XenEro | Pony | NSFW

Version notes

v1.8 (v1.7 + Male-Female fix + SatPony IOR composition mix)

Many thanks to @Sateluco and SatPony IOR checkpoint. I used it as a second layer of composition / characters fix.

Merge Recipe: v1.8 = [Male_Female_Nature dataset] + [MoodySky] + [AutumnVibes] + [RiverSide] + [SatPony IOR]

Male-Female-Nature LoRa (unreleased) was specially trained on deliberately crafted and captioned dataset (Male - Female - Cybernetic / Mechanic parts - Nature). My recent LoRas (MoodySky, Autumn, Riverside) was also mixed in with small weight.

SatPony IOR merged in MBW mode with careful blocks tuning.

• Male / Female adherence fix using Training.

• Hands / fingers. Additional minor fixes.

• Much improved nature (MoodySky + Autumn + RiverSide LoRas)

• Major adjustment of CFG sensitivity. From 3-4 to 7+

• Major improvement in smallest details rendering (hairs, micro elements of constructions, etc)

• Colors adjustment to increase realism (saturation reduced)

• Same bokeh + slightly better cinematic style

• Slightly more diverse faces

v1.7 1st try to fix v1.6

The best version to switch to v1.8 ;)

v1.6 (Refiner + CRealism = CRealism2)

Fixes for v1.5: too contrast unnatural liquids, flawy fingers, high bias into naked feminine body, bias to simple backgrounds.

Based on v1.5 and v1.4. Added few SDXL / Pony resources (CLIP + Unet IN** blocks) to fix Prompt Adherence. This version should show better adherence (v1.4 Refiner fix) with good creativity (v1.4 CRealism replication).

This version also has a tendency to closer portraits (if you give high weight to "eye" related portions of the prompt).

v1.5 (Refiner)

IMPORTANT: read details about Samplers and other parameters.

Fixes for hands, anatomy, details, more diverse faces, less contrast palette for better realism.

This version aimed to be a better 2nd pass model in comparison with v1.4. I added few datasets with turn table 3D scan of artistic females and male models (fantasy theme; mage, battle stances with and without weapons). To bring back the creativity, several SDXL / Pony resources was added via MBW merges (careful precise tuning). Most influential resources was CinEro and XenEro models, my own unreleased LoRas. Ability to render cosmos, stars, magic light might be weaker than creativity of XenEro v1.4.

This version is created to be used in pair with Flux models to render better skin and other fine details textures.

About samplers and params

FOR Image-to-Image

General recommendation is to use Image-to-Image with v1.5. Use VERY LOW denoise (like 0.1..0.16 with Beta scheduler or 0.25..0.35 with Karras or SGM Uniform) combined with VERY HIGH CFG (like 3.5 for Beta scheduler or 12..15 with other schedulers). DPM++ samplers shows higher CFG sensitivity. I recommend to lower the CFG with DPM++ SDE and DPM++ 2M samplers.

Euler a with Karras or SGM Unifirm gives smoother results. This sampler more safe in sense of realism and contrast.

For Text-To-Image
1. DPM++ 2Sa => CFG 3.5 .. 4.5 + Karras or SGM Uniform

2. Euler A => CFG 4 and more + Karras or SGM Uniform

3. With Beta scheduler: the lower CFG the higher realism and smoother picture. If you can edit the inputs in A1111 WebUI via Chrome's F12 Developer Tools, then you can try using the CFG like 0.5 or lower. I suspect Beta will give good results this way.

v1.4 (CRealism)

IMPORTANT: read details about Samplers and other parameters.

This version is based on v1.3 with addition of many different LoRas (my own and others).

Improvements: significantly improved realism (skin and non organic textures), creativity (liquids, light particles, body and background details, expressions), render quality (eyes, hands, intimate parts, anatomy).

About samplers and params

For better realism, sensitivity to CFG was significantly increased. Euler A (smooth by its nature) works quite well with Karras and SGM Uniform schedulers, BUT...
DPM++ 2M sampler with Karras or SGM Uniform schedulers AND most Samplers with Beta scheduler shows much higher CFG sensitivity. For DPM ++ 2M (and probably Heun) sampler OR with Beta scheduler you probably should use 2-time lower CFG as for Euler A.

Additionally, Beta scheduler shows better realism with lower CFGs. I tested Beta scheduler with CFG 2.0..3.0.

DPM++ SDE based samplers should work good up to CFG 4.5.

DPM++ 2S a didn't test at the moment of release, but it should give a better realism with Karras at CFG 3.5..4.5.

CFG higher than 4.5 to be used with extra small samplers. With higher CFG final picture with have a simpler shapes and less small details with high contrast.

I also strongly recommend to use ADetailer because this version (and v1.3 as well) have a ability to produce very good portraits and crispy textures at close shots.

Pleasurable Neuro Drawing to every one! Thanks for support!

v1.3 (Realism)

This version's goal is to achieve good realism, improve skin texture, improve structural consistency and anatomy correctness. Secondary goal: add more creativity but keep DucHaiten's style triggers alive as much as possible (probably they are weak now). This version should be able to draw in illustration style despite the overall focus on realism.

Dynamic range of CFG and prompt weights has being improved (feel free to report how it works for high CFGs).

IMPORTANT NOTE:

FP8 (3.6Gb) model is not working in vanilla A1111 WebUI because it doesn't know how to decode FP8 e4m3fn. Please test this version in ComfyUI or another SDXL UI which can be started in FP8 mode. Maybe A1111 WebUI also have this mode via command line arguments... I don't know if FP8 can be used with WebUI.

v1.2 (DucHaiten)

This version made using a precision multi-stage merge of CinEro v1.3 Ephoch 40 (not released) + XenEro v1.1 + DucHaiten Pony (by @DucHaiten ).

Recipe: MBW merge mode with careful, precise balancing of CLIP and UNet related block weights.

Goal: keep style triggers from DucHaiten Pony v7.0, add more realism from CinEro and add more adherence from XenEro v1.1; ultimate goal is to create a model capable to produce flat 2D artist styled images and realistic/semi-realistic images with hints of artist style.

--< STYLES from DucHaiten Pony v7.0 >--

--< Semi-Realism and Creativity from CinEro v1.3 Epoch 40 and XenEro v1.1 >--

--< Fixes stability issues detected on many PONY models >--

(find "DH7.0-CinEro1.3" in comparison grids in this post for DucHaiten Pony v7.0)

Stability issues: when positive prompt contains a small weight (<0.8) and high weights (>1.2) many Pony models like CinEro v1.3, Crow & Pony, DucHaiten Puny v7.0 may produce overburned or exploded image. This version of XenEro adresses it (chance to get exploded image very low) according to comparison grids.

PS: all BUZZ earned will be split 50/50 with DucHaiten after Early Access end.

v1.1

Improved:

• cyan tint removed (chance to get whole picture shifted to cyans dramatically reduced);

• CFG sensitivity corrected back to 5-9 range (this version works with usual CFGs like 5..7..10);

• zoom issue rate reduced (chance to get extreme zoom lowered);

• adherence improved (see comparison grids);

• anatomy issues rate reduced;

• big region texture made smoother (for this version CLIP blocks has been adjusted to make it produce smoother textures);

• other minor fixes and improvements.

v1.0

Initial merge.

Has a tendency to produce smoother picture at CFGs lower that 14. This version is tested with at least CFG 20. It does not burn out at CFG 17. So, CFG range is much higher that it usually is on the Pony Models.

It was carefully tuned to give smoother texture for big objects and background while producing the crispy micro-details in small objects and small objects parts.

Bad things: slightly higher chance to get bad or misaligned hands relatively to CinEro v1.3 E18.

Some details

This model is a result of a request to merge CinEro Pony with XenoDiffusion (SDXL).

XenoDiffusion license does not allow to share the merges with this model, so intermediate merges with this model was rolled back. I decided to replace it with similarly styled SDXL models.

SDXL models with good adherence was merged into CinEro Pony v1.3 (Epochs 18, 20, 21) in block-wise mode. Block weights was configured to transfer mostly CLIP weights of LOWRES and HIGHRES stages. Weight setup was similar to the following.

Initial merges done in Tripple Sum and Add Difference (Extract) modes.

Model B was one of the selected SDXL and Pony models with good adherence.

Model C was my CinEro Pony trained models.

During merge iterations I noticed that merge result drifts away from Pony toward SDXL. At some iteration merge of SDXL B with Pony C produced damaged model. So, last 2-3 iterations of the merges was done with Pony as a B model and CinEro v1.3 E23 trained Pony.

As a result I've got a merge with better adherence closer to SDXL models and inherited creative features of CinEro v1.3.