AstolfoMix-XL (DGMLA-216 / TGMD-192 / TGMD / TSD / TIES-SOUP / Extended-FP64 / Baseline)

• Unsolved merge, even with experience on SD1 and SD2.
• See full article in Github
• Since "Extended-FP64", dedicated merger sd-mecha is used. It is the SD version of mergekit.
• Preview here are powered by the top session of README.md and converting SDXL model from A1111 standalone file into diffuers via convert_sdxl_to_diffusers.py. Settings may not being optimal (no CFG / PAG / FreeU etc). I'll replace the preview diffuser as soon as I upload the main model file.

DGMLA-216

• DGMLA Merge

• Recipe, E2E merger.

• Current version: x215a-AstolfoMix-24101101-6e545a3.safetensors

• Recommended version: "x215a"

• Recommended CFG: 4.5 (with CHG = 1.0), 3.0 (with PAG = 1.0)

• Prompt is minimal. Even empty.

parameters

[halloween], (astolfo:0.98), [[[[cemetery]]]]
Steps: 256, Sampler: Euler, Schedule type: Automatic, CFG scale: 3, Seed: 2123095857, Size: 1344x768, Model hash: bdb9f136b6, Model: x215a-AstolfoMix-24101101-6e545a3, VAE hash: 235745af8d, VAE: sdxl-vae-fp16-fix.vae.safetensors, Denoising strength: 0.7, 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, Hires upscale: 2, Hires steps: 64, Hires upscaler: Latent, 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 SANF: True, PAG Scale: 1, PAG Start Step: 0, PAG End Step: 150, Version: v1.10.1

TGMD-192

• Scaled up version from TGMD (from 117 to 192)

• Recipe, E2E merger.

• Current version: x191a-AstolfoMix-24083001-3360d18.safetensors

• Recommended version: "x191a"

• Recommended CFG: 4.5 (with CHG = 1.0), 3.0 (with PAG = 1.0)

• Prompt is minimal. Even empty.

parameters

(car:0), [[lamborghini]], (1boy:0), [astolfo]
Steps: 256, Sampler: Euler, Schedule type: Automatic, CFG scale: 3, Seed: 2435649982, Size: 1344x768, Model hash: 4c118beaa8, Model: x191a-AstolfoMix-24083001-3360d18, VAE hash: 235745af8d, VAE: sdxl-vae-fp16-fix.vae.safetensors, Denoising strength: 0.7, 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, Hires upscale: 1.5, Hires steps: 64, Hires upscaler: Latent, 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 SANF: True, PAG Scale: 1, PAG Start Step: 0, PAG End Step: 2048, Version: v1.10.1

TGMD (TIES-GeometricMedian w/ DROP)

• TGMD Merge of 116 SDXL models, unfiltered. TGMD is an algorithm modified from Model Stock.

• Recipe, E2E merger.

• Current version: x116a-AstolfoMix-24060702-01823a9.safetensors

• Recommended version: "x116a"

• Recommended CFG: 4.5 (with CHG = 1.0), 3.0 (with PAG = 1.0)

• Prompt is minimal. Even empty.

parameters

[[striped thighhighs]], [[midriff]], [[striped shirt]], [[hoodie]], [[braid]], [astolfo], [[[[eiffel tower, france]]]]
Steps: 256, Sampler: Euler, Schedule type: Automatic, CFG scale: 3, Seed: 1526207600, Size: 1344x768, Model hash: bc747cafd1, Model: x116a-AstolfoMix-24060702-01823a9, VAE hash: 235745af8d, VAE: sdxl-vae-fp16-fix.vae.safetensors, Denoising strength: 0.7, 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, Hires upscale: 2, Hires steps: 64, Hires upscaler: Latent, 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: v1.9.4

TSD (TIES-SOUP w/ DROP)

• TSD Merge of 102 SDXL models, unfiltered. TSD is an algorithm modified from DARE Merge (ICML 2024).

• Recipe, E2E merger.

• Current version: x101a-AstolfoMix-24050903-4edc67c.safetensors

• Recommended version: "x101a"

• Recommended CFG: 4.5 (with CHG = 1.0), 3.0 (with PAG = 1.0)

• Prompt is minimal. Even empty.

parameters

(hippogriff:0.98), [braid], [[cape]], [[astolfo]], [[[[greece]]]]
Steps: 256, Sampler: Euler, Schedule type: Automatic, CFG scale: 3, Seed: 3134096594, Size: 1344x768, Model hash: 7668681e22, Model: x101a-AstolfoMix-24050903-4edc67c, VAE hash: 235745af8d, VAE: sdxl-vae-fp16-fix.vae.safetensors, Denoising strength: 0.7, 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, Hires upscale: 2, Hires steps: 64, Hires upscaler: Latent, 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

TIES-SOUP

• TIES-SOUP of 73 SDXL models, unfiltered. TIES-SOUP is an algorithm modified from TIES merging (NeurIPS 2023).

• Recipe, E2E merger.

• Current version: x72a-AstolfoMix-240421-feefbf4.safetensors

• Recommended version: "x72a"

• Recommended CFG: 4.5 (with CHG = 1.0), 3.0 (with PAG = 1.0)

• Prompt is minimal. Even empty.

parameters

(car:0), [[mclaren]], [astolfo]
Steps: 256, Sampler: Euler, Schedule type: Automatic, CFG scale: 3, Seed: 1504757665, Size: 1344x768, Model hash: e276a52700, Model: x72a-AstolfoMix-240421-feefbf4, VAE hash: 26cc240b77, VAE: sd_xl_base_1.0.vae.safetensors, Denoising strength: 0.7, 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, Hires upscale: 1.5, Hires steps: 64, Hires upscaler: Latent, 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

Extended-FP64

• Uniform merge of 52 UNETS + (61+42) CLIPS (from 70 models discovered).

• Current version: x51-AstolfoMix-x60te0x41te1-e2e-240407-feefbf4.safetensors

• Recommended version: "x51"

• Recommended CFG: 4.5 (with CHG = 1.0), 3.0 (with PAG = 1.0)

• Prompt is minimal.

parameters

(car:0), [[mclaren]], (1boy:0), [astolfo]
Steps: 256, Sampler: Euler, Schedule type: Automatic, CFG scale: 4.5, Seed: 1841382272, Size: 1344x768, Model hash: a52eba463d, Model: x51-AstolfoMix-x60te0x41te1-e2e-240407-feefbf4, VAE hash: 26cc240b77, VAE: sd_xl_base_1.0.vae.safetensors, Denoising strength: 0.7, Clip skip: 2, CHG: "{'RegS': 1, 'RegR': 1, 'MaxI': 50, 'NBasis': 0, 'Reuse': 1, 'Tol': -4, 'IteSS': 1, 'ASpeed': 0.4, 'AStrength': 0.5, 'AADim': 2, 'CMode': 'More ControlNet', 'StartStep': 0, 'StopStep': 1}", 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, Hires upscale: 1.5, Hires steps: 64, Hires upscaler: Latent, Dynamic thresholding enabled: True, Mimic scale: 1, Separate Feature Channels: False, Scaling Startpoint: MEAN, Variability Measure: AD, Interpolate Phi: 0.3, Threshold percentile: 100, Refiner switch by sampling steps: True, Version: v1.9.0

Baseline

• Uniform merge of 32 UNETS + (19+26) CLIPS (from 21 models). Discovered model count: 42. It is a spinoff of Uniform Soup.

• Current version: x17-AstolfoMix-x13te0x14te1.safetensors

• Recommended version: "x17" for full experience, or "x11c" for human focused.

• Recommended CFG: 4.5

• Prompt is minimal.

parameters

(solo:0), (boy:0), (qipao:0.98), [astolfo_\(fate\)], [[lunar new year]], [[[[kowloon]]]]
Steps: 192, Sampler: Euler, CFG scale: 4.5, Seed: 2213673007, Size: 1344x768, Model hash: 82f53a8fe1, Model: x17-AstolfoMix-x13te0x14te1, VAE hash: 26cc240b77, VAE: sd_xl_base_1.0.vae.safetensors, Denoising strength: 0.7, 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, Hires upscale: 1.5, Hires upscaler: Latent, Dynamic thresholding enabled: True, Mimic scale: 1, Separate Feature Channels: False, Scaling Startpoint: MEAN, Variability Measure: AD, Interpolate Phi: 0.3, Threshold percentile: 100, Version: v1.7.0

Recipes / Model selection logs / Models involved

• See full article in Github

Round	Algo	Model Name	RAW	UNET	TE0	TE1	Recipe
01	Uniform Soup	x17-AstolfoMix-x13te0x14te1.safetensors	42	32	14	21	json
02	Uniform Soup	x43-AstolfoMix-x22te0x31te1.safetensors	50	44	22	31	mecha
03	Uniform Soup	x45-AstolfoMix-x39te0x39te1-e2e-240222-60d0764.safetensors	52	46	40	40	mecha
04	Uniform Soup	x63-AstolfoMix-x60te0x41te1-e2e-240407-feefbf4.safetensors	70	52	61	42	mecha
05	TIES-SOUP	x72a-AstolfoMix-240421-feefbf4.safetensors	73	73	73	73	mecha
06	TIES-SOUP w/ DROP	x101a-AstolfoMix-24050903-4edc67c.safetensors	102	102	102	102	mecha
07	TGMD: TIES-GeometricMedian w/ DROP	x116a-AstolfoMix-24060702-01823a9.safetensors	117	117	117	117	mecha
08	TGMD-192: Scaled up from TGMD	x191a-AstolfoMix-24083001-3360d18.safetensors	192	192	192	192	mecha
09	DGMLA-216: Drop w/ GeoMedian and LA	x215a-AstolfoMix-24101101-6e545a3.safetensors	216	216	216	216	mecha

DGMLA in nutshell

• This algorithm should be linear a.k.a O(N) in space, and O(NlogN) in time complexity. However layerwise merging introduce some constants skewing the actual experience.

Date	Algo	Model counts	Threads	RAM Usage (TB, FP64)	Time used (Hours, Xeon 8358 x2)
240607	TGMD	117	16	1.214	14
240622	TGMD	133	8	< 1.0	12.5
240830	TGMD	192	8	1.446	41.5
241002	DGMLA	192	16	1.452	39.1
241006	DGMLA	20	48	0.358	2.33
241011	DGMLA	216	48	3.500	36.2