AutismMix NSFW Guide: Versions, Settings, Prompts (2026)

AutismMix is a free Pony and SDXL-based anime checkpoint family (AutismMix SDXL, Confetti, and Pony) prized for predictable, artist-aware output with lean negatives. Use Pony score tags, CFG 5 to 7, 25 to 30 steps, Euler a, 832×1216. We tested all three variants on an RTX 4090 and a 3060.

AutismMix NSFW Guide: Versions, Settings, Prompts (2026)

AutismMix is a long-running anime checkpoint family that earned a loyal following for one reason: it makes Pony predictable. Where raw Pony Diffusion V6 demands giant negative prompts and a lot of trial and error, AutismMix tames that while keeping Pony’s deep artist knowledge and explicit NSFW understanding. If you want to preview the anime look before downloading, you can try our on-site generator as a no-install option first.

We ran every AutismMix variant through our standard prompt battery. This guide explains the differences between the versions, the settings that work, and how to prompt each one. If you are new to this family, read it alongside our anime checkpoint roundup and, since one branch is Pony-based, our Pony NSFW guide.

The AutismMix variants explained

AutismMix ships as a small family rather than a single file. All of it lives on the Civitai page. The three that matter:

• AutismMix_pony is the Pony-based variant. It keeps Pony Diffusion V6’s strong anatomy and artist comprehension but is tuned to be more stable and less dependent on schizo negative prompts. This is the explicit-content workhorse of the family.
• AutismMix_confetti is the headline release. The author’s goal with Confetti was to make Pony more predictable and less reliant on giant negatives while keeping comprehension and artist knowledge. In our testing it gives the cleanest default look of the three.
• AutismMix SDXL Lightning is a 4-step distilled variant for speed. It trades some quality for very fast generation. Use it for rapid iteration, not finals.

The key thing to understand: AutismMix is Pony-lineage, so it uses Pony’s score tag system, not the Illustrious quality tags used by WAI or NoobAI. Mixing those vocabularies up is the most common beginner mistake.

It is also worth knowing why this family has the name it does. The author built it as a personal project to make Pony generations more consistent and less random, the kind of reliability that lets you focus on the image instead of fighting the model. That design goal shows up everywhere in how it behaves: predictable composition, stable anatomy, and a strong response to clean prompts rather than chaotic walls of negatives. If you have used raw Pony V6 and found it temperamental, the difference on AutismMix is immediately obvious from the first generation.

Hardware and software

We benchmarked AutismMix on two GPUs.

On an RTX 4090 (24 GB) a standard Confetti image at 832×1216, 28 steps, takes about 4 to 6 seconds. The Lightning variant at 6 steps lands in roughly 1.5 to 2 seconds. On an RTX 3060 (12 GB) the standard variant takes about 18 to 25 seconds, while Lightning drops that to around 5 to 7 seconds. AutismMix is SDXL-class, so 8 GB is the floor and 12 GB is comfortable.

For software, any modern UI works:

• Automatic1111 for the standard interface.
• Forge for speed and memory efficiency.
• ComfyUI for node graphs, batching, and detailers.

No special prediction mode is required. If you would rather skip the install, the no-install generator is the quick path.

Best settings for the standard variants (Confetti and Pony)

These are the settings we landed on for AutismMix_confetti and AutismMix_pony after testing.

• Sampler: Euler a is the safe default. DPM++ 2M Karras gives sharper output.
• Steps: 25 to 30. We use 28 for finals.
• CFG scale: 5 to 7. Six is a good middle. Confetti tolerates the higher end better than raw Pony.
• Resolution: 832×1216 portrait, 1216×832 landscape, or 1024×1024.
• Clip skip: 2.
• VAE: the SDXL VAE; the baked-in version is fine, or load sdxl_vae.safetensors if colors look off.

For hires fix, upscale 1.5x with R-ESRGAN 4x+ Anime6B and denoise 0.35 to 0.45.

Best settings for the Lightning variant

Lightning is a different beast and needs different numbers, or it produces mush.

• Sampler: Euler a SGMUniform, or DPM++ 2M SDE SGMUniform. The SGMUniform scheduler matters here.
• Steps: 4 to 8, with 6 recommended.
• CFG scale: 1 to 3, with around 2 to 2.5 recommended. Pushing CFG higher on Lightning degrades quality unless you add steps.
• Resolution: 832×1216.
• Clip skip: 2.

Lightning is for fast drafts. We use it to find a composition in seconds, then re-render the locked seed on Confetti at full steps for the final.

Prompting AutismMix

Because AutismMix is Pony-based, you prompt with Pony score tags. The standard quality block is the score ladder. The author’s notes are clear that schizo negatives and the old SDXL/Pony negative embeddings actually make Confetti results worse, so keep negatives lean.

Here is a starting prompt we used in testing.

Positive:
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,
source_anime, BREAK,
1girl, solo, long hair, looking at viewer, detailed face,
nude, large breasts, lying on bed, indoors, soft lighting,
detailed background, depth of field

Negative:
score_6, score_5, score_4, worst quality, low quality, lowres,
bad anatomy, bad hands, missing fingers, extra digits,
watermark, signature, blurry, censored

Notes from our runs:

• Always lead with the full score ladder. On Pony-lineage models the score tags are load-bearing, not optional.
• Use “source_anime” to steer toward the anime aesthetic; “source_pony” or “source_furry” shift the look.
• The BREAK keyword separates the quality block from the content block, which helps the model weight them independently.
• Keep negatives lean. This is the whole point of Confetti. Do not paste a giant negative wall; it degrades output.
• Artist tags work strongly thanks to Pony’s deep artist comprehension. Use them ethically.

AutismMix versus the other anime NSFW checkpoints

Here is how AutismMix compares against the models we benchmark, on a single 832×1216 image at 28 steps on the RTX 4090.

Checkpoint	Base	Quality tags	Best sampler	Steps	CFG	4090 gen time
AutismMix Confetti	Pony	score_9, score_8_up…	Euler a	25-30	5-7	~5s
AutismMix Lightning	Pony	score_9, score_8_up…	Euler a SGMUniform	4-8	1-3	~2s
Prefect Pony XL	Pony	score_9, score_8_up…	Euler a	25-30	5-7	~5s
WAI-NSFW Illustrious	Illustrious	masterpiece, best quality…	Euler a	25-40	5-7	~5s
NoobAI-XL	Illustrious	masterpiece, best quality…	Euler a	25-30	5-6	~5s

The takeaway: AutismMix and Prefect Pony are the Pony-lineage options with score tags and deep artist knowledge, while WAI and NoobAI are the Illustrious-lineage options with cleaner default anatomy. AutismMix Confetti is the most predictable Pony you can run, which is why people who love Pony’s artist comprehension but hate its quirks gravitate to it.

A practical AutismMix workflow

This is the loop we use for finished images.

1. Draft fast on the Lightning variant at 6 steps to find a composition.
2. Lock the seed and re-render on Confetti at 28 steps, CFG 6, Euler a.
3. Run hires fix at 1.5x, R-ESRGAN Anime6B, denoise 0.4.
4. Run ADetailer on the face at inpaint denoise around 0.3, and on hands if needed.
5. Optional final upscale pass in img2img at denoise 0.2 for 4K.

The Lightning-then-Confetti hop gives you speed for exploration and quality for the final.

LoRA compatibility

AutismMix is Pony-based, so it works with Pony LoRAs, which are abundant. A few notes:

• Pony character and concept LoRAs load cleanly; start around weight 0.8.
• Illustrious and NoobAI LoRAs do not transfer to AutismMix, since the base is Pony, not Illustrious.
• Confetti’s tuned base means you often need fewer LoRAs to get a clean result.
• Keep stacked LoRA weight under about 1.5 to preserve coherence.

If you have a library of Pony LoRAs, AutismMix is a natural home for them.

Common problems and fixes

Output is generic or low quality. You probably dropped the score ladder. On Pony-lineage models the score_9, score_8_up chain is essential; without it quality collapses.

Image looks worse the more negatives you add. That is expected on Confetti. Remove the negative wall and the embeddings; lean negatives are by design.

Lightning output is mushy. You are using too many steps with too high CFG, or the wrong scheduler. Use SGMUniform, 6 steps, CFG around 2.

Hands and fingers are wrong. Universal SDXL issue. Use hires fix plus ADetailer with a hand model.

Censoring bars appear. Add “censored, mosaic censoring, bar censor” to the negative.

Colors look washed. Try loading sdxl_vae.safetensors as an external VAE and confirm Clip skip is 2.

Understanding the score tag ladder

The score ladder is the single most important concept for any Pony-lineage model, AutismMix included, so it is worth understanding rather than just copying. Pony Diffusion was trained with aesthetic score tags from score_1 up to score_9, where score_9 marks the highest-rated images in the training set. When you put the full ascending chain in your positive prompt, you are telling the model to draw from its highest-quality distribution.

The full chain, score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, is the conventional block because each “_up” tag pulls in everything at that level and above, reinforcing the quality signal. Dropping even part of this chain visibly lowers quality on AutismMix. Conversely, putting low score tags like score_4 or score_5 in the negative prompt pushes the model further away from mediocre output, which is a clean way to lift quality without bloating the positive.

A few additional source tags shape the overall aesthetic:

• source_anime steers toward the anime illustration look, which is what most readers here want.
• source_pony shifts toward the show-inspired style the model is named for.
• source_furry and source_cartoon move the look accordingly.
• source_photo nudges toward semi-realism, though Illustrious-lineage models are better at photoreal.

Getting the score ladder and a source tag right is 80 percent of prompting AutismMix well.

Tag weighting and emphasis

After the score ladder, attention weighting is your main steering tool. In Automatic1111 and Forge syntax, wrap a tag in parentheses with a number, like (detailed face:1.2) to strengthen or (background:0.8) to weaken. AutismMix responds well, though Pony-lineage models can be slightly more sensitive than Illustrious ones, so stay conservative.

Patterns we rely on:

• Strengthen the focal subject modestly, for example (long hair:1.1).
• Weaken an element that dominates, such as (depth of field:0.8).
• Lock a pose or act with a small bump rather than a large one; over-weighting an explicit tag distorts anatomy fast on Pony models.
• Keep individual weights under about 1.25 on AutismMix; it warps sooner than the Illustrious models.

Inpainting and fixing AutismMix output

Running AutismMix locally lets you repair one region rather than reroll the whole image. The repair loop:

1. Send the image to the inpaint tab.
2. Mask only the broken area.
3. Set inpaint area to “only masked” for full resolution on that patch.
4. Use denoise 0.4 to 0.6, and keep the score ladder in the inpaint prompt, since dropping it lowers quality even on a small patch.
5. Trim the rest of the prompt to just the masked subject.

Keeping the score tags in the inpaint prompt is a Pony-specific detail people forget, and it is why inpainted patches sometimes come back lower quality than the surrounding image.

Running AutismMix on lower-end hardware

Below 12 GB of VRAM, AutismMix still runs with care. On 8 GB cards, use Forge for its better SDXL memory handling, generate at 832×1216, keep batch size at 1, and run hires fix as a separate img2img pass. The Lightning variant is a genuine gift here: at 6 steps it slashes generation time, so even an 8 GB card can iterate quickly, then you finish the chosen seed on Confetti.

Below 8 GB, local SDXL gets painful. That is where the on-site generator shines as the zero-setup option, running the heavy compute remotely so an older card or laptop can still produce anime output. We treat it as the no-install path and local AutismMix as the power-user path.

Settings recap and final verdict

In one place: for Confetti and the Pony variant, set Euler a, 28 steps, CFG 5 to 7 with 6 as default, 832×1216, Clip skip 2. For Lightning, switch to a SGMUniform scheduler, 6 steps, CFG around 2. Always lead with the full score ladder plus source_anime and BREAK, keep negatives lean, and finish with hires fix at 1.5x, R-ESRGAN Anime6B, denoise 0.4, then ADetailer on the face.

Our verdict after testing all three variants: AutismMix Confetti is the most predictable Pony-lineage anime checkpoint you can run in 2026. If you love Pony’s artist comprehension and strong anatomy but hate its chaos and giant negative prompts, this is the model that fixes that. The Lightning variant turns it into a fast iteration tool, and the standard Confetti finish quality is excellent.

The alternatives each have a niche: Prefect Pony for a different polished Pony look, WAI and NoobAI for the Illustrious lineage with cleaner default anatomy, and Hassaku for step efficiency. But for the most predictable, low-fuss Pony experience, AutismMix Confetti is the one we recommend.

AutismMix is the answer for people who want Pony’s enormous artist comprehension without Pony’s chaos. Learn the score ladder, keep negatives lean, draft on Lightning, finish on Confetti, and you have a fast, predictable anime NSFW pipeline. Grab it from the Civitai page, and when you just want one quick image, the on-site generator is always available.

Frequently asked questions

What is the difference between AutismMix Confetti, Pony, and Lightning?

AutismMix_pony is the Pony-based variant with strong anatomy and artist knowledge, tuned to need fewer schizo negatives. AutismMix_confetti is the refined headline release that makes Pony the most predictable, with the cleanest default look. AutismMix Lightning is a 4-step distilled variant for speed, best for rapid drafts rather than finals. All three are Pony-lineage and use the same score tag system, so prompts carry over between them.

What quality tags does AutismMix use?

Because AutismMix is Pony-based, it uses Pony score tags, not Illustrious quality tags. Lead your positive prompt with score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, then source_anime and BREAK. The score ladder is load-bearing on Pony-lineage models, so quality collapses without it. This is different from WAI and NoobAI, which use masterpiece, best quality. Mixing the two vocabularies is the most common beginner mistake.

Why do my AutismMix results get worse when I add more negatives?

That is by design on the Confetti variant. The author tuned Confetti specifically so it does not need the giant schizo negative prompts and old SDXL or Pony negative embeddings that raw Pony relied on. Those embeddings actively degrade Confetti output. Keep your negative prompt lean, with a few score tags, worst quality, bad anatomy, and bad hands. Removing the negative wall almost always improves results on this model.

What settings should I use for AutismMix Lightning?

Lightning needs very different numbers from the standard variants. Use a SGMUniform scheduler such as Euler a SGMUniform or DPM++ 2M SDE SGMUniform, 4 to 8 steps with 6 recommended, and a low CFG of 1 to 3 with around 2 to 2.5 recommended. Higher CFG on Lightning degrades quality unless you add more steps. Use Lightning for fast drafts, then re-render the locked seed on Confetti at full steps for the final.

What are the best settings for AutismMix Confetti?

Use Euler a as the sampler, 25 to 30 steps with 28 for finals, CFG 5 to 7 with 6 as a middle value, and a resolution of 832×1216. Clip skip should be 2. The baked-in SDXL VAE is fine, though you can load sdxl_vae.safetensors if colors look off. For finishing, add hires fix at 1.5x with R-ESRGAN 4x+ Anime6B and a denoise of 0.35 to 0.45.

Do Illustrious or NoobAI LoRAs work with AutismMix?

No. AutismMix is built on the Pony base, so it works with Pony LoRAs, not Illustrious or NoobAI LoRAs. Those use a different conditioning scheme and will not bind correctly. If you have a library of Pony character or concept LoRAs, AutismMix is a natural home for them; start around weight 0.8. For Illustrious LoRAs, use an Illustrious checkpoint like WAI or NoobAI instead.

Can AutismMix run on an RTX 3060 12GB?

Yes. On our RTX 3060 12GB, a standard Confetti image at 832×1216 and 28 steps took about 18 to 25 seconds, and the Lightning variant at 6 steps dropped that to around 5 to 7 seconds. The 12 GB of VRAM is comfortable for SDXL-class models including hires fix. Cards with 8 GB work with longer waits. For quick images without any GPU, the on-site generator is the easier path.

Is AutismMix better than Prefect Pony XL?

Both are excellent Pony-lineage anime checkpoints and use the same score tags, so it comes down to aesthetic preference. AutismMix Confetti is famous for predictability and lean negatives, while Prefect Pony XL has its own polished anime look and very low recommended CFG on merged versions. We keep both installed and pick by project. If you value the most stable, low-fuss Pony experience, AutismMix Confetti is hard to beat.