Rtx 2060 qwen image edit black output *never used sageattention*

[realisticdreamer114514]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

Image able to produce normal preview from step 2 of inference & normal output saved after vae decode

Actual Behavior

Preview is black thruout inference & only black output is generated

Steps to Reproduce

1. Install calcuis/gguf city96/ComfyUI-GGUF extension for gguf loading and pollockjj/ComfyUI-MultiGPU for manual offloading, offloading ext is necessary to infer on my device as disabling will consistently OOM, gguf might be replaced by safetensors on another device & would likely still replicate since it's unlikely te at fault
2. Obtain model files: diffusion model, te te mmproj, vae, lightning lora
3. python main.py --preview-method auto --force-fp16 --disable-cuda-malloc --windows-standalone-build
4. Run workflow from the screencap: (I avoid uploading image with metadata or json)

Debug Logs

Console (cmd in conda environment):

(comfyui) PS D:\webui-forge\ComfyUI_windows_portable\ComfyUI> .\run

D:\webui-forge\ComfyUI_windows_portable\ComfyUI>python main.py --preview-method auto --force-fp16 --disable-cuda-malloc --windows-standalone-build
Adding extra search path checkpoints D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\Stable-diffusion
Adding extra search path configs D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\Stable-diffusion
Adding extra search path vae D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\VAE
Adding extra search path loras D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\Lora
Adding extra search path loras D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\LyCORIS
Adding extra search path upscale_models D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\ESRGAN
Adding extra search path upscale_models D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\RealESRGAN
Adding extra search path upscale_models D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\SwinIR
Adding extra search path embeddings D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\embeddings
Adding extra search path hypernetworks D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\hypernetworks
Adding extra search path controlnet D:\webui-forge\ComfyUI_windows_portable\ComfyUI\path\to\stable-diffusion-webui\models\ControlNet
Adding extra search path checkpoints D:\webui-forge\webui\models\Stable-diffusion
Adding extra search path configs D:\webui-forge\webui\models\Stable-diffusion
Adding extra search path vae D:\webui-forge\webui\models\VAE
Adding extra search path loras D:\webui-forge\webui\models\Lora
Adding extra search path loras D:\webui-forge\webui\models\LyCORIS
Adding extra search path upscale_models D:\webui-forge\webui\models\ESRGAN
Adding extra search path upscale_models D:\webui-forge\webui\models\RealESRGAN
Adding extra search path upscale_models D:\webui-forge\webui\models\SwinIR
Adding extra search path embeddings D:\webui-forge\webui\embeddings
Adding extra search path hypernetworks D:\webui-forge\webui\models\hypernetworks
Adding extra search path controlnet D:\webui-forge\webui\models\ControlNet
Adding extra search path clip D:\webui-forge\webui\models\text_encoder

Checkpoint files will always be loaded safely.
Total VRAM 6144 MB, total RAM 32605 MB
pytorch version: 2.8.0+cu126
xformers version: 0.0.32.post2
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 2060 : native
Enabled pinned memory 14672.0
Using xformers attention
Python version: 3.11.13 | packaged by Anaconda, Inc. | (main, Jun  5 2025, 13:03:15) [MSC v.1929 64 bit (AMD64)]
ComfyUI version: 0.3.68
ComfyUI frontend version: 1.28.8
[Prompt Server] web root: C:\Users\USER\anaconda3\envs\comfyui\Lib\site-packages\comfyui_frontend_package\static
2025-11-06 18:27:17.392247: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
From C:\Users\USER\anaconda3\envs\comfyui\Lib\site-packages\tf_keras\src\losses.py:2976: The name tf.losses.sparse_softmax_cross_entropy is deprecated. Please use tf.compat.v1.losses.sparse_softmax_cross_entropy instead.

Unable to parse pyproject.toml due to lack dependency pydantic-settings, please run 'pip install -r requirements.txt': Illegal character '\n' (at line 3, column 101)
ComfyUI-GGUF: Allowing full torch compile
[MultiGPU Core Patching] Patching mm.soft_empty_cache for Comprehensive Memory Management (VRAM + CPU + Store Pruning)
[MultiGPU Core Patching] Patching mm.get_torch_device, mm.text_encoder_device, mm.unet_offload_device
[MultiGPU DEBUG] Initial current_device: cuda:0
[MultiGPU DEBUG] Initial current_text_encoder_device: cuda:0
[MultiGPU DEBUG] Initial current_unet_offload_device: cpu
[MultiGPU] Initiating custom_node Registration. . .
-----------------------------------------------
custom_node                   Found     Nodes
-----------------------------------------------
ComfyUI-LTXVideo                  N         0
ComfyUI-Florence2                 N         0
ComfyUI_bitsandbytes_NF4          N         0
x-flux-comfyui                    N         0
ComfyUI-MMAudio                   N         0
ComfyUI-GGUF                      Y        18
PuLID_ComfyUI                     N         0
ComfyUI-WanVideoWrapper           N         0
-----------------------------------------------
[MultiGPU] Registration complete. Final mappings: CheckpointLoaderAdvancedMultiGPU, CheckpointLoaderAdvancedDisTorch2MultiGPU, UNetLoaderLP, UNETLoaderMultiGPU, VAELoaderMultiGPU, CLIPLoaderMultiGPU, DualCLIPLoaderMultiGPU, TripleCLIPLoaderMultiGPU, QuadrupleCLIPLoaderMultiGPU, CLIPVisionLoaderMultiGPU, CheckpointLoaderSimpleMultiGPU, ControlNetLoaderMultiGPU, DiffusersLoaderMultiGPU, DiffControlNetLoaderMultiGPU, UNETLoaderDisTorch2MultiGPU, VAELoaderDisTorch2MultiGPU, CLIPLoaderDisTorch2MultiGPU, DualCLIPLoaderDisTorch2MultiGPU, TripleCLIPLoaderDisTorch2MultiGPU, QuadrupleCLIPLoaderDisTorch2MultiGPU, CLIPVisionLoaderDisTorch2MultiGPU, CheckpointLoaderSimpleDisTorch2MultiGPU, ControlNetLoaderDisTorch2MultiGPU, DiffusersLoaderDisTorch2MultiGPU, DiffControlNetLoaderDisTorch2MultiGPU, UnetLoaderGGUFDisTorchMultiGPU, UnetLoaderGGUFAdvancedDisTorchMultiGPU, CLIPLoaderGGUFDisTorchMultiGPU, DualCLIPLoaderGGUFDisTorchMultiGPU, TripleCLIPLoaderGGUFDisTorchMultiGPU, QuadrupleCLIPLoaderGGUFDisTorchMultiGPU, UnetLoaderGGUFDisTorch2MultiGPU, UnetLoaderGGUFAdvancedDisTorch2MultiGPU, CLIPLoaderGGUFDisTorch2MultiGPU, DualCLIPLoaderGGUFDisTorch2MultiGPU, TripleCLIPLoaderGGUFDisTorch2MultiGPU, QuadrupleCLIPLoaderGGUFDisTorch2MultiGPU, UnetLoaderGGUFMultiGPU, UnetLoaderGGUFAdvancedMultiGPU, CLIPLoaderGGUFMultiGPU, DualCLIPLoaderGGUFMultiGPU, TripleCLIPLoaderGGUFMultiGPU, QuadrupleCLIPLoaderGGUFMultiGPU
Nvidia APEX normalization not installed, using PyTorch LayerNorm

Import times for custom nodes:
   0.0 seconds: D:\webui-forge\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-GGUF
   0.1 seconds: D:\webui-forge\ComfyUI_windows_portable\ComfyUI\custom_nodes\gguf
   0.1 seconds: D:\webui-forge\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-MultiGPU

Context impl SQLiteImpl.
Will assume non-transactional DDL.
No target revision found.
Starting server

To see the GUI go to: http://127.0.0.1:8188
got prompt
Using xformers attention in VAE
Using xformers attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.float16
gguf qtypes: Q6_K (1), F32 (141), IQ4_XS (166), Q5_K (31)
Using mmproj 'mmproj-qwen2.5-vl-7b-it-q4_0.gguf' for 'qwen2.5-vl-7b-it-iq4_xs.gguf'.
gguf qtypes: Q4_0 (192), F32 (291), F16 (34), Q8_0 (2)
[MultiGPU Core Patching] text_encoder_device_patched returning device: cuda:0 (current_text_encoder_device=cuda:0)
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load WanVAE
0 models unloaded.
loaded partially; 128.00 MB usable, 128.00 MB loaded, 114.00 MB offloaded, lowvram patches: 0
Requested to load QwenImageTEModel_
loaded completely; 3541.80 MB usable, 4469.60 MB loaded, full load: True
[MultiGPU Core Patching] Successfully patched ModelPatcher.partially_load
gguf qtypes: F16 (1093), Q8_0 (30), Q4_K (692), Q5_K (118)
model weight dtype torch.float16, manual cast: None
model_type FLUX
[MultiGPU DisTorch V2] Full allocation string: #cuda:0;8.0;cpu
[MultiGPU DisTorch V2] GGUFModelPatcher missing 'model_patches_models' attribute, using 'model_patches_to' fallback.
Requested to load QwenImage
===============================================
    DisTorch2 Model Virtual VRAM Analysis
===============================================
Object   Role   Original(GB) Total(GB)  Virt(GB)
-----------------------------------------------
cuda:0   recip       6.00GB   14.00GB   +8.00GB
cpu      donor      31.84GB   23.84GB   -8.00GB
-----------------------------------------------
model    model      11.62GB    3.62GB   -8.00GB


[MultiGPU DisTorch V2] Model size (11.62GB) is larger than 90% of available VRAM on: cuda:0 (5.40GB).
[MultiGPU DisTorch V2] To prevent an OOM error, set 'virtual_vram_gb' to at least 6.22.


==================================================
[MultiGPU DisTorch V2] Final Allocation String:
cuda:0,0.6040;cpu,0.2512
==================================================
    DisTorch2 Model Device Allocations
==================================================
Device    VRAM GB    Dev %   Model GB    Dist %
--------------------------------------------------
cuda:0       6.00    60.4%       3.62     31.2%
cpu         31.84    25.1%       8.00     68.8%
--------------------------------------------------
    DisTorch2 Model Layer Distribution
--------------------------------------------------
Layer Type         Layers   Memory (MB)   % Total
--------------------------------------------------
Linear                846      12010.58    100.0%
RMSNorm               241          0.07      0.0%
LayerNorm             241          0.00      0.0%
--------------------------------------------------
DisTorch2 Model Final Device/Layer Assignments
--------------------------------------------------
Device             Layers   Memory (MB)   % Total
--------------------------------------------------
cuda:0 (<0.01%)       484          0.82      0.0%
cuda:0                264       3874.96     32.3%
cpu                   580       8134.87     67.7%
--------------------------------------------------
[MultiGPU DisTorch V2] DisTorch loading completed.
[MultiGPU DisTorch V2] Total memory: 12010.65MB
100%|████████████████████████████████████████████████████████████████████████| 3/3 [01:23<00:00, 27.73s/it]
[MultiGPU DisTorch V2] ModelPatcher missing 'model_patches_models' attribute, using 'model_patches_to' fallback.
Requested to load WanVAE
0 models unloaded.
loaded partially; 128.00 MB usable, 128.00 MB loaded, 114.00 MB offloaded, lowvram patches: 0
Prompt executed in 399.30 seconds

Other

This is not a duplicate of previous similar bugs with workaround:

• Since my gpu is too old to support sageattention this incompatibility is bypassed (many threads here & on reddit point to this as the cause of a bug that has identical output)
• I never added --fast during inference & testing
• I tested these flag combinations --force-fp16 --force-fp16 --fp32-unet --force-fp16 --fp32-vae, only --force-fp16 --fp32-unet produced a normal preview and output, this implies it's likely qwen diffusion model's issue with fp16 inference

As seen in the successful test run, the way to bypass this is --fp32-unet, at tremendous speed cost (~2-5x)

Possibly this is an issue with qwen model's fp16 inference, or implementation of fp16 inference in comfy, or quantization issue (might be from either fp32 or bf16), or specific rtx20 series-only fp16 issues, might need some more narrowing down

(Apologies for having to include some exts dealing with hardware constraints, someone might test on rtx 20 series cards with more vram)

[realisticdreamer114514]

Ps 1 there is a redundant lightning lora for the merge I used for the op, bypassing it doesnt change anything

Ps 2 tested again with QuantStack/Qwen-Image-Edit-2509-GGUF and only ComfyUI-GGUF + ComfyUI-MultiGPU, can replicate

I'm very confused at what works & doesn't work with my hardware & sorry for not being that clean with my testing

[rattus128]

Can you try without multi-gpu and GGUF node packs, start with --no-vram and see if it works? It will be slower, but at least you should get an image. If you still dont get an image, we can take it as a pure comfy core bug. If you do get an image, re-introduce your custom nodes and see whats making the difference.

Offloads are reasonably dangerous with --disable-cuda-malloc. You might find removing that changes things too.

[realisticdreamer114514]

Can you try without multi-gpu and GGUF node packs, start with --no-vram and see if it works

I tested again with the official workflow, fp8 diffusion model & te from comfy-org's hf repo and python main.py --preview-method auto --force-fp16 --windows-standalone-build --disable-all-custom-nodes --novram

te was loaded & encoded prompt successfully, but sadly it crashed while loading the diffusion model & quit without any console/popup error, this is probably a distinct ram issue that should get help on its own before I go on & test again

--disable-cuda-malloc

I dont usually have this flag when I use comfy, this was for testing to see if cuda-malloc is the issue, & whether cuda-malloc is enabled or not changes nothing about the black output after more testing (with the exts back on tho, if it was really just cuda-malloc whether the ext are enabled wont matter)

[realisticdreamer114514]

Btw speaking of the crash I downloaded the portable version before cu128 was a thing, unzipped it where it is, & set up a conda env + git pull on the comfy git repo anyway. But I tested a clean portable install (nvidia_cu128 one) + clean conda venv (cu128 py312 torch290) git pulled & the crash is still there to stop me from testing the fp8 diffusion model in any way

Can anyone with rtx20 cards & more vram help out with testing?

[realisticdreamer114514]

QwenLM/Qwen-Image#68 (comment)

I suspect --force-fp16 and/or --fp8_e4m3fn-unet

For me this started happening after adding --force-fp16 flag, so looking more like a model problem

[realisticdreamer114514]

Not even --force-upcast-attention fixed this, wow
I'm reminded of a solution that's better off in a new issue