gguf-org

/

mageflow-gguf

like 15

Follow

gguf 217

you don't need clip_l clip_g and t5xxl for this model; try to use mmproj f16 instead of q8; set the steps to 4 and cfg-scale to 1

@calcuis
Thanks! But I get an error that way.

gguf-diffusion engine -- \
--ref-image cat.png \
--diffusion-model models/mageflow-edit-turbo-nvfp4.gguf \
--vae models/pig_mageflow_vae_fp32-f16.gguf \
--llm models/Qwen3VL-4B-Thinking-Q4_K_M.gguf \
--llm_vision models/mmproj-Qwen3VL-8B-Thinking-F16.gguf \
-W 1024 -H 704 \
-p 'A cat in sunglasses' \
--steps 4 --cfg-scale 1 \
-v --clip-on-cpu \
-o out.png

[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
0x00007aec1ff10813 in __GI___wait4 (pid=43476, stat_loc=0x0, options=0, usage=0x0) at ../sysdeps/unix/sysv/linux/wait4.c:30
warning: 30	../sysdeps/unix/sysv/linux/wait4.c: Нет такого файла или каталога
#0  0x00007aec1ff10813 in __GI___wait4 (pid=43476, stat_loc=0x0, options=0, usage=0x0) at ../sysdeps/unix/sysv/linux/wait4.c:30
30	in ../sysdeps/unix/sysv/linux/wait4.c
#1  0x000056b3fe007823 in ggml_print_backtrace ()
#2  0x000056b3fe0079c7 in ggml_abort ()
#3  0x000056b3fe00d846 in ggml_concat ()
#4  0x000056b3fd00b1b6 in LLM::splice_image_embeds(GGMLRunnerContext*, ggml_tensor*, std::vector<std::pair<int, ggml_tensor*>, std::allocator<std::pair<int, ggml_tensor*> > > const&) ()
#5  0x000056b3fd15b22e in LLM::LLMRunner::build_graph(sd::Tensor<int> const&, sd::Tensor<float> const&, std::vector<std::pair<int, sd::Tensor<float> >, std::allocator<std::pair<int, sd::Tensor<float> > > > const&, std::set<int, std::less<int>, std::allocator<int> >, bool) ()
#6  0x000056b3fd15bf71 in std::_Function_handler<ggml_cgraph* (), LLM::LLMRunner::compute(int, sd::Tensor<int> const&, sd::Tensor<float> const&, std::vector<std::pair<int, sd::Tensor<float> >, std::allocator<std::pair<int, sd::Tensor<float> > > > const&, std::set<int, std::less<int>, std::allocator<int> >, bool, bool, bool, bool)::{lambda()#1}>::_M_invoke(std::_Any_data const&) ()
#7  0x000056b3fd1d8c65 in std::optional<sd::Tensor<float> > GGMLRunner::compute<float>(std::function<ggml_cgraph* ()>, int, bool, bool, bool, bool) ()
#8  0x000056b3fd207fd2 in LLMEmbedder::encode_prompt(int, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::pair<int, int> const&, int, int, std::vector<std::pair<int, sd::Tensor<float> >, std::allocator<std::pair<int, sd::Tensor<float> > > > const&, std::set<int, std::less<int>, std::allocator<int> > const&, int, bool, int) [clone .isra.0] ()
#9  0x000056b3fd209608 in LLMEmbedder::get_learned_condition(int, ConditionerParams const&) ()
#10 0x000056b3fd02f7fa in generate_image ()
#11 0x000056b3fcec81e1 in main ()
[Inferior 1 (process 43454) detached]

mmproj-Qwen3VL-8B-Thinking-F16 is not right, you need to use 4b; 4b te and 8b mmproj cannot be mixed

@calcuis
Thanks for your reply! But I tried that too. There are strange elements in the image.

gguf-diffusion engine -- \
--ref-image cat.png \
--diffusion-model models/mageflow-edit-turbo-nvfp4.gguf \
--vae models/pig_mageflow_vae_fp32-f16.gguf \
--llm models/Qwen3VL-4B-Thinking-Q4_K_M.gguf \
--llm_vision models/mmproj-Qwen3VL-4B-Thinking-F16.gguf \
-W 1024 -H 704 \
-p 'A cat in sunglasses' \
--steps 4 --cfg-scale 1 \
-v --clip-on-cpu \
-o out4.png

tested with instruct instead of thinking clip; seems no problem

gguf-diffusion engine -- --ref-image cat.png --diffusion-model models/mageflow-edit-turbo-nvfp4.gguf --vae models/pig_mageflow_vae_fp32-f16.gguf --llm models/qwen3vl-4b-it-q4_k_m.gguf --llm_vision models/mmproj-qwen3vl-4b-it-f16.gguf -W 1024 -H 704 -p 'A cat in sunglasses' --steps 4 --cfg-scale 1 -v --clip-on-cpu -o out.png