"AstolfoRF" Unconditional Image Generation up to SDXL (and more)

parameters

,
Negative prompt: ,
Steps: 64, Sampler: Euler, Schedule type: Automatic, CFG scale: 4, Seed: 1703748712, Size: 1024x1024, Model hash: 82e7f6bcb3, Model: x1a-AstolfoRF-3ep, VAE hash: 235745af8d, VAE: sdxl-vae-fp16-fix.vae.safetensors, Denoising strength: 0.7, Hires CFG Scale: 4, Hires upscale: 1.5, Hires upscaler: Latent, advanced_sampling_enabled: True, advanced_sampling_mode: SD3, sd3_shift: 3, dynthres_enabled: True, dynthres_simple_mode: False, dynthres_mimic_scale: 4, dynthres_threshold_percentile: 1, dynthres_mimic_mode: Constant, dynthres_mimic_scale_min: 0, dynthres_cfg_mode: Constant, dynthres_cfg_scale_min: 0, dynthres_sched_val: 1, dynthres_separate_feature_channels: enable, dynthres_scaling_startpoint: MEAN, dynthres_variability_measure: AD, dynthres_interpolate_phi: 0.3, sag_enabled: True, sag_scale: 0.5, sag_blur_sigma: 2, Version: f1.0.0v2-v1.10.1RC-latest-2502-g630cf49b

An informal research about unconditional image generation with Stable Diffusion, or "AI", up to UNet based SDXL. Such observation may be useful for "data visualization" to show that how the "number" works. Please be skeptic on this repo. My explanantion in CivitAI.

parameters

Steps: 48, Sampler: Euler, Schedule type: Automatic, CFG scale: 3, Seed: 3649863581, Size: 1024x1024, Model hash: e276a52700, Model: x72a-AstolfoMix-240421-feefbf4, VAE hash: 26cc240b77, VAE: sd_xl_base_1.0.vae.safetensors, Clip skip: 2, FreeU Stages: "[{\"backbone_factor\": 1.1, \"skip_factor\": 0.6}, {\"backbone_factor\": 1.2, \"skip_factor\": 0.4}]", FreeU Schedule: "0.0, 1.0, 0.0", FreeU Version: 2, Dynamic thresholding enabled: True, Mimic scale: 1, Separate Feature Channels: False, Scaling Startpoint: MEAN, Variability Measure: AD, Interpolate Phi: 0.3, Threshold percentile: 100, PAG Active: True, PAG Scale: 1, Version: v1.9.3

Negative prompt: (bad:0), (comic:0), (cropped:0), (error:0), (extra:0), (low:0), (lowres:0), (speech:0), (worst:0)
Steps: 32, Sampler: Euler, CFG scale: 10.5, Seed: 1337, Size: 512x512, Model hash: 925997e9, Clip skip: 2

Pixiv album for storing the images

(New) Observation on PonyXL V6

Major contents

No explaination. Read the articles instead.

• Generic research methods (CFG-STEP scan) when an unknown anime model is received.
• My hands-on experience on txt2img only. Another prompting research.
• Generic prompting guide for a webui compatable "anime model". Core concept (heck what application will support negative prompts?) is viable.
• Docuementry, journal and ranting. Read for drama. Actually some of them are primary sources.
• Astolfo is a good boy.
• Beware of random docuement style because I don't have time to explain or even expand it.
• Also beware model hasing algorithm has been changed into SHA256 for entire model. I'll use the new hash, but old hash remains (usually they are famous model isn't it?)

Index

Too lazy to update constantly. Just iterlate the directories. You will find the pattern.

• ch00: My content is probably not popular / legit and even completely non-sense. You shold leave if you want nice AI art. Why I'm writing all these stuffs.
• ch01: Common content across models. Most theory / explaination / derive goes there.
• ch02: Model specific contents. Assumed you've already read ch01 and know the context.
• ch03: Data analysis. Usually involves model comparasion.
• ch04: A (very dumb) batch script for multiple WebUI instances.
• ch05: Astolfo mix. Existing technology, original idea.
• ch06: Finetuning in vanilla approach. Guess what? Original content! What?
• ch97: Uncategorized contents. Usually "not article".
• ch98: Backup from discord server because I think it is not safe to leave them there forever. Removed. Merged / moved to different articles.
• ch99: A "House-Tree-Person" test towards SD, as well as the oldest articles. Still following this pracice today.

So where to start?

• Take the blue pill to return the major comminuty and continue drawing. Take the red pill if you're prepared to my observation with some legit ML / NLP / AI knowledge complete non-sense (or somewhat closest to the reality of bugs / expolits / Undefined behavior dystopia of the released AI models).

(260504) Capping of research topics

• I am not going to continue exploring Diffusion / AutoRegressive models based from DiT e.g. Anima or MoT e.g. SenseNova-U1 because of lacking of research support. Math properies, scientitic study and engineering application should be studied from scratch. Relying on designer tools / creator based community will not be sustainable. I will fade out with A1111 which he is a great inventor by blending many useful tools within math properties.

Contact

Seriously? I'm no different than a random anon in this field.

• Discord: 6DammK9#2533
• GMail: 6DammK9@gmail.com
• Pixiv: 6DammK9