<h1 id="actually-played-with-it-and-found-7ouikfd9k">Actually played with it and found</h1>This model actually depicts on a 5x5 grid, and the attention is split on the 5x5 lines.Behold, the rule of 3 is broken with this model. Long live the rule of 5.This is the first Flux model I've seen with bottom-line 5x5 split attention control.<h1 id="update-v1512-is-released:-2152025-h2y44wjgl">Update - V1512 is released: 2/15/2025</h1>I'm doing an early release due to needing at least a few days to develop the captioning software to create the next version iteration; and I simply do not have the time of day.I'm releasing mostly because of the discovery with Hunyuan, and I'm also releasing a working Hunyuan merge with Simulacrum Schnell that produces very fun stuff... and it really shouldn't.This version's core features;<ul><li>does not require negative prompting</li><li>works with Flux1S models</li><li>works with Hunyuan -&gt; throws some errors just ignore them</li><li>works with the various clip_l_omega versions</li><li>unfinished, but far more robust</li><li>tons of nsfw depiction and control</li><li>absolute ton of spatial understand and entirely new methods of reasoning for the model to understand it</li><li>up to 20 identifiable characters simultaneously per image, this was not done yet so the outcomes are very hit or miss at times.</li><li>Grid and screen depiction control baked into the core.</li><li>Full comics on request, also wasn't done yet.</li><li>Tremendous amount of outfits, screen control, truncated sectioning, grid control, rotation control, offset control, size control, and even more; hit or miss due to not being completely trained yet.</li></ul>Configuration:Load using the CLIP_24_L_OMEGA for full effect, as this version of Simulacrum Schnell was heavily trained using the post SDXL CLIP_L version.I'm uncertain if this CLIP_L is the exact same version, but it's definitely pretty similar if it isn't and should work with other versions of CLIP_L omega.<h1 id="for-flux-1s-4irleomx7">For Flux 1S</h1>Positive prompt is plain English peppered with booru tags, and then use booru tags as solidifiers later on.Prompt using a mix of styles, offsets, whatever. Just talk to the thing, it'll probably understand what you want. I gave it additional intelligence for many sectors in the human world.<h1 id="whatever-you-do-don't-put-anything-into-the-prompt-that-you-don't-want-to-see.-it-is-much-more-intelligent-and-more-literal-than-it's-flux-counterpart.-c380dxjtx">Whatever you do, don't put anything into the prompt that you don't want to see. It is much more intelligent and more literal than it's Flux counterpart.</h1>This is NOT Flux how you know it. It's far more unchained and far more flexible.There's a reason why these things take so long to come out, and you'll see it here if you play with it too much.This thing IS DEFINITELY NOT a safe model, nor is it meant for all ages. It was never given a full controlled polish finetune for v2, so it will produce what you want, and you WILL get the monkey paw with it.This is NOT Schnell. This is finetuned so heavily that it barely reflects the original at times, and other times it's basically base SchnellBe VERY WARY. You WILL see monsters.<pre><code>Positive Prompt TLDR:

&lt;setting, placement, location, situation&gt;
&lt;context caption&gt;

&lt;quality&gt;&lt;styles...&gt;
&lt;subject counts&gt;

&lt;action caption&gt;

add outfits, clothing, interactions, whatever here

masterpiece, most aesthetic, very aesthetic,

&lt;superimposed captions&gt;

&lt;t5 captions for UI and overlays&gt;




Mix and match these for negative;

Negative Prompt TLDR:

sex, nsfw, explicit, questionable, safe, 
anime, 3d, realistic, 
line drawing, digital artwork, 
interpolated frame, blurry, grid_, depicted-, size_, behind, side, front, 
bad anatomy, bad hands, mutated, extra limbs, missing limbs, amputee, quadruple amputee, blood, gore, guro, 
humanoid, anthro, furry, censored, uncensored, 
lowres, good aesthetic, very displeasing, disgusting, </code></pre><pre><code>Steps: 12-58 -&gt; 32
CFG: 2.5-9 -&gt; 3.5
DCFG: 0 (flux guidance)

Samplers:
Euler -&gt; Simple
DPM-2M -&gt; Beta/Simple
DPM-2S -&gt; Beta/Simple
DEIS -&gt; SGM Uniform &lt;&lt;&lt; Just found it, so good.

Resolutions: 
1308x1308 (really close to this can't remember), very big.
1216x1216, 1216x832, 832x1216
1024x1024
1024x832, 832x1024, 832x832
768x768
512x512, 512x768, 768x512
</code></pre><h1 id="for-hunyuan-jfdopc64a">For Hunyuan</h1>It's not the easiest setup yet.This wasn't TRAINED for Hunyuan, but the next version WILL BE trained for Hunyuan. Direct Hunyuan interpolation control.However, this thing works for some reason and I still haven't figured out why.<pre><code>Steps: 12-64 -&gt;
12 works okay, but it tends to get blurry or pixelated
24 works probably the best but takes a while with many frames

Frames:
65-200
I did most of my testing around 130 or so, so don't think this is a strict rule.

CFG:
3.5 - 9
They all produce interesting outcome. 9 is actually really good sometimes.
My go-to is 6.4


LORA STRENGTH:
You will likely need two or three lora loader nodes to make it work, but it does work.
- Single Blocks 0.80
- Double Blocks 0.20
- CLIP - 1.0</code></pre>Consult the SDXL-Tag Guide for a full list of trained bbox data.Interpolative video training data used for 3d and realistic.<h1 id="update-v2-is-still-brewing-:-262025-qmcvc2nyq">Update - V2 is still brewing : 2/6/2025</h1>I've turned up the NSFW knob and broke it off. If this thing doesn't produce high grade high complexity NSFW I'll be shocked.It'll probably take another week of cooking to fully reach maturation, so bare with the time.<h1 id="update-v2-is-brewing-:1302025-sqkw9sd23">Update - V2 is brewing :1/30/2025</h1>V2 will not need a negative prompt. I've been running the same data that I ran through SDXL for an epoch to see what it'll do. It's already starting to take, and the need for a negative prompt is going away rapidly.It's about 300,000 or so images, roughly a third with plain English prompts.So either the 5x5 grid will work, or the model will burn to a crisp.It's attached to CLIP_L OMEGA V4, so be aware it'll behave a little... a little differently than you expect compared to the first version. This CLIP_L is 10 million samples smarter.<hr />I made everything so you can download it while you aren't logged in, so no login validation or keys required to auto-download it.As of V129 DEPICTION OFFSET works in a substantial way. Experimentation required.I can officially declare V122 ADVANCED PROMPT GRADE NSFW. You should be able to create the majority of common NSFW related acts and detailed situations with the current released version. I fed 1D these exact images in MULTIPLE epochs and it laughed at me.The further along the training goes, the more plain English depiction it'll be able to handle. Currently this is not a simple process, but you can work it out if you give it a bit of elbow grease.I'll be writing three articles soon based on using this model, because it's quite different than Flux1D and it's VERY VERY underestimated. The power here is substantial and responsive to training, while Flux1D often fell apart during training.<ul><li>Simple Schnell subject fixation using the rule of 3</li><li>Complex scene interactions and careful caption planning for Schnell NSFW training</li><li>Prompting NSFW interactions and adult depictions with Simulacrum Schnell V1</li></ul>I guarantee this model is far more powerful than is expected of it, and the outcomes from training are far more powerful than expected. The QUALITY is suffering a bit currently, but the additional training is showing certain traits are most definitely clearing up over time. This is only going to compound when I provide it with more training and more information for the requests.I STRONGLY advise using "Shuttle 3 Diffusion" Schnell to inference this lora. It amplifies the capabilities a large amount with less prompting. Shuttle v3.1 is okay but doesn't work as well with this lora, it's more compliant with it's own thing.Standard Flux Schnell FP16 and FP8 depict a FAIR QUALITY with the same settings now that we've hit Epoch 5/10. Many details that Shuttle 3 Diffusion is hiding or replacing with it's own training appeared from the training as emergent traits in standard Schnell, while Shuttle is still hiding the effects. FP8 is a little lesser but not by much. I actually ran the first epoch on FP8 at a higher learn rate for the baeline, so it should respond pretty well with FP8. The additional 4 were on BF16 mixed training however, which makes them substantially more powerful with the BF16 and FP16 versions of Flux Schnell. I haven't tried the BF16 version yet, but I assume it's good.Schnell FP16 requires a bit of a balancing act with prompts to make the dataset pop out, but it's not too bad. You can usually generate some fair quality stuff with a few tries and some prompt tinkering.<h3 id="be-sure-to-use-the-simv4-clip_l-no-matter-which-model-you-use-as-it's-required-for-a-proper-experience.-peut0weo3">Be sure to use the SimV4 CLIP_L no matter which model you use, as it's required for a proper experience.</h3><h3 id="you-must-use-negative-prompt-for-the-full-experience.-7vd6nzhyy">You MUST use NEGATIVE PROMPT for the full experience.</h3><pre><code>Euler -&gt; Simple
DPM2M -&gt; Simple 

Steps 28
CFG 3.5</code></pre><pre><code>V122 Epoch 5 - Generation Settings:

The model .safetensors says e4 but I mislabeled it. It's definitely e5. 

Inference:
1024x1024, 1216x832, 832x1216
1216x1216, 
1024x768, 768x1024,
768x512, 512x768, 
768x832, 832x768

rule34.xxx and rule34.us tags for 3d.
danbooru/gelbooru tags for anime.
plain English for realistic.

Nothing special REQUIRED for positive prompt, but these do help.

Positive Prompt:
anime, realistic, real, 3d \(artwork\), 3d,

&lt;CAPTION HERE&gt;

very aesthetic, aesthetic, masterpiece

#########################################
### BASE SCHNELL FP16 Negative Prompt ###
#########################################
censored, censor, bar censor, blur censor,
lowres, bad quality, low quality, bad anatomy, 
blur, depth of field, distorted, pixelated, 
bad hands, blurry hands, extra digits, missing digits, missing hands, extra hands, unexplained hands, merging, 
penis, erection, sex toy, dildo, pussy, cameltoe, 
multi penis, deformed, mutated, monster, vore,
disembodied, floating object, 
disembodied hand, disembodied foot, disembodied head,
extra feet, unexplained feet, unexplained arm, 3 legs, missing leg, missing arm, 
simple background, blurry background, cave,


####################################
### SHUTTLE BF16 Negative Prompt ###
####################################
nsfw, explicit, 
censored, censor, bar censor, blur censor,
lowres, bad quality, low quality, 
blur, depth of field, distorted, pixelated, 
monochrome, greyscale, comic, 2koma, doujin, manga,
bad hands, blurry hands, extra digits, missing digits, missing hands, extra hands, unexplained hands,
penis, erection, flaccid, pussy, cameltoe,
multi penis, deformed, mutated, monster, vore, pregnant,
cum, ejaculation, messy, unexplained white liquid,
disembodied, floating object, disembodied penis, disembodied hand, disembodied foot, disembodied head, jumping, floating, extra feet, unexplained feet, unexplained penis, unexplained arm,
simple background, blurry background, cave, </code></pre><hr />The home for the Simulacrum Schnell model zoo.The article with more detailed information about the training and process can be found <a target="_blank" rel="ugc" href="https://civarchive.com/articles/9954/schnell-v1-the-first-50k-greater-training-the-50mil-sample-flux-simulacrum-v5">here.</a>The Simulacrum Schnell versions require the Simulacrum V4 CLIP_L to function properly.The tagging template is the same as <a target="_blank" rel="ugc" href="https://civarchive.com/models/950382?modelVersionId=1064359">Simulacrum V4.</a><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/093e2fc4-d08d-4df7-89fe-ec15847028fe/width=525/093e2fc4-d08d-4df7-89fe-ec15847028fe.jpeg" />Simulacrum Schnell is protected under a slightly modified <a target="_blank" rel="ugc" href="https://www.apache.org/licenses/LICENSE-2.0.html">Apache Open Source 2.0</a> license.<pre><code>Copyright 2025 Abstract Powered

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this model or model zoo compliant component except in compliance with the License.
You may obtain a copy of the License at

 http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

model zoo compliant component
Any code, component, image, derived image, Schnell based ai model released unto, Schnell based AI model released as derived and trained by Abstract Powered directly posted and hosted on Huggingface, Civit, or any other legal hosting service herein.

I hereby grant this direct exception to this license:
I grant the individual, small business, influencer, researcher, research facility, research groups, and small non-corporate entity direct and free of use to inference, train, replicate, modify, alter, or derive their own personal works based on all Schnell Simulacrum versions indefinitely without monetary contribution. You are free to use this model within the constraints of applicable law within your country of residence.

Special Exceptions:
Huggingface and Civit are both exempt from this rule and can profit monetarily without contribution.

Compliance:
Corporate entities, derived corporate entities, subset business entities, and for-profit research groups, or any similar group that fits the for-profit model, are to contact me directly for commercial and monetary use unless they are exempt via the exception rules.

By downloading the Simulacrum Schnell or any of it's derivatives trained and uploaded for distribution and sharing directly by Abstract Powered, you hereby accept this license.</code></pre>I'm not a lawyer. Just know that my intent is for the individual, small business, and influencer to monetarily gain from this model.Have fun everyone. I'll be posting many models.

v1512-E12 - Simulacrum Schnell Model Zoo

First attempt:THIS VERSION DOES NOT WORK VERY WELL, but it does show promise.REQUIRES: Sim_v4 CLIP_L<pre><code>Inference:
1024x1024, 1216x832, 832x1216

Positive Prompt:
realistic, real, 3d \(artwork\), 3d,

&lt;CAPTION HERE&gt;

very aesthetic, aesthetic, masterpiece

Negative Prompt:
nsfw, explicit, disgusting, bad quality, bad anatomy, mutation, mutated, disembodied, deformed, monochrome, greyscale, patreon logo, watermark, bad hands, extra digits, anthro, fur, furry, feral, humanoid

Euler -&gt; Simple

Steps 16
CFG 5

Steps 24
CFG 4

Steps 32
CFG 3

Steps 50
CFG 2</code></pre><pre><code>BF16 mixed lora training

FP16 Schnell Model
FP16 t5xxl
FP16 Sim-v4-CLIP_L

UNET_LR 0.0001
TE_LR 0.000001
Cosine with 2 restarts &lt;- my go-to for flux

AdamW8Bit
Dim 64 &lt;- high number means high potential to learn derived patterns
Alpha 64 &lt;- high number means better pattern retention

Flux Guidance 0 
Flux Shift
Euler Flow 1
Prediction Type Raw
Use t5 attention mask

255 max tokens for t5

Noise Offset 0.05
Offset Type Original</code></pre>

Files with SHA256: cf143f343b6e19e5316084a8ec6a091839d3a47e4f388c64a62db0ab5ba5ffda