Awful Flux Fill results. (blurry, grainy results)

[JackDainzh]

Expected Behavior

A proper image being generated without the said artifact (blurry, grainy). Like on the image above when generating on Replicate Space.

Actual Behavior

Generates images with awful, mushy quality.

Steps to Reproduce

Use a standard (recommended) workflow for flux fill. You can load either a standard 22 gb dev model, or use a GGUF one, the results are nearly the same.

Debug Logs

no errors during and after generation.

Other

Have encountered another person on CivitAI using an community made FP8 version of this model, who describes that the results are exactly the same as I am encountering.

[yincangshiwei]

It seems that not all GPUs are compatible, and there may be issues under certain seeds, only a few specific seeds are normal.
#5715

[yincangshiwei]

It should be related to this:
black-forest-labs/flux#195

[JackDainzh]

It should be related to this: black-forest-labs/flux#195

Strange, I have RTX 4070 gpu, should work fine I think? Because 4070 has Ada Lovelace architecture, which should work as expected.

[Mozer]

I have 2080ti-22GB and flux fill outpaint works bad (grainy) with high resolutions of the final image (more than 700x700). With lower res (632x696) it looks almost OK but still not perfect. With my 3060-12GB it works perfect. I tried fp16, fp8 and gguf_q4_0 checkpoints. Clip is fp8. Inpaint also suffers when high res.
I tried with fp16 checkpoint in default mode - same bad results.

Python version: 3.11.9
pytorch version: 2.5.1+cu121
comfyUI v0.3.4

[taiyouZhang]

me too, always wierd blurry

[catbaba777]

I tried many times, depending on the resolution and expanding the area of ​​the image. Same goes for partial repaints. Noise appears when large areas are redrawn. Simple compositions and lower raw images are easier to achieve good results at 768 or 512 resolution. Additionally, it relies on random seeds. In short, very uncontrollable. .

[Mozer]

I have successfully used Flux fill 1024x1024 20 steps on my 2080ti-22GB using this diffusers-bnb-nf4 code https://huggingface.co/sayakpaul/FLUX.1-Fill-dev-nf4
No noise on the sides that i have with comfy. But nf4 quality in the center is degraded a bit, compared to fp8 version that i have on my 3060. Inpaint and outpaint works. Inpaint looks better. uses 16 GB of vram. Original flux dev in diffusers version is needed to run this code.

python 3.11.10, cuda 11.8, torch: 2.5.1+cu118

[willhsmit]

Buried in a Swarmui commit
mcmonkeyprojects/SwarmUI@d83fe84

is the comment:

• Boosting the Flux Guidance Scale way up to eg 30 may improve quality

I can confirm that making this change is a substantial improvement for me on a few test generations.

[JackDainzh]

All these generations that are shown in my examples were made with Guidance of 30.

[willhsmit]

Are you inpainting or outpainting? I ask because normally I expect a flux fill output to contain a mix of content leftover from the original image and generated content, and yours look uniformly mushy everywhere.

[JackDainzh]

That was inpainting with the whole image masked, to show the issue more clearly. As for generating with small patches, it's the same, new generated objects are mushy too

[comfyanonymous]

Should be fixed now. The flux inpaint model does not work in fp16 unlike the other flux models. The fix is to make it inference in fp32 on cards that do not support bf16 which is slow but will make it work.

[Mozer]

Thx, I can confirm it is working for my 2080ti now, though slower.