FLUX.1 Fill dev inpainting not working - mask completely ignored

[kulvil7988-svg]

Problem

FLUX.1 Fill dev model completely ignores the mask during inpainting. The output image is identical to the input - no changes are applied to the masked area.

Setup

• Model: FLUX.1-Fill-dev_flux1-fill-dev.safetensors (23.2 GB)
• CLIP: t5xxl_fp16.safetensors + clip_l.safetensors (type: flux)
• VAE: ae.safetensors (319 MB)
• ComfyUI version: Latest (June 2026)
• GPU: NVIDIA

Workflow

Using official workflow from ComfyUI_examples/flux/flux_fill_inpaint_example.png

Settings:

• Flux Guidance: 30.0
• KSampler: cfg=1.0, scheduler=normal, euler sampler, 20 steps, denoise=1.0
• Differential Diffusion [BETA]: intensity 1.0
• InpaintModelConditioning: noise_mask=false

What I tried

1. Different prompts ("a woman wearing a pink shirt", "nude woman", etc.)
2. Recreated mask via MaskEditor
3. Saved mask to disk and reloaded
4. Removed LoRA
5. Tried different scheduler (simple, normal, beta)
6. Changed noise_mask to true

Result

No errors in console, generation completes in ~28 seconds, but output is identical to input. The masked area (white coat) remains white instead of being replaced with "pink shirt" as per prompt.

Logs

[INFO] Model Flux prepared for dynamic VRAM loading. 22701MB Staged.
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:18<00:00, 1.08it/s]
[INFO] Prompt executed in 27.85 seconds

No errors, only warnings about clip missing 'text_projection.weight'

Question

Is this a bug in the latest ComfyUI version? Or am I missing something in the workflow configuration?

[aleph23]

I just ran an outpaint that did what it should, with one caveat: It turned to included background into a wall-poster on an absolute-black wall. The clear over-cropped parts I was trying to recover did recover just fine though.

I'm guessing from this behavior and your, that the model needs something to attach to. Try making your mask itself noisy. Like make your mask 127 grey and then apply a gaussian-shaped variance across luminance (or any other method to get a uniformly-distributed (obviously on a higher than per-pixel level), visible, scattered variance).

Like either left or, since you want pink, right:

[guimaraesdeborah]

same problem here, how can I solve this problem?

[rskickflip1-cmd]

ComfyUI version? You are using the mask editor to make the mask and not uploading an image with mask on it? He should not be getting an image returned with a white mask on it, the load image node does not pass the mask layered on the image output.

[ricesana]

Same here, Comfy Desktop 1.0.34.
"You are using the mask editor to make the mask and not uploading an image with mask on it?" Tried 'both':

• edited and feed it directly in Image Inpainting (Flux.1 Fill Dev) = Same image as result, nothing changed
• edited, feed to a MaskToImage to check and feet to Image Inpainting (Flux.1 Fill Dev) through a ImageToMask = Same image as result, nothing changed