Adetailer not working for Flux D. Alternatives?

[evanheckert]

My loras are incredible for face close-ups, but very hit and miss when further away. Adetailer was always the perfect solution for SD1.5, but I only get black boxes when trying to use it for Flux.

Does anyone have a good alternative that works in Forge?

[freakstritch]

adetailer works for me in forge using flux1-dev-Q8_0.gguf and loras. i have 3090 and 96GB system ram

[freakstritch]

evanheckert = try a new simple prompt that has a face and see if that works.

vagrin - try lowering strength of LORAs and you may be able to use them - i use several loras with flux at the same time

[evanheckert]

Interesting, error in the logs. It looks like it's actually processing the face, but then hits an error, presumably when merging the result. Here's everything from the original image completion on:

Moving model(s) has taken 4.29 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 22/22 [00:25<00:00,  1.16s/it]
[Unload] Trying to free 4355.84 MB for cuda:0 with 0 models keep loaded ... Current free memory is 11081.72 MB ... Done.
[Memory Management] Target: IntegratedAutoencoderKL, Free GPU: 11081.72 MB, Model Require: 159.87 MB, Previously Loaded: 0.00 MB, Inference Require: 4148.00 MB, Remaining: 6773.85 MB, All loaded to GPU.
Moving model(s) has taken 0.44 seconds

0: 640x480 1 face, 65.8ms
Speed: 2.0ms preprocess, 65.8ms inference, 15.1ms postprocess per image at shape (1, 3, 640, 480)
WARNING:root:Sampler Scheduler autocorrection: "Euler" -> "Euler", "None" -> "Automatic"
[Unload] Trying to free 4148.00 MB for cuda:0 with 1 models keep loaded ... Current free memory is 10881.18 MB ... Done.
Skipping unconditional conditioning when CFG = 1. Negative Prompts are ignored.
[Unload] Trying to free 16682.57 MB for cuda:0 with 0 models keep loaded ... Current free memory is 10898.06 MB ... Unload model KModel Current free memory is 23017.60 MB ... Done.
[Memory Management] Target: JointTextEncoder, Free GPU: 23017.60 MB, Model Require: 9641.98 MB, Previously Loaded: 0.00 MB, Inference Require: 4148.00 MB, Remaining: 9227.62 MB, All loaded to GPU.
Moving model(s) has taken 5.85 seconds
Distilled CFG Scale: 3.5
[Unload] Trying to free 19903.36 MB for cuda:0 with 0 models keep loaded ... Current free memory is 13370.87 MB ... Unload model IntegratedAutoencoderKL Current free memory is 13530.74 MB ... Unload model JointTextEncoder Done.
[Memory Management] Target: KModel, Free GPU: 23172.72 MB, Model Require: 12119.51 MB, Previously Loaded: 0.00 MB, Inference Require: 4148.00 MB, Remaining: 6905.21 MB, All loaded to GPU.
Moving model(s) has taken 5.37 seconds
100%|████████████████████████████████████████████████████████████████████████████████████| 9/9 [00:10<00:00,  1.12s/it]
[Unload] Trying to free 4355.84 MB for cuda:0 with 0 models keep loaded ... Current free memory is 11050.93 MB ... Done.
[Memory Management] Target: IntegratedAutoencoderKL, Free GPU: 11050.93 MB, Model Require: 159.87 MB, Previously Loaded: 0.00 MB, Inference Require: 4148.00 MB, Remaining: 6743.05 MB, All loaded to GPU.
Moving model(s) has taken 0.15 seconds
C:\Users\evanh\git\forge\webui\modules\processing.py:1010: RuntimeWarning: invalid value encountered in cast
  x_sample = x_sample.astype(np.uint8)
Total progress: 100%|██████████████████████████████████████████████████████████████████| 22/22 [00:48<00:00,  2.20s/it]
Total progress: 100%|██████████████████████████████████████████████████████████████████| 22/22 [00:48<00:00,  1.20s/it]

[evanheckert]

Per copilot's suggestion, inserted a check for NaN's or Infinity at that part of the process_images_inner(p) function:

        for i, x_sample in enumerate(x_samples_ddim):
                p.batch_index = i

                x_sample = 255. * np.moveaxis(x_sample.cpu().numpy(), 0, 2)

                 # Check for NaNs or infinities
                if np.isnan(x_sample).any() or np.isinf(x_sample).any():
                    raise ValueError("x_sample contains NaNs or infinities")

                # Clamp values to the range [0, 255]
                x_sample = np.clip(x_sample, 0, 255)

                x_sample = x_sample.astype(np.uint8)

and shore'nuf:

ValueError: x_sample contains NaNs or infinities 

I'm comfortable with python, but haven't done anything AI/tensors with it, so not much use beyond that to get to an actual solution, I'm afraid.

    ╭──────────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
    │                                                   System info                                                    │
    │ ┏━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓ │
    │ ┃             ┃ Value                                                                                          ┃ │
    │ ┡━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩ │
    │ │    Platform │ Windows-10-10.0.22631-SP0                                                                      │ │
    │ │      Python │ 3.10.6 (tags/v3.10.6:9c7b4bd, Aug  1 2022, 21:53:49) [MSC v.1932 64 bit (AMD64)]               │ │
    │ │     Version │ f2.0.1v1.10.1-previous-501-g668e87f9                                                           │ │
    │ │      Commit │ 668e87f920be30001bb87214d9001bf59f2da275                                                       │ │
    │ │ Commandline │ ['launch.py', '--listen', '--port', '7862', '--cuda-malloc']                                   │ │
    │ │   Libraries │ {'torch': '2.3.1+cu121', 'torchvision': '0.18.1+cu121', 'ultralytics': '8.2.82', 'mediapipe':  │ │
    │ │             │ '0.10.14'}                                                                                     │ │
    │ └─────────────┴────────────────────────────────────────────────────────────────────────────────────────────────┘ │
    │    

[Vagrin]

vagrin - try lowering strength of LORAs and you may be able to use them - i use several loras with flux at the same time

Great idea. I gave it a try and it didn't work but it was a valid concept. I brought the lora down to .25. No dice. Anything lower than that would be pointless. Thanks for the idea.

[evanheckert]

Curious what version of Torch you guys are using, and which command line args?

[CCpt5]

uDetailer is an alternative fork for Adetailer. I haven't had a chance to test it w/ Forge nor Flux yet.

That said, the developer is very receptive/active and will address bugs/issues if mentioned in the issue page. He also coded "Model Mixer" and fixed issues I'd have w/ that from time to time after updates very quickly.

https://github.com/wkpark/uddetailer