What sampler?

#53
by brennanisgreat - opened

I managed to get the model working, but am getting absolutely terrible results from it. What sampler, scheduler, steps, etc. are you guys using to actually get good results? Just as an example, here's what I'm getting out of the prompts "Verner Herzog's sad beige toys for sad beige children:"
2024-06-12_12-22-06-sd3_medium_incl_clips-1122010006.png
2024-06-12_12-26-35-sd3_medium_incl_clips-202626488.png

Edit - fixed word choice error

Check out the example ComfyUI workflows. They will allow you to reproduce the examples.

"Node name for S&R": "KSampler"
      },
      "widgets_values": [
        945512652412924,
        "fixed",
        28,
        4.5,
        "dpmpp_2m",
        "sgm_uniform",
        1
      ]
    },

I replicated the options exactly and ended up with this. What am I missing here?
ComfyUI_00002_.png

I've had similar results.
Can't create any text like in the cookie example.

Hand and finger anomalies, I tried different settings. What to do?

Same. I'm getting LOTS of severe deformation on humans. I've had some good results with a cityscape and text. The sample images are 1) SDXL 1.0, 2) Copax Timeless 12.2, 3) Bing (DallE 3), and 4) SD3m. There's a huge difference in image quality over SDXL 1.0 and even Bing. If there's any improvement over Timeless Copax 12.2, it's going to be a matter of taste. Based on these few samples, Bing handled text better still, but 3m has a much better image quality. I'm going to play around with some human figures next to see what I come up with.
SDXL 1.0
2024-06-12_15-53-15-sd_xl_base_1.0-1883309831.png
Copax Timeless 12.2
2024-06-12_15-51-47-copaxTimelessxlSDXL1_v122-74516430.png
Bing
_be70a29a-ec0b-4e1d-a313-f20d01a0e9d9.jpg
SD3m
2024-06-12_15-56-14-sd3_medium_incl_clips-1111326293.png

I've tried to make some human scenarios, and SD3m is much, much worse at people than scapes. So here I have 1) SDXL 1.0, 2) Copax Timeless 12.2, 3) Bing/DallE 3, and 4) SD3m. As you can see, SDXL does a decent job of human shape but is shit at text, Copax timeless has good image and text quality, but the scenario is a little static, and Bing excels at text but its images just always looks hazy. SD3m just completely struggled with people. All the human images were massively oversaturated and while it had a better scene composition, the image just looks weird and unrealistic. This has been consistent across any prompt I give that involves humans. Clearly the sample images prove that it's possible to make realistic looking human images, but how?

SDXL 1.0
2024-06-12_16-39-01-sd_xl_base_1.0-2126079882.png

Copax Timeless 12.2
2024-06-12_16-37-28-copaxTimelessxlSDXL1_v122-1610323061.png

Bing
_57d95213-316b-4a3b-b991-5a09c4b5b6a3.jpg

SD3m
2024-06-12_16-44-19-sd3_medium_incl_clips-1802453620.png

can u share .json for workflow please?

I have bad results with what is in the repository

drop_aneonsigninthewindowread_50_20240620_124311.jpg

SD3
"A neon sign in the window reads "Shock Massagers" in a retro, diner-style font. Inside, two women chat behind a gleaming chrome bar counter. Separate robot bartenders with glowing red eyes zip around behind them, mixing drinks with robotic precision."

Sign up or log in to comment