Comfy-Org

/

Mage-Flow

Like 123

Follow

Comfy Org 13.3k

well you are using the turbo version which should have 1 cfg and 4-6 steps

For me step 1 has somewhat coherent image and then it devolves into a cloudy mess with each further step.

For txt2img turbo model, not edit.

For me step 1 has somewhat coherent image and then it devolves into a cloudy mess with each further step.

For txt2img turbo model, not edit.

This happened to me too, but when I tried euler sampler and linear_quadratic scheduler, the problem was solved, although the quality is still unsatisfactory.

Try just euler/simple for the turbo, that gives the exact schedule the original code uses:

well you are using the turbo version which should have 1 cfg and 4-6 steps.

Thanks, I forgot I am using turbo version.

1024 x 1024 seems not a good resolution.

get the much better result at 2048 x 2048

Try just euler/simple for the turbo, that gives the exact schedule the original code uses:

Euler/simple was the first thing I tried, that's where the issue is. I'm using the normal turbo model so I have omitted the TextEncodeMageFlow node and went with the default ComfyUI workflow.

Euler/linear_quadratic like Job0115 said seems to produce a proper image after 4 steps (compared to euler/simple producing clouds), I can't tell if that's the best quality the model is capable of.

I thought it might be an issue with int8 convrot so I tried with bf16 but it's still the same result.

Euler/linear_quadratic 4 steps

Euler/simple 4 steps

Euler/simple 1 step

GPU: GTX 1660 Super

I really don't know what to tell you... are you using any startup arguments?

I really don't know what to tell you... are you using any startup arguments?

Other than --enable-manager, nothing else

Are all of you having the issue using GPU that does not support BF16? I know GTX 1660 Super doesn't.

I've confirmed it breaks if compute dtype is fp32, which it will be on GPUs not supporting bf16, this PR should fix the fully corrupt outputs: https://github.com/Comfy-Org/ComfyUI/pull/15081

Just proof of concept samples for the —turbo models...

Some findings (maybe?) in very limited testing:

• euler_ancestral_cfg_pp + linear_quadratic gave good results with about 7 steps.
• er_sde + linear_quadratic as a refiner with about 2 steps seemed to clean up the image even better.
• Don't remember which ones I tried, but some other schedulers besides linear_quadratic caused the image to fall apart completely.
• In editing, er_sde + linear_quadratic with about 9 steps and CFG 1.2 stayed fairly true to the input image and corrected the replacement text.
• Perhaps a happy accident, but CLIP skip does something ...-1, -2, -3 all gave variations, while different seeds and other settings hardly did; however, less than -3 and the image deconstructed.

I've confirmed it breaks if compute dtype is fp32, which it will be on GPUs not supporting bf16, this PR should fix the fully corrupt outputs: https://github.com/Comfy-Org/ComfyUI/pull/15081

That fixed it, thanks!

Dis not good model, qwen nunchaku or flux2 are way better quality and actually faster ! this is running meh with convrot on 3090 compared to flux2 and qwen edit , sure it is small but... with comfy dynbamic vram imo skip it use qwen edit or flux2 if you want cleaner vae, cause this models vae is SD 1.5 level

Dis not good model, qwen nunchaku or flux2 are way better quality and actually faster ! this is running meh with convrot on 3090 compared to flux2 and qwen edit , sure it is small but... with comfy dynbamic vram imo skip it use qwen edit or flux2 if you want cleaner vae, cause this models vae is SD 1.5 level

bro this model literally uses flux 2 vae space...

if you're getting nonsense results it could be your attention backend.
https://huggingface.co/Comfy-Org/Mage-Flow/discussions/11