[BUG] Flux.2 : strange output

[olivbrau]

Hi, I've tried a CPU backend with Flux.2, here is my command :
sd.exe --diffusion-model ..\Flux.2\flux2-dev-Q2_K.gguf --vae "..\Flux.2\ae.safetensors" --llm ..\Flux.2\Mistral-Small-3.2-24B-Instruct-2506-Q3_K_M.gguf -p "The word 'ZAZ' written as an assembly of many Lego bricks of various colors" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 512 -W 512 -t 20 --steps 20 -s 123456 --preview vae
and here is the output :

Are there limitations on image size (maybe 512 is too small) or may be is there a pb with flux2-dev-Q2 ? Thanks for your help ! Olivier

[olivbrau]

(By the way, this image is quite interesting, like an abstract oil painting ...)

[olivbrau]

and here is the text output :

C:\Users\braultoli\Desktop\sd-master-9578fdc-bin-win-avx2-x64\sd CPU 2025-12-01>sd.exe --diffusion-model ..\Flux.2\flux2-dev-Q2_K.gguf --vae "..\Flux.2\ae.safetensors" --llm ..\Flux.2\Mistral-Small-3.2-24B-Instruct-2506-Q3_K_M.gguf -p "The word 'ZAZ' written as an assembly of many Lego bricks of various colors" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 512 -W 512 -t 20 --steps 20 -s 123456 --preview vae
Option:
n_threads: 20
mode: img_gen
model_path:
wtype: unspecified
clip_l_path:
clip_g_path:
clip_vision_path:
t5xxl_path:
llm_path: ..\Flux.2\Mistral-Small-3.2-24B-Instruct-2506-Q3_K_M.gguf
llm_vision_path:
diffusion_model_path: ..\Flux.2\flux2-dev-Q2_K.gguf
high_noise_diffusion_model_path:
vae_path: ..\Flux.2\ae.safetensors
taesd_path:
esrgan_path:
control_net_path:
embedding_dir:
photo_maker_path:
pm_id_images_dir:
pm_id_embed_path:
pm_style_strength: 20.00
output_path: output.png
init_image_path:
end_image_path:
mask_image_path:
control_image_path:
ref_images_paths:
control_video_path:
auto_resize_ref_image: true
increase_ref_index: false
offload_params_to_cpu: true
clip_on_cpu: false
control_net_cpu: false
vae_on_cpu: false
diffusion flash attention: true
diffusion Conv2d direct: false
vae_conv_direct: false
control_strength: 0.90
prompt: The word 'ZAZ' written as an assembly of many Lego bricks of various colors
negative_prompt:
clip_skip: -1
width: 512
height: 512
sample_params: (txt_cfg: 1.00, img_cfg: 1.00, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: 0.00, shifted_timestep: 0)
high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 3, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: -1, eta: 0.00, shifted_timestep: 0)
moe_boundary: 0.875
prediction: default
lora_apply_mode: auto
flow_shift: inf
strength(img2img): 0.75
rng: cuda
sampler rng: NONE
seed: 123456
batch_count: 1
vae_tiling: false
force_sdxl_vae_conv_scale: false
upscale_repeats: 1
chroma_use_dit_mask: true
chroma_use_t5_mask: false
chroma_t5_mask_pad: 1
video_frames: 1
easycache: disabled (threshold=0.200, start=0.15, end=0.95)
vace_strength: 1.00
fps: 16
preview_mode: vae (denoised)
preview_interval: 1
System Info:
SSE3 = 1
AVX = 1
AVX2 = 1
AVX512 = 0
AVX512_VBMI = 0
AVX512_VNNI = 0
FMA = 1
NEON = 0
ARM_FMA = 0
F16C = 1
FP16_VA = 0
WASM_SIMD = 0
VSX = 0
[DEBUG] stable-diffusion.cpp:189 - Using CPU backend
[INFO ] stable-diffusion.cpp:234 - loading diffusion model from '..\Flux.2\flux2-dev-Q2_K.gguf'
[INFO ] model.cpp:378 - load ..\Flux.2\flux2-dev-Q2_K.gguf using gguf format
[DEBUG] model.cpp:420 - init from '..\Flux.2\flux2-dev-Q2_K.gguf'
[INFO ] stable-diffusion.cpp:281 - loading llm from '..\Flux.2\Mistral-Small-3.2-24B-Instruct-2506-Q3_K_M.gguf'
[INFO ] model.cpp:378 - load ..\Flux.2\Mistral-Small-3.2-24B-Instruct-2506-Q3_K_M.gguf using gguf format
[DEBUG] model.cpp:420 - init from '..\Flux.2\Mistral-Small-3.2-24B-Instruct-2506-Q3_K_M.gguf'
[INFO ] stable-diffusion.cpp:295 - loading vae from '..\Flux.2\ae.safetensors'
[INFO ] model.cpp:381 - load ..\Flux.2\ae.safetensors using safetensors format
[DEBUG] model.cpp:511 - init from '..\Flux.2\ae.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:318 - Version: Flux.2
[INFO ] stable-diffusion.cpp:346 - Weight type stat: f32: 457 | q2_K: 136 | q3_K: 161 | q4_K: 140 | q5_K: 4 | bf16: 11
[INFO ] stable-diffusion.cpp:347 - Conditioner weight type stat: f32: 81 | q3_K: 161 | q4_K: 116 | q5_K: 4
[INFO ] stable-diffusion.cpp:348 - Diffusion model weight type stat: f32: 128 | q2_K: 136 | q4_K: 24 | bf16: 11
[INFO ] stable-diffusion.cpp:349 - VAE weight type stat: f32: 248
[DEBUG] stable-diffusion.cpp:351 - ggml tensor size = 400 bytes
[DEBUG] llm.hpp:386 - vocab size: 131072
[DEBUG] llm.hpp:401 - merges size 269443
[INFO ] flux.hpp:1256 - Flux blocks: 8 double, 48 single
[INFO ] stable-diffusion.cpp:535 - Using flash attention in the diffusion model
[DEBUG] ggml_extend.hpp:1877 - mistral_small3.2 params backend buffer size = 12695.02 MB(RAM) (362 tensors)
[DEBUG] ggml_extend.hpp:1877 - flux params backend buffer size = 12262.56 MB(RAM) (299 tensors)
[DEBUG] ggml_extend.hpp:1877 - vae params backend buffer size = 94.72 MB(RAM) (140 tensors)
[DEBUG] stable-diffusion.cpp:683 - loading weights
[DEBUG] model.cpp:1359 - using 20 threads for model loading
[DEBUG] model.cpp:1381 - loading tensors from ..\Flux.2\flux2-dev-Q2_K.gguf
|================> | 299/909 - 118.42it/s←[K
[DEBUG] model.cpp:1381 - loading tensors from ..\Flux.2\Mistral-Small-3.2-24B-Instruct-2506-Q3_K_M.gguf
|====================================> | 661/909 - 88.35it/s←[KK
[DEBUG] model.cpp:1381 - loading tensors from ..\Flux.2\ae.safetensors
|==================================================| 909/909 - 118.27it/s←[K
[INFO ] model.cpp:1590 - loading tensors completed, taking 7.69s (process: 0.00s, read: 6.86s, memcpy: 0.00s, convert: 0.15s, copy_to_backend: 0.00s)
[INFO ] stable-diffusion.cpp:782 - total params memory size = 25052.30MB (VRAM 0.00MB, RAM 25052.30MB): text_encoders 12695.02MB(RAM), diffusion_model 12262.56MB(RAM), vae 94.72MB(RAM), controlnet 0.00MB(VRAM), pmid 0.00MB(RAM)
[INFO ] stable-diffusion.cpp:866 - running in Flux2 FLOW mode
[DEBUG] stable-diffusion.cpp:908 - finished loaded file
[DEBUG] stable-diffusion.cpp:3138 - generate_image 512x512
[INFO ] stable-diffusion.cpp:3169 - sampling using Euler method
[DEBUG] denoiser.hpp:660 - Flux2FlowDenoiser: set shift to 1.916
[INFO ] denoiser.hpp:364 - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:3282 - TXT2IMG
[INFO ] stable-diffusion.cpp:1167 - apply at runtime
[DEBUG] conditioner.hpp:1701 - parse '[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object
attribution and actions without speculation.[/SYSTEM_PROMPT][INST]The word 'ZAZ' written as an assembly of many Lego bricks of various colors[/INST]' to [['[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object
attribution and actions without speculation.[/SYSTEM_PROMPT][INST]', 1], ['The word 'ZAZ' written as an assembly of many Lego bricks of various colors', 1], ['[/INST]', 1], ]
[DEBUG] llm.hpp:259 - split prompt "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object
attribution and actions without speculation.[/SYSTEM_PROMPT][INST]" to tokens ["[SYSTEM_PROMPT]", "You", "─áare", "─áan", "─áAI", "─áthat", "─áreasons", "─áabout", "─áimage", "─ádescriptions", ".", "─áYou", "─ágive", "─ástructured", "─áresponses", "─áfocusing", "─áon", "─áobject", "─árelationships", ",", "─áobject", "─è", "att", "ribution", "─áand", "─áactions", "─áwithout", "─áspeculation", ".", "[/SYSTEM_PROMPT]", "[INST]", ]
[DEBUG] llm.hpp:259 - split prompt "The word 'ZAZ' written as an assembly of many Lego bricks of various colors" to tokens ["The", "─áword", "─á'", "Z", "AZ", "'", "─áwritten", "─áas", "─áan", "─áassembly", "─áof", "─ámany", "─áL", "ego", "─ábricks", "─áof", "─ávarious", "─ácolors", ]
[DEBUG] llm.hpp:259 - split prompt "[/INST]" to tokens ["[/INST]", ]
[DEBUG] ggml_extend.hpp:1691 - mistral_small3.2 compute buffer size: 17.73 MB(RAM)
[DEBUG] conditioner.hpp:1896 - computing condition graph completed, taking 5973 ms
[INFO ] stable-diffusion.cpp:2917 - get_learned_condition completed, taking 6123 ms
[INFO ] stable-diffusion.cpp:3028 - generating image: 1/1 - seed 123456
[DEBUG] ggml_extend.hpp:1691 - flux compute buffer size: 770.33 MB(RAM)
[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|==> | 1/20 - 260.95s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|=====> | 2/20 - 263.98s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|=======> | 3/20 - 258.97s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|==========> | 4/20 - 256.11s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|============> | 5/20 - 254.03s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|===============> | 6/20 - 252.57s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|=================> | 7/20 - 251.48s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|====================> | 8/20 - 250.60s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|======================> | 9/20 - 249.96s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|=========================> | 10/20 - 249.48s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|===========================> | 11/20 - 249.06s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|==============================> | 12/20 - 248.72s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|================================> | 13/20 - 248.47s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|===================================> | 14/20 - 248.24s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|=====================================> | 15/20 - 248.04s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|========================================> | 16/20 - 247.83s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|==========================================> | 17/20 - 247.67s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|=============================================> | 18/20 - 247.55s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|===============================================> | 19/20 - 247.43s/it←[K[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
|==================================================| 20/20 - 247.30s/it←[K
[INFO ] stable-diffusion.cpp:3069 - sampling completed, taking 4946.08s
[INFO ] stable-diffusion.cpp:3077 - generating 1 latent images completed, taking 4947.86s
[INFO ] stable-diffusion.cpp:3080 - decoding 1 latents
[DEBUG] ggml_extend.hpp:1691 - vae compute buffer size: 1664.00 MB(RAM)
[DEBUG] stable-diffusion.cpp:2286 - computing vae decode graph completed, taking 12.15s
[INFO ] stable-diffusion.cpp:3090 - latent 1 decoded, taking 12.15s
[INFO ] stable-diffusion.cpp:3094 - decode_first_stage completed, taking 12.15s
[INFO ] stable-diffusion.cpp:3390 - generate_image completed in 4966.16s
save result PNG image to 'output.png' (success)

C:\Users\braultoli\Desktop\sd-master-9578fdc-bin-win-avx2-x64\sd CPU 2025-12-01>REM --vae-tiling

[wbruna]

As with #1031 , quantization issues could cause this. I do get "abstract" images on some other models with very low quants; for instance, this is "a cat" from a locally quantized (likely broken) Q2_K Z-Image:

But I'd suggest comparing the sha256 of your downloaded files with the ones shown on the Hugginface page, just to be sure it's not simply a damaged model file.

[olivbrau]

I will try to check the file.
If I took such a small file is because my amount of RAM, which is 32 GB.
I notice that when the processing is running, sd use only the size of the model (in this case Q2 is 12.5 GB), but during the initial sate, sd try to load the model AND the llm. And 32 GB is not enough with less quantized model.
Is this a normal behaviour ? Is there a way to load and process image model and llm separatly ?

[wbruna]

Is this a normal behaviour ? Is there a way to load and process image model and llm separatly ?

It is expected, yes (see #908 and #979 for related discussions).

[olivbrau]

Argh, so in reality, the RAM needed to use a model is necessarily the sum of the imager model + size of the LLM, but just at the begining ...
Which force me to use very quantized model ...
I will try Q3 to see

[olivbrau]

So I tried with flux2-dev-Q3_K_M.gguf, and here is the kind of image in the first iterations :