Flux output is blurry compared to webui-forge

[Foul-Tarnished]

I use your workflow (https://comfyanonymous.github.io/ComfyUI_examples/flux/) but with the GGUF extension (https://github.com/city96/ComfyUI-GGUF) Q6_K and t5xxl fp8

I tried to replicate a gen on forge to compare, and it seems it's not just a model/prompt issue: (Check in fullscreen to better see)

ComfyUI

SDWebUI-Forge

It's strange because even though half the generations are blurry on both (model/prompt issue), comfy does appear blurrier, always*. It might be even more visible on sharper image (better prompt)

Other

• Fresh ComfyUI standalone install, then upgraded to Pytorch 12.4, on RTX 4080
• I tried everything I could think of to replicate (CPU noise on forge, same sampling method and schedule type, same sampling steps)
• Could be related to the GGUF extension, as I didn't check on FP8/FP16

[RandomGitUser321]

I commented on your post in the GGUF repo: https://github.com/city96/ComfyUI-GGUF/issues/54#issuecomment-2300129278

[deepfree2023]

Have you tried other resolutions? I tested some SDXL recommended resolutions without blurry results.

[deepfree2023]

Tried some higher resolutions, blurry occured. Need be confirmed.

[ltdrdata]

When I reproduced using the CPU seed, I obtained completely identical results.

ComfyUI:

Forge:

Both were reproduced in the latest version.

Please share more detailed reproduction conditions.

[deepfree2023]

Replacing RandomNoise node by RandomNoise (Inspire) node in Foul-Tarnished's comfyUI workflow, and set noise_mode to cpu, the result was just the same as using RandomNoise node.

So It's already cpu seed in the comfyUI workflow.

[deepfree2023]

This workflow can be loaded to replicate the blurry ComfyUI image:

It would be interesting to find out how can forge produce a sharper image without much detail difference to the blurry one?

I've heard some people reporting higher probability of producing blurry images using dev model, and encounted this kind of problem myself.

[ltdrdata]

This workflow can be loaded to replicate the blurry ComfyUI image:

It would be interesting to find out how can forge produce a sharper image without much detail difference to the blurry one?

I've heard some people reporting higher probability of producing blurry images using dev model, and encounted this kind of problem myself.

Although I haven't tested it, disabling the ModelSamplingFlux node should result in a configuration identical to Forge. And is it really true that the beta schedule was applied in Forge?

[deepfree2023]

This workflow can be loaded to replicate the blurry ComfyUI image: It would be interesting to find out how can forge produce a sharper image without much detail difference to the blurry one? I've heard some people reporting higher probability of producing blurry images using dev model, and encounted this kind of problem myself.

Although I haven't tested it, disabling the ModelSamplingFlux node should result in a configuration identical to Forge. And is it really true that the beta schedule was applied in Forge?

Disabled ModelSamplingFlux node , by using normal, simple or sgm_uniform, the result would be sharper, although not the same as Tarnished's Forge image, and still somewhat blurry.

[ltdrdata]

This workflow can be loaded to replicate the blurry ComfyUI image: It would be interesting to find out how can forge produce a sharper image without much detail difference to the blurry one? I've heard some people reporting higher probability of producing blurry images using dev model, and encounted this kind of problem myself.

Although I haven't tested it, disabling the ModelSamplingFlux node should result in a configuration identical to Forge. And is it really true that the beta schedule was applied in Forge?

Disabled ModelSamplingFlux node , by using normal, simple or sgm_uniform, the result would be sharper, although not the same as Tarnished's Forge image, and still somewhat blurry.

In the beta schedule, both resulted in blurred images, while in the simple schedule, completely identical images were obtained.

When applying the beta schedule, I'm not sure what additional steps Forge takes internally, but there were subtle differences in detail. While ComfyUI produced a slightly more blurred image, Forge's output was also blurred.

It's not that the beta schedule has a tendency to create blurred images. According to my tests, FLUX simply has a tendency to generate blurred images with certain prompts. And when the results are not exactly the same, it just leads to differently blurred images.

[deepfree2023]

Made a few more tests in ComfyUI(based on the previous workflow)：

1. Dev models, at 976 x 1328, the blurry image ratio is higher(about 3/8-4/8 in my tests), using different scheduler may increase or decrease the blurry level, but can not make the image really sharp, ModelSamplingFlux does not affect the blurry level obviously. Adjust pompt seems cannot solve the blurry problem(need more tests).
2. Dev models, at 768 x 1280, the blurry image ratio is lower (about 1/8 in my tests).
3. Schnell, at 976 x 1328 or 768 x 1280, the blurry image ratio is 0.

(need much more tests to reach any serious conclusion)

[deepfree2023]

"FLUX simply has a tendency to generate blurred images with certain prompts." This could be the case. It might be that to get better image quality, there's a tendency to use a shallow depth of field, which makes it easier to go out of focus?

[deepfree2023]

Tested some fine-tuned model, claiming solved blurry problem, and yes, very sharp images without blurry up to now. But they are still in the early stage, and have other problems. This kind of fine-tuning models or loras might be the solution to the blurry image problem.

@Foul-Tarnished Could you take more tests to confirm if Forge has lower blurry image rate with your configuration?

[frankchieng]

I use your workflow (https://comfyanonymous.github.io/ComfyUI_examples/flux/) but with the GGUF extension (https://github.com/city96/ComfyUI-GGUF) Q6_K and t5xxl fp8

I tried to replicate a gen on forge to compare, and it seems it's not just a model/prompt issue: (Check in fullscreen to better see)

ComfyUI

SDWebUI-Forge

It's strange because even though half the generations are blurry on both (model/prompt issue), comfy does appear blurrier, always*. It might be even more visible on sharper image (better prompt)

Other

• Fresh ComfyUI standalone install, then upgraded to Pytorch 12.4, on RTX 4080
• I tried everything I could think of to replicate (CPU noise on forge, same sampling method and schedule type, same sampling steps)
• Could be related to the GGUF extension, as I didn't check on FP8/FP16

choose sgm_uniform or beta of ksampler schduler instead of simple,btw,diffusers has FlowMatchEulerDiscreteScheduler which i think flow match will be another option

[shaun-ba]

I still get this, what's the fix?

[deepfree2023]

I still get this, what's the fix?

Have you tried antiblur lora or something like that?