---

license: apache-2.0

base_model:

• black-forest-labs/flux-3-image-klein-pup

base_model_relation: quantized

language:

• en

tags:

• gguf
• llama.cpp
• prompt-upsampling
• prompt-rewriting
• text-to-image
• image-editing
• flux
• flux-3

pipeline_tag: image-text-to-text

---

FLUX 3 Image Klein Prompt Upsampler — GGUF

GGUF build of black-forest-labs/flux-3-image-klein-pup,

the 9B vision-language model that expands a short image request into the detailed prompt

FLUX 3 Image Klein works best with. It is a

fine-tune of Qwen3.5-9B. Everything about what the model does, how it

was trained and its limitations is on the original model card; this repository only changes the file format,

for llama.cpp and the tools built on it.

Files

| File | Content | Size | SHA-256 |

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

| flux-3-image-klein-pup-Q4_K_M.gguf | the language model, Q4_K_M | 5.63 GB | c20bcdb1282d1a750199eea1ac8f23860d999e151c748c8ba44556e57761783f |

| mmproj-flux-3-image-klein-pup-F16.gguf | the vision tower and projector, F16 (only needed for edits with reference images) | 0.92 GB | 2f92e35102fce9d976e78d3e6da0122de2911052d9e09413a652316f0431db36 |

| system_prompts.json | the four system prompts, unchanged from the original repository: t2i, i2i, t2i_bbox, i2i_bbox | 13 kB | b5da3ed8bc04ab59dbdb747930f7dbff847e868903fed8028f85ee1a4ac82fb4 |

The chat template is embedded in the language-model GGUF.

How it was made

• Converted from the original model.safetensors (bfloat16) with llama.cpp's convert_hf_to_gguf.py at commit

4df29be: the language

model to a bfloat16 GGUF (architecture qwen35), the vision tower with --mmproj to F16.

• Converted with --no-mtp. The original config.json declares one multi-token-prediction layer

(mtp_num_hidden_layers: 1), but the checkpoint ships no MTP weights; without the flag the GGUF declares 33

blocks and llama.cpp refuses to load it (tensor 'blk.32.attn_norm.weight' not found).

• Quantized from that bfloat16 GGUF to Q4_K_M with llama.cpp's quantizer, no importance matrix.

Requires a llama.cpp build with Qwen3.5 (qwen35) support.

Usage

Use the original card's settings: the system prompt that matches the task, the chat template rendered with

enable_thinking=False, greedy decoding with top_p 0.95 and top_k 20, presence penalty 1.5 for

text-to-image and 0 for edits, and up to 3584 new tokens.

Text-to-image (llama-cpp-python)

The chat template is rendered with enable_thinking=False and sent as a plain completion, because

create_chat_completion does not pass that flag to the template:

import json
from huggingface_hub import hf_hub_download
from llama_cpp import Llama
from llama_cpp.llama_chat_format import Jinja2ChatFormatter

repo = "rekuen/flux-3-image-klein-pup-GGUF"
llm = Llama(model_path=hf_hub_download(repo, "flux-3-image-klein-pup-Q4_K_M.gguf"),
            n_gpu_layers=-1, n_ctx=8192, flash_attn=True)
prompts = json.load(open(hf_hub_download(repo, "system_prompts.json")))

fmt = Jinja2ChatFormatter(template=llm.metadata["tokenizer.chat_template"],
                          eos_token=llm.detokenize([llm.token_eos()], special=True).decode(),
                          bos_token="")
rendered = fmt(messages=[{"role": "system", "content": prompts["t2i"]},
                         {"role": "user", "content": "A man walking with his dog, afternoon sun"}],
               enable_thinking=False)
out = llm.create_completion(rendered.prompt, temperature=0.0, top_p=0.95, top_k=20, min_p=0.0,
                            repeat_penalty=1.0, presence_penalty=1.5, max_tokens=3584,
                            stop=["<|im_end|>", "<|endoftext|>"])
print(out["choices"][0]["text"].strip())

Edits with reference images

Load mmproj-flux-3-image-klein-pup-F16.gguf as the multimodal projector (for example --mmproj in

llama-mtmd-cli / llama-server), use prompts["i2i"], put the reference images first in the user message and

then this text, three lines:

Reference images follow in order (the first image, the second image, ...).
User request: put a red bandana around the dog's neck
Output the caption.

The original inference code resizes each reference before the model sees it (at most 1 MP for one reference,

724×724 for two, 512×512 for three or more, sides on a multiple of 32, flattened onto white and re-encoded as

JPEG); see flux3/pup.py in

black-forest-labs/flux-3-image-klein.

License

The weights are released under the Apache License 2.0, as the original model is. The original model is

by Black Forest Labs and derived from Qwen3.5-9B by the Qwen team at Alibaba Cloud, released under the Apache

License 2.0. system_prompts.json is copied unchanged from

black-forest-labs/flux-3-image-klein-pup.

This is an unofficial conversion, not affiliated with or endorsed by Black Forest Labs or the Qwen team. Use of

the Black Forest Labs and FLUX trademarks is subject to their owners' policies.