[Bug] ERNIE - white blank (diffusion-fa and gen-size affected) #1447

Activity

1. SolicTous

   added

   bugSomething isn't workingSomething isn't working

   on Apr 20, 2026
2. maxi1134 commented on Apr 20, 2026

   

   maxi1134

   on Apr 20, 2026

   Last edited by maxi1134

   More actions

   I am also still getting this issue with z-image-Q4_K_M.gguf

   I am piping this tool into Home-Assistant and it returns a white image

   

   And the logs I get

   [INFO ] stable-diffusion.cpp:776  - Using flash attention in the diffusion model
     |====================>                             | 453/1095 - 4.71GB/s
     |======================================>           | 851/1095 - 3.25GB/s
     |==================================================| 1095/1095 - 3.04GB/s
   [INFO ] model.cpp:1006 - loading tensors completed, taking 2.37s (process: 0.00s, read: 0.49s, memcpy: 0.00s, convert: 0.15s, copy_to_backend: 0.55s)
   [INFO ] stable-diffusion.cpp:897  - total params memory size = 8549.12MB (VRAM 8549.12MB, RAM 0.00MB): text_encoders 3555.38MB(VRAM), diffusion_model 4833.74MB(VRAM), vae 160.00MB(VRAM), controlnet 0.00MB(VRAM), pmid 0.00MB(VRAM)
   [INFO ] stable-diffusion.cpp:981  - running in FLOW mode
   [INFO ] main.cpp:142  - listening on: 0.0.0.0:1234
   [INFO ] stable-diffusion.cpp:3160 - generate_image 512x512
   [INFO ] denoiser.hpp:499  - get_sigmas with discrete scheduler
   [INFO ] stable-diffusion.cpp:2736 - sampling using Euler method
   [INFO ] stable-diffusion.cpp:3090 - get_learned_condition completed, taking 17.68s
   [INFO ] stable-diffusion.cpp:3194 - generating image: 1/1 - seed 42
     |==================================================| 20/20 - 1.28s/it
   [INFO ] stable-diffusion.cpp:3225 - sampling completed, taking 25.69s
   [INFO ] stable-diffusion.cpp:3243 - generating 1 latent images completed, taking 25.69s
   [INFO ] stable-diffusion.cpp:3114 - decoding 1 latents
   [INFO ] stable-diffusion.cpp:3130 - latent 1 decoded, taking 0.38s
   [INFO ] stable-diffusion.cpp:3134 - decode_first_stage completed, taking 0.38s
   [INFO ] stable-diffusion.cpp:3255 - generate_image completed in 43.77s
   
   
3. wbruna commented on Apr 20, 2026

   

   wbruna

   on Apr 20, 2026

   Contributor

   More actions

   Please post the full verbose generation log, including the command line and backend initialization part. It's also useful to know where did the model files came from, and if it affects other model types on the same backend.

   It's also useful to know if turning on/off certain flags help with the issue: --diffusion-fa (or --fa), --offload-to-cpu, --clip-on-cpu, --vae-on-cpu.
4. SolicTous commented on Apr 21, 2026

   

   SolicTous

   on Apr 21, 2026

   Last edited by SolicTous

   Author

   More actions

   Turning off --diffusion-fa helped.
   Result looks some artifact (text, cat legs), but any way high quality.
   

   main model - https://huggingface.co/unsloth/ERNIE-Image-Turbo-GGUF/blob/main/ernie-image-turbo-Q8_0.gguf
   original vae from ERNIE-Image-Turbo rep
   llm - https://huggingface.co/unsloth/Ministral-3-3B-Instruct-2512-GGUF/blob/main/Ministral-3-3B-Instruct-2512-UD-Q8_K_XL.gguf
   (the same if using https://huggingface.co/Comfy-Org/ERNIE-Image/blob/main/text_encoders/ministral-3-3b.safetensors)

   Below will be result without diffusion-fa. Also will add in the top message log from white blank with diffusion-fa.
   "../sd-cli.exe" --diffusion-model ../../Models/Diffusers/ERNIE-Image/ernie-image-turbo-Q8_0.gguf --vae ../../Models/Diffusers/Vae/vae_ernie_image_turbo.safetensors --llm ../../Models/Diffusers/Ministral/Ministral-3-3B-Instruct-2512-UD-Q8_K_XL.gguf -p "a lovely cat holding a sign says 'ernie.cpp'" --cfg-scale 1.0 --steps 8 -H 1024 -W 1024 -v -o ../output/cli_out.jpg
   [DEBUG] main.cpp:547 - version: stable-diffusion.cpp version unknown, commit 44cca3d
   [DEBUG] main.cpp:548 - 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] main.cpp:549 - SDCliParams {
   mode: img_gen,
   output_path: "../output/cli_out.jpg",
   image_path: "",
   metadata_format: "text",
   verbose: true,
   color: false,
   canny_preprocess: false,
   convert_name: false,
   preview_method: none,
   preview_interval: 1,
   preview_path: "preview.png",
   preview_fps: 16,
   taesd_preview: false,
   preview_noisy: false,
   metadata_raw: false,
   metadata_brief: false,
   metadata_all: false
   }
   [DEBUG] main.cpp:550 - SDContextParams {
   n_threads: 8,
   model_path: "",
   clip_l_path: "",
   clip_g_path: "",
   clip_vision_path: "",
   t5xxl_path: "",
   llm_path: "../../Models/Diffusers/Ministral/Ministral-3-3B-Instruct-2512-UD-Q8_K_XL.gguf",
   llm_vision_path: "",
   diffusion_model_path: "../../Models/Diffusers/ERNIE-Image/ernie-image-turbo-Q8_0.gguf",
   high_noise_diffusion_model_path: "",
   vae_path: "../../Models/Diffusers/Vae/vae_ernie_image_turbo.safetensors",
   taesd_path: "",
   esrgan_path: "",
   control_net_path: "",
   embedding_dir: "",
   embeddings: {
   }
   wtype: NONE,
   tensor_type_rules: "",
   lora_model_dir: ".",
   photo_maker_path: "",
   rng_type: cuda,
   sampler_rng_type: NONE,
   offload_params_to_cpu: false,
   enable_mmap: false,
   control_net_cpu: false,
   clip_on_cpu: false,
   vae_on_cpu: false,
   flash_attn: false,
   diffusion_flash_attn: false,
   diffusion_conv_direct: false,
   vae_conv_direct: false,
   circular: false,
   circular_x: false,
   circular_y: false,
   chroma_use_dit_mask: true,
   qwen_image_zero_cond_t: false,
   chroma_use_t5_mask: false,
   chroma_t5_mask_pad: 1,
   prediction: NONE,
   lora_apply_mode: auto,
   force_sdxl_vae_conv_scale: false
   }
   [DEBUG] main.cpp:551 - SDGenerationParams {
   loras: "{
   }",
   high_noise_loras: "{
   }",
   prompt: "a lovely cat holding a sign says 'ernie.cpp'",
   negative_prompt: "",
   clip_skip: -1,
   width: 1024,
   height: 1024,
   batch_count: 1,
   init_image_path: "",
   end_image_path: "",
   mask_image_path: "",
   control_image_path: "",
   ref_image_paths: [],
   control_video_path: "",
   auto_resize_ref_image: true,
   increase_ref_index: false,
   pm_id_images_dir: "",
   pm_id_embed_path: "",
   pm_style_strength: 20,
   skip_layers: [7, 8, 9],
   sample_params: (txt_cfg: 1.00, img_cfg: 1.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 8, eta: inf, shifted_timestep: 0, flow_shift: inf),
   high_noise_skip_layers: [7, 8, 9],
   high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: inf, shifted_timestep: 0, flow_shift: inf),
   custom_sigmas: [],
   cache_mode: "",
   cache_option: "",
   cache: disabled (threshold=inf, start=0.15, end=0.95),
   moe_boundary: 0.875,
   video_frames: 1,
   fps: 16,
   vace_strength: 1,
   strength: 0.75,
   control_strength: 0.9,
   seed: 42,
   upscale_repeats: 1,
   upscale_tile_size: 128,
   vae_tiling_params: { 0, 0, 0, 0.5, 0, 0 },
   }
   [DEBUG] stable-diffusion.cpp:175 - Using CUDA backend
   [INFO ] ggml_extend.hpp:81 - ggml_cuda_init: found 1 CUDA devices (Total VRAM: 12281 MiB):
   [INFO ] ggml_extend.hpp:81 - Device 0: NVIDIA GeForce RTX 4070 Ti, compute capability 8.9, VMM: yes, VRAM: 12281 MiB
   [INFO ] stable-diffusion.cpp:269 - loading diffusion model from '../../Models/Diffusers/ERNIE-Image/ernie-image-turbo-Q8_0.gguf'
   [INFO ] model.cpp:229 - load ../../Models/Diffusers/ERNIE-Image/ernie-image-turbo-Q8_0.gguf using gguf format
   [DEBUG] model.cpp:278 - init from '../../Models/Diffusers/ERNIE-Image/ernie-image-turbo-Q8_0.gguf'
   [INFO ] stable-diffusion.cpp:316 - loading llm from '../../Models/Diffusers/Ministral/Ministral-3-3B-Instruct-2512-UD-Q8_K_XL.gguf'
   [INFO ] model.cpp:229 - load ../../Models/Diffusers/Ministral/Ministral-3-3B-Instruct-2512-UD-Q8_K_XL.gguf using gguf format
   [DEBUG] model.cpp:278 - init from '../../Models/Diffusers/Ministral/Ministral-3-3B-Instruct-2512-UD-Q8_K_XL.gguf'
   [INFO ] stable-diffusion.cpp:330 - loading vae from '../../Models/Diffusers/Vae/vae_ernie_image_turbo.safetensors'
   [INFO ] model.cpp:232 - load ../../Models/Diffusers/Vae/vae_ernie_image_turbo.safetensors using safetensors format
   [DEBUG] model.cpp:307 - init from '../../Models/Diffusers/Vae/vae_ernie_image_turbo.safetensors', prefix = 'vae.'
   [INFO ] stable-diffusion.cpp:355 - Version: Ernie Image
   [INFO ] stable-diffusion.cpp:383 - Weight type stat: f32: 203 | f16: 26 | q8_0: 410 | bf16: 254
   [INFO ] stable-diffusion.cpp:384 - Conditioner weight type stat: f32: 53 | f16: 26 | q8_0: 157
   [INFO ] stable-diffusion.cpp:385 - Diffusion model weight type stat: f32: 150 | q8_0: 253 | bf16: 6
   [INFO ] stable-diffusion.cpp:386 - VAE weight type stat: bf16: 248
   [DEBUG] stable-diffusion.cpp:388 - ggml tensor size = 400 bytes
   [DEBUG] mistral_tokenizer.cpp:23 - vocab size: 131072
   [DEBUG] mistral_tokenizer.cpp:31 - merges size 269443
   [DEBUG] llm.hpp:697 - llm: num_layers = 26, vocab_size = 131072, hidden_size = 3072, intermediate_size = 9216
   [INFO ] ernie_image.hpp:383 - ernie_image: layers = 36, hidden_size = 4096, heads = 32, ffn_hidden_size = 12288, in_channels = 128, out_channels = 128
   [DEBUG] ggml_extend.hpp:2050 - ministral3.3b params backend buffer size = 4285.00 MB(VRAM) (236 tensors)
   [DEBUG] ggml_extend.hpp:2050 - ernie_image params backend buffer size = 8292.08 MB(VRAM) (409 tensors)
   [INFO ] stable-diffusion.cpp:681 - using VAE for encoding / decoding
   [INFO ] auto_encoder_kl.hpp:517 - vae decoder: ch = 128
   [DEBUG] ggml_extend.hpp:2050 - vae params backend buffer size = 94.72 MB(VRAM) (140 tensors)
   [DEBUG] stable-diffusion.cpp:805 - loading weights
   [DEBUG] model.cpp:755 - using 8 threads for model loading
   [DEBUG] model.cpp:777 - loading tensors from ../../Models/Diffusers/ERNIE-Image/ernie-image-turbo-Q8_0.gguf
   |======================> | 409/893 - 3.18GB/s
   [DEBUG] model.cpp:777 - loading tensors from ../../Models/Diffusers/Ministral/Ministral-3-3B-Instruct-2512-UD-Q8_K_XL.gguf
   |====================================> | 645/893 - 3.14GB/s
   [DEBUG] model.cpp:777 - loading tensors from ../../Models/Diffusers/Vae/vae_ernie_image_turbo.safetensors
   |==================================================| 893/893 - 3.00GB/s
   [INFO ] model.cpp:1012 - loading tensors completed, taking 4.12s (process: 0.00s, read: 2.45s, memcpy: 0.00s, convert: 0.03s, copy_to_backend: 1.09s)
   [DEBUG] stable-diffusion.cpp:845 - finished loaded file
   [INFO ] stable-diffusion.cpp:912 - total params memory size = 12671.79MB (VRAM 12671.79MB, RAM 0.00MB): text_encoders 4285.00MB(VRAM), diffusion_model 8292.08MB(VRAM), vae 94.72MB(VRAM), controlnet 0.00MB(VRAM), pmid 0.00MB(VRAM)
   [INFO ] stable-diffusion.cpp:981 - running in FLOW mode
   [INFO ] stable-diffusion.cpp:3160 - generate_image 1024x1024
   [INFO ] denoiser.hpp:499 - get_sigmas with discrete scheduler
   [INFO ] stable-diffusion.cpp:2736 - sampling using Euler method
   [DEBUG] conditioner.hpp:1699 - parse 'a lovely cat holding a sign says 'ernie.cpp'' to [['a lovely cat holding a sign says 'ernie.cpp'', 1], ]
   [DEBUG] bpe_tokenizer.cpp:183 - split prompt "a lovely cat holding a sign says 'ernie.cpp'" to tokens ["a", "Ġlovely", "Ġcat", "Ġholding", "Ġa", "Ġsign", "Ġsays", "Ġ'", "ern", "ie", ".cpp", "'", ]
   [DEBUG] ggml_extend.hpp:1862 - ministral3.3b compute buffer size: 1.42 MB(VRAM)
   [DEBUG] conditioner.hpp:1953 - computing condition graph completed, taking 87 ms
   [INFO ] stable-diffusion.cpp:3090 - get_learned_condition completed, taking 0.09s
   [INFO ] stable-diffusion.cpp:3194 - generating image: 1/1 - seed 42
   [DEBUG] ggml_extend.hpp:1862 - ernie_image compute buffer size: 2452.20 MB(VRAM)
   |==================================================| 8/8 - 3.56s/it
   [INFO ] stable-diffusion.cpp:3225 - sampling completed, taking 28.96s
   [INFO ] stable-diffusion.cpp:3245 - generating 1 latent images completed, taking 29.09s
   [INFO ] stable-diffusion.cpp:3114 - decoding 1 latents
   [DEBUG] ggml_extend.hpp:1862 - vae compute buffer size: 6658.00 MB(VRAM)
   [DEBUG] vae.hpp:206 - computing vae decode graph completed, taking 0.96s
   [INFO ] stable-diffusion.cpp:3130 - latent 1 decoded, taking 0.99s
   [INFO ] stable-diffusion.cpp:3134 - decode_first_stage completed, taking 0.99s
   [INFO ] stable-diffusion.cpp:3255 - generate_image completed in 30.66s
   [INFO ] main.cpp:438 - save result image 0 to '../output/cli_out.jpg' (success)
   [INFO ] main.cpp:487 - 1/1 images saved

   👀React with 👀1Reacted by Erik Scholz
5. SolicTous

   changed the title [-][Bug] Ernie - white blank[/-] [+][Bug] Ernie - white blank (with diffusion fa)[/+] on Apr 21, 2026
6. SolicTous commented on Apr 21, 2026

   

   SolicTous

   on Apr 21, 2026

   Author

   More actions

   "a lovely cat holding a sign says 'ernie.cpp'"

   the problem in unsloth weights, after manual download from orig and converting to q8 via sdcpp all works fine

   

   👀React with 👀2Reacted by Erik Scholz and KuhnChris
7. SolicTous

   closed this as completedon Apr 21, 2026
8. mpulukkinen commented on Apr 21, 2026

   

   mpulukkinen

   on Apr 21, 2026

   Contributor

   More actions

   Lot's of paws with that cat 🤣
9. SolicTous commented on Apr 21, 2026

   

   SolicTous

   on Apr 21, 2026

   Author

   More actions

   Addidtional way getting white blank

   Without setting Sampler/Scheduler - White Blank

   "sample_method": 'euler', "scheduler": 'simple' --> Normal generation

   I will check sample_methods and schedulers and provide here an info
10. SolicTous

    reopened this on Apr 21, 2026
11. kuhnchris commented on Apr 21, 2026

    

    kuhnchris

    on Apr 21, 2026

    More actions

    "a lovely cat holding a sign says 'ernie.cpp'"

    the problem in unsloth weights, after manual download from orig and converting to q8 via sdcpp all works fine

    

    Unsloth Weights + diffusion-fa, or is it solely unsloth regardless of diffusion-fa?
12. SolicTous commented on Apr 21, 2026

    

    SolicTous

    on Apr 21, 2026

    Last edited by SolicTous

    Author

    More actions

    @kuhnchris

    1. Unsloth Weights works only without -diffusion-fa, and some bad artifacts

    2. Orig Weights working with -diffusion-fa and without it. Manual and accurate convert work as well.

    3. Now Investigated that some bad generation failed several samplers and schedulers
       currently - "discrete", "kl_optimal", "karras", "exponential" with 1024 - white blank, others - good

    

    👀React with 👀1Reacted by KuhnChris
13. SolicTous commented on Apr 21, 2026

    

    SolicTous

    on Apr 21, 2026

    Author

    More actions

    512 - size - No white blanks
    
14. SolicTous

    changed the title [-][Bug] Ernie - white blank (with diffusion fa)[/-] [+][Bug] Ernie - white blank (diffusion fa and gen-size affected)[/+] on Apr 21, 2026
15. SolicTous

    changed the title [-][Bug] Ernie - white blank (diffusion fa and gen-size affected)[/-] [+][Bug] ERNIE - white blank (diffusion-fa and gen-size affected)[/+] on Apr 21, 2026
16. wbruna commented on Apr 21, 2026

    

    wbruna

    on Apr 21, 2026

    Contributor

    More actions

    I'll take a look at the samplers code later. Could you see if the preview images get blank right away, or at the end (or middle) of the generation?
17. SolicTous commented on Apr 22, 2026

    

    SolicTous

    on Apr 22, 2026

    Last edited by SolicTous

    Author

    More actions

    @wbruna
    sampler/scheduler - "default"
    steps = 8
    size = 1024x1024

    	Vae	Proj
    Step - 1
    Step - 2
    Step - 3
    Step - 4
    Step - 5
    Step - 6
    Step - 7
    Step - 8
    Final Image
18. SolicTous commented on Apr 22, 2026

    

    SolicTous

    on Apr 22, 2026

    Author

    More actions

    The problem on last steps looks same as mentioned here #1427 (comment)
19. wbruna commented on Apr 22, 2026

    

    wbruna

    on Apr 22, 2026

    Contributor

    More actions

    This is the sigma schedule for 8 steps, omitting the first (always 1) and last (always 0) values:

    scheduler	1	2	3	4	5	6	7
    discrete	0.960045	0.909207	0.842338	0.750437	0.616212	0.401677	0.00398828
    kl_optimal	0.798406	0.629932	0.483682	0.352475	0.231242	0.116136	0.00398804
    karras	0.566517	0.305216	0.154863	0.0730525	0.0314801	0.0120888	0.00398803
    exponential	0.454204	0.206302	0.0937031	0.0425604	0.0193311	0.00878027	0.00398804
    bong_tangent	0.971458	0.929704	0.863435	0.745319	0.501994	0.21626	0.0733563
    smoothstep	0.988892	0.955832	0.896461	0.800639	0.649923	0.426921	0.155477
    simple	0.965517	0.923077	0.869565	0.8	0.705882	0.571429	0.363636
    sgm_uniform	0.965555	0.923172	0.869747	0.80032	0.706436	0.572407	0.365484

    The model seems to be choking on small sigma values. Happens even with the LCM sampler, which just takes the plain model output for the last step.

    👀React with 👀1Reacted by Erik Scholz
20. wbruna commented on Apr 22, 2026

    

    wbruna

    on Apr 22, 2026

    Contributor

    More actions

    Interesting: Vulkan without Flash Attention works for me. With FA, the second step already produces scrambled output. Disabling FA doesn't seem to make any difference for ROCm.
21. Green-Sky commented on Apr 22, 2026

    

    Green-Sky

    on Apr 22, 2026

    Contributor

    More actions

    Sounds like some value is sensitive to getting cast to f16.
22. SolicTous commented on Apr 23, 2026

    

    SolicTous

    on Apr 23, 2026

    Last edited by SolicTous

    Author

    More actions

    As I investigated it's need a manual convert from original weights because some layers too sensitive to flash attention. With disabling diffusion-fa and manual tensor-type-rules I've got normal generation for all samplers/schedulers (1024 size).

    below example default sampler/scheduler, size 1536x1024
    on 12 gb vram I can't reach more sizes...
    

    🚀React with 🚀1Reacted by Erik Scholz
23. Green-Sky commented on Apr 23, 2026

    

    Green-Sky

    on Apr 23, 2026

    Contributor

    More actions

    and manual tensor-type-rules

    Can you share your rule with us?
24. SolicTous commented on Apr 23, 2026

    

    SolicTous

    on Apr 23, 2026

    Last edited by SolicTous

    Author

    More actions

    disable diffusion-fa
    +
    fixed model
    https://huggingface.co/Solictous/ERNIE-Image-Turbo-GGUF/blob/main/ernie-image-turbo_q4_0.gguf
25. SolicTous

    closed this as completedon Apr 23, 2026