01bb2e320dea412fa657d701491b443 |

Heretic GGUF was converted from the verified BF16 source with llama.cpp 620d2225cc436ab0b001e3200e590e3b286ef9ab, using pinned official PUP configuration and PE protocols. Source and artifact SHA256 values are recorded in each manifest.

The v0.2.2 Heretic runtime passed 481 unit tests, 40/40 real-model tasks, seven canvas executions, four workflow save/reopen checks and PNG restoration, including a 512×512 PE-to-FLUX3 image and six official-model regressions. Validation record · Actual canvas. Inference code and workflows are unchanged in v0.2.3.

Heretic model tests used Windows / RTX 5090 Laptop / llama-cpp-python 0.3.36 CUDA 13.2. The official PUP backend also passed six Linux CPU model tasks. Heretic inference on Linux/macOS was not tested. Structural and sampled text checks do not guarantee every prompt's semantics.

Credits and licenses

GGUF conversion and node integration: T8star. Official PUP: Black Forest Labs, Apache-2.0. Heretic checkpoint: tuanlai, with its upstream license declarations retained separately. See LICENSES.md, NOTICE and conversion details.

T8star

Bilibili · YouTube · API · Free gallery

Online AI apps · ComfyUI bundle · Hugging Face