Update JuggernautXL model to v8

[GuidoBartoli]

KandooAI has released Juggernaut XL v7. Can it replace the default v6 used in Fooocus?

[mashb1t]

it might be replaced after testing, see https://github.com/lllyasviel/Fooocus/discussions/1053 and https://github.com/lllyasviel/Fooocus/issues/1166(https://github.com/lllyasviel/Fooocus/issues/1166#issuecomment-1872540771)

[Amit30swgoh]

KandooAI has released Juggernaut XL v7. Can it replace the default v6 used in Fooocus?

Nice good to know! Whats new in the v7?

[GuidoBartoli]

KandooAI has released Juggernaut XL v7. Can it replace the default v6 used in Fooocus?

Nice good to know! Whats new in the v7?

Quoted from Civitai:

“For Version 7, I've slightly reduced the RunDiffusion Photo Model (from 0.35 to 0.3) and added a new Cinematic Set with 120k training steps. Additionally, I tinkered with some small details. Version 7 doesn't differ greatly from Version 6 but, thanks to its new update, it has significantly increased contrast (there were a few complaints about the desaturated look), and the lighting conditions appear somewhat more natural.”

[c4dee]

Juggernaut XL V8 is about to be released this weekend.

[mashb1t]

Juggernaut XL V8 is about to be released this weekend.

fyi further information on V8: https://www.reddit.com/r/StableDiffusion/comments/18fta92/juggernautxl_v8_early_training_hand_shots/

[hswlab]

It would be really nice to see 5 fingers on one hand more regularly. You would also need something like this for the teeth on your feet :)

[mashb1t]

aaand it's out now: https://civitai.com/models/133005/juggernaut-xl

[hswlab]

I have tested it a bit tonight :) Does fooocus actually do an automatic update to the juggernaut model? When I compare my downloaded version 8 and the version 6 which was downloaded by fooocus, I don't see any difference in the file size, only the name looks different.

[mashb1t]

@hswlab the file size depends on the amount of base images used for training and a few more things (algorithm, precision, optimiser, etc.) for model training. I assume the total hasn't changed => still the same. Fooocus will pull the newest model after the code has been adjusted in a PR and users use entry_with_update.py, but will never delete your old model.

[Den41k92]

I've also tried other models and noticed that Juggernaut V7 is not so good to adhere to different styles and V8 is even worse, however increasing a guidance scale might help bit.

Looks like the styles were fine tuned for the original v6 model.

Here are some comparisons, same settings and seed but different models. From left to right: v6, v7 and v8

Click to expand

lady in the hat : Artstyle Graffiti

```{ "Prompt": "lady in the hat", "Negative Prompt": "", "Fooocus V2 Expansion": "lady in the hat, sharp focus, intricate, elegant, dynamic illumination, highly detailed, colorful, advanced, professional color, strong contrasted, cinematic, fine detail, full background, beautiful, symmetry, clear, pretty, attractive, classy, complex, elaborate, artistic, fancy, dramatic, charming, illuminated, cute, lovely, modern, light, magical, fair", "Styles": "['Fooocus V2', 'Artstyle Graffiti']", "Performance": "Speed", "Resolution": "(1152, 896)", "Sharpness": 3, "Guidance Scale": 4, "ADM Guidance": "(1.5, 0.8, 0.3)", "Base Model": "juggernautXL_version6Rundiffusion.safetensors", "Refiner Model": "None", "Refiner Switch": 0.5, "Sampler": "dpmpp_2m_sde_gpu", "Scheduler": "karras", "Seed": 7018044226794326698, "LoRA 1": "sd_xl_offset_example-lora_1.0.safetensors : 0.1", "Version": "v2.1.862" } ```

  

a house in the forest, white and blue colors : Simple Vector Art, Artstyle Art Nouveau

``` { "Prompt": "a house in the forest, white and blue colors", "Negative Prompt": "", "Fooocus V2 Expansion": "a house in the forest, white and blue colors, cinematic, heavenly atmosphere, stunning, highly detailed, professional, complex, color, cool, intricate, awesome, creative, calm, relaxed, beautiful, symmetry, light, iconic, rich deep lucid, epic, best, pure, brilliant, vibrant, shiny, perfect, fine detail, peaceful, amazing, pristine, very", "Styles": "['Fooocus V2', 'Simple Vector Art', 'Artstyle Art Nouveau']", "Performance": "Quality", "Resolution": "(1344, 704)", "Sharpness": 3.2, "Guidance Scale": 14.44, "ADM Guidance": "(1.5, 0.8, 0.3)", "Base Model": "juggernautXL_version6Rundiffusion.safetensors", "Refiner Model": "None", "Refiner Switch": 0.5, "Sampler": "dpmpp_2m_sde_gpu", "Scheduler": "karras", "Seed": 8531870041007880211, "LoRA 1": "sd_xl_offset_example-lora_1.0.safetensors : 0.11", "Version": "v2.1.862" } ```

  

mythical cat : Artstyle Art Nouveau, Artstyle Psychedelic

``` { "Prompt": "mythical cat", "Negative Prompt": "", "Fooocus V2 Expansion": "mythical cat, full color, highly detailed, excellent composition, cinematic dramatic atmosphere, dynamic light, aesthetic, very inspirational, glowing, majestic, inspiring, stunning, creative, winning, epic, fine detail, clear, perfect, artistic, beautiful, novel, surreal, awarded, best, awesome, singular, sharp, focus, colorful background, peaceful, amazing, illuminated", "Styles": "['Fooocus V2', 'Artstyle Art Nouveau', 'Artstyle Psychedelic']", "Performance": "Speed", "Resolution": "(1152, 896)", "Sharpness": 3.724, "Guidance Scale": 8.15, "ADM Guidance": "(1.5, 0.8, 0.3)", "Base Model": "juggernautXL_version6Rundiffusion.safetensors", "Refiner Model": "None", "Refiner Switch": 0.5, "Sampler": "dpmpp_2m_sde_gpu", "Scheduler": "karras", "Seed": 279914321689178715, "LoRA 1": "sd_xl_offset_example-lora_1.0.safetensors : 0.1", "Version": "v2.1.861" } ```

  

a mythic dog : Artstyle Pop Art

``` { "Prompt": "a mythic dog", "Negative Prompt": "", "Fooocus V2 Expansion": "a mythic dog, detailed, dynamic, dramatic, vibrant colors, cinematic, winning, perfect, artistic, sharp focus, fair, beautiful, emotional, highly detail, pretty, inspired, intricate, innocent, light, iconic, fine, atmosphere, professional, composition, elite, expressive, elegant, very inspirational, colorful, epic, best, stunning, symmetry, illuminated", "Styles": "['Fooocus V2', 'Artstyle Pop Art']", "Performance": "Speed", "Resolution": "(1152, 896)", "Sharpness": 3.724, "Guidance Scale": 8.15, "ADM Guidance": "(1.5, 0.8, 0.3)", "Base Model": "juggernautXL_version6Rundiffusion.safetensors", "Refiner Model": "None", "Refiner Switch": 0.5, "Sampler": "dpmpp_2m_sde_gpu", "Scheduler": "karras", "Seed": 8179423517990499100, "LoRA 1": "sd_xl_offset_example-lora_1.0.safetensors : 0.1", "Version": "v2.1.861" } ```

  

[GuidoBartoli]

I replicated exactly your results and can confirm what you noted, including the fact that increasing the Guidance improves the situation a bit. In fact the tests I had done with version 8 were all with the default styles "Fooocus V2", "Fooocus Sharp" and "Focus Enhance" turned on, I didn't realize the effect of the other styles on the final result.

This is unfortunate, because on a general level I really noticed a tangible increase in image quality from v6 to v8, particularly in overall contrast and hand/face quality. I will do some more testing with other styles to see if this depends on the type (SAI, MRE, DIVA, etc...) or applies to all.

[GuidoBartoli]

I've also tried other models and noticed that Juggernaut V7 is not so good to adhere to different styles and V8 is even worse, however increasing a guidance scale might help bit.

Looks like the styles were fine tuned for the original v6 model.

Here are some comparisons, same settings and seed but different models. From left to right: v6, v7 and v8

Seems that https://github.com/lllyasviel/Fooocus/commit/7b26b292260ac4f1c9acc5a95109e5e3765b9a5c commit introduced v8 as default model, I will make some more tests with styles.