Hassaku XL NSFW Guide: Best Settings for Anime (2026)

Hassaku XL (Illustrious) is a free, clean anime SDXL checkpoint that produces great NSFW output with few steps and minimal tuning. Recommended settings: Euler a, 20 to 28 steps, CFG 3 to 7 (6 is ideal), 832×1216, Clip skip 2, VAE inherited from Illustrious. We tested v3.4 on an RTX 4090 and a 3060.

Hassaku XL NSFW Guide: Best Settings for Anime (2026)

Hassaku XL is one of the most efficient anime NSFW checkpoints you can run in 2026. The Illustrious-based releases are known for a clean, vibrant anime aesthetic that looks finished with surprisingly few steps, which makes Hassaku a favorite for people on mid-range GPUs. If you want to see the style before committing a download, try our on-site generator first as a no-install option.

We put Hassaku XL through our standard prompt battery, most recently the v3.x line. This guide covers the settings, the prompt structure, and the quirks. If this is your first checkpoint in the family, read it next to our Illustrious base guide and our anime checkpoint roundup.

What Hassaku XL is

Hassaku XL (Illustrious) is a fine-tuned merge built on the Illustrious-XL base, with additional training on newer characters. The latest releases, in the v3 line, merge earlier Hassaku work back into the Illustrious base and add fresh character data. The model inherits its VAE directly from Illustrious-v1.1, which is why its color palette, saturation, and overall tone feel consistent with the Illustrious family.

The practical signature of Hassaku is efficiency. In testing it produces a clean, usable anime image at low step counts, where some checkpoints still look rough. It is fully uncensored, so explicit content needs no special unlock.

The official model lives on its Civitai page. Grab the latest version listed there for the best character coverage.

Hardware and software

We benchmarked Hassaku on two GPUs.

On an RTX 4090 (24 GB) an 832×1216 image at 24 steps renders in about 4 seconds. Because Hassaku tolerates low steps, you can push iteration even faster. On an RTX 3060 (12 GB) the same image takes roughly 15 to 22 seconds, and Hassaku’s low-step tolerance is a genuine advantage here: dropping to 20 steps keeps quality high while shaving time. As an SDXL model, 8 GB is the floor and 12 GB is comfortable.

For software, any modern UI runs Hassaku with no special configuration:

• Automatic1111 for the classic interface.
• Forge for better speed and memory handling.
• ComfyUI for node-based workflows, batching, and detailers.

No v-prediction, no Zero Terminal SNR. If you would rather not install anything, the no-install generator handles quick images.

Recommended settings for Hassaku XL

These are the settings we use after testing, aligned with the author’s recommendations.

• Sampler: Euler a. This is the recommended sampler and gives Hassaku its clean look. DPM++ 2M Karras works for sharper output.
• Steps: 20 to 28. Hassaku is unusually efficient; it can produce a decent image in as few as 11 steps, but 20 to 28 is the sweet spot for finals. We default to 24.
• CFG scale: 3 to 7, with 6 ideal when paired with Euler a.
• Resolution: 832×1216 is the most recommended portrait bucket. 1216×832 for landscape, 1024×1024 for square.
• Clip skip: 2.
• VAE: inherited from Illustrious-v1.1 and baked in; no external VAE needed.

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

Prompting Hassaku XL

Hassaku is a booru-tag, Illustrious-lineage model, so prompt in tags using the Illustrious quality vocabulary, not Pony score tags. The author gives a specific tag order that matters: number of characters first, then character tags, then all remaining tags. Example structure: “1girl, rem (re:zero), standing, masterpiece, upper body.” The author also advises avoiding Danbooru metadata tags and franchise tags, and adding “signature” to negatives if logos or signatures leak in.

Here is a starting prompt we used in testing.

Positive:
1girl, solo, long hair, detailed eyes, looking at viewer,
masterpiece, best quality, amazing quality, very aesthetic, absurdres,
nude, medium breasts, on bed, indoors, window light,
depth of field, detailed background

Negative:
worst quality, low quality, lowres, bad anatomy, bad hands,
missing fingers, extra digits, signature, watermark,
text, jpeg artifacts, blurry, censored

Notes from our runs:

• Lead with character count and character tags, as the author recommends, before the quality block. Hassaku binds character tags more reliably this way.
• Avoid Danbooru metadata tags and franchise tags; they muddy results on this model.
• Add “signature” to the negative if you see stray logos or signatures, a known leak from the dataset.
• Keep negatives lean. Like the rest of the Illustrious family, Hassaku does not need a bloated negative wall.
• Use “censored” in the negative to suppress mosaic or bar censoring.

Hassaku XL versus the other anime NSFW checkpoints

Here is how Hassaku compares against the models we benchmark, on a single 832×1216 image at the model’s recommended steps on the RTX 4090.

Checkpoint	Base	Best sampler	Steps	CFG	Step efficiency	4090 gen time
Hassaku XL	Illustrious	Euler a	20-28	3-7 (6 ideal)	Excellent	~4s
WAI-NSFW Illustrious	Illustrious	Euler a	25-40	5-7	Good	~5s
NoobAI-XL	Illustrious	Euler a	25-30	5-6	Good	~5s
AutismMix Confetti	Pony	Euler a	25-30	5-7	Good	~5s
Prefect Pony XL	Pony	Euler a	25-30	5-7	Good	~5s

The takeaway: Hassaku’s standout trait is step efficiency. It reaches a clean result at lower steps than most peers, which is a real advantage on slower cards. WAI is the most forgiving overall, NoobAI has the deepest raw vocabulary, and the Pony-lineage models bring score-tag artist comprehension, but if you want fast, clean anime on a budget GPU, Hassaku is the pick.

A practical Hassaku workflow

This is the loop we use for finished images.

1. Generate at 832×1216, Euler a, 20 steps, CFG 6, to iterate fast thanks to Hassaku’s low-step tolerance.
2. When the composition is right, lock the seed and bump to 24 to 28 steps for the final.
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 img2img upscale pass at denoise 0.2 for 4K.

The low-step iteration loop is what makes Hassaku feel fast even on a 3060.

LoRA compatibility

Hassaku is Illustrious-based, so it works with Illustrious and NoobAI LoRAs, but with one caveat the author calls out: because v3 adds significant extra training, some LoRAs trained on plain Illustrious or NoobAI bind less precisely than they would on the raw base. In practice:

• NoobAI LoRAs tend to work better than plain Illustrious LoRAs on the v3 line.
• Start character LoRAs around weight 0.7 to 0.8 and adjust, since Hassaku’s extra training can fight a too-strong LoRA.
• Pony LoRAs do not transfer; keep those for Pony checkpoints like Prefect Pony XL or AutismMix.
• Keep stacked LoRA weight under about 1.5 for coherence.

Because Hassaku already has a strong, clean default look, you frequently get the result you want with a single character LoRA.

Common problems and fixes

Character tag is not binding. Reorder your prompt so character count and character tags come first, before the quality block, as the author recommends.

Stray logos or signatures appear. Add “signature” to the negative prompt, a documented Hassaku quirk.

Results look muddy. Remove Danbooru metadata and franchise tags; they degrade output on this model. Confirm Clip skip is 2.

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

A LoRA looks weak or distorted. Hassaku v3’s extra training can fight LoRAs. Try a NoobAI-trained LoRA, or lower the weight to around 0.7.

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

Tag weighting and emphasis in Hassaku

With the prompt order sorted, attention weighting is how you fine-tune. In Automatic1111 and Forge syntax, wrap a tag in parentheses with a number, like (detailed eyes:1.2) to strengthen or (background:0.8) to weaken. Hassaku, like the rest of the Illustrious family, responds smoothly to this.

Patterns we use often:

• Strengthen a focal feature, for example (looking at viewer:1.1), so the model commits to it.
• Tame an element that dominates, such as (depth of field:0.8) when the background is too blurred.
• Lock a mood with weighted lighting, like (window light:1.1).
• Keep individual weights under about 1.3; above that, faces and anatomy start to warp.

Because Hassaku binds character tags strongly when they lead the prompt, you usually need less weighting on the subject than on other models. Reserve your weighting budget for lighting, mood, and background instead.

Inpainting and editing Hassaku output

Running Hassaku locally lets you repair a single region rather than reroll the whole image. The loop:

1. Send the finished image to the inpaint tab.
2. Mask only the broken area, such as a hand or an eye.
3. Set inpaint area to “only masked” so the model uses full resolution on that patch.
4. Use denoise 0.4 to 0.6, lower to preserve more, higher to redraw more.
5. Keep the prompt with character tags first, or trim it to the masked subject.

For faces, ADetailer automates this loop, and because Hassaku is so step-efficient, ADetailer passes are fast even on a 3060. Manual inpainting still wins on tricky overlapping hands where you want precise control.

Getting the most from Hassaku’s step efficiency

Hassaku’s defining trait is that it reaches a clean result at low steps, and there is a workflow that exploits this fully. Run your exploration phase at 14 to 18 steps, which is fast enough to feel almost interactive on a 4090 and still very quick on a 3060. Generate a grid of seeds, pick the composition you like, then re-render only that seed at 24 to 28 steps for the final.

This two-tier approach, low steps to explore and higher steps to finish, gets you more good compositions per minute than running everything at full steps. On slower cards it is the difference between a frustrating wait between every image and a fluid creative session. We consider it the main reason to choose Hassaku over a heavier-stepping checkpoint when GPU time is limited.

A related tip: because Hassaku is efficient, you can afford a slightly higher hires-fix denoise, around 0.45, to let the upscaler add real detail, since the base render is already clean. On slower checkpoints we keep hires denoise lower to avoid drift, but Hassaku’s strong base tolerates a more aggressive finishing pass.

Running Hassaku on lower-end hardware

Below 12 GB of VRAM, Hassaku is one of the friendlier SDXL anime checkpoints precisely because of its step efficiency. On 8 GB cards, use Forge for better memory handling, generate at 832×1216, keep batch size at 1, and run hires fix as a separate img2img pass. Because you can drop to 18 to 20 steps without much quality loss, generation times stay reasonable even at 8 GB, often in the 25 to 40 second range.

Below 8 GB, local SDXL becomes a chore. That is when the on-site generator is the better tool: it runs the compute remotely so an older card or a laptop can still produce clean anime output. We treat it as the zero-setup option and local Hassaku as the power-user option for anyone with a capable GPU.

Why Hassaku is a great budget-GPU pick

The headline reason to choose Hassaku is efficiency. It produces clean, vibrant anime NSFW output at lower step counts than most of its peers, which translates directly into faster iteration on mid-range cards like the 3060. Combine that with a forgiving default aesthetic and the large Illustrious LoRA ecosystem, and you have a checkpoint that punches above its weight. We still reach for WAI when we want maximum forgiveness and NoobAI for raw vocabulary, but for fast, clean anime on a budget, Hassaku XL earns its place.

Settings recap and final verdict

All in one place: set Euler a, 20 to 28 steps with 24 as default, CFG 3 to 7 with 6 ideal, 832×1216, Clip skip 2, and use the inherited Illustrious VAE with no external file. Lead prompts with character count and character tags first, then the masterpiece quality block. Add signature to negatives if logos appear, avoid Danbooru metadata and franchise tags, and finish with hires fix at 1.5x, R-ESRGAN Anime6B, denoise around 0.4, then ADetailer on the face.

Our verdict after testing the v3 line: Hassaku XL is the best step-efficient anime NSFW checkpoint for budget GPUs in 2026. It reaches a clean, vibrant result at lower steps than its peers, which makes a real difference on a 3060, and its default aesthetic is polished enough that you often do not need style LoRAs. The tradeoff is that some plain Illustrious LoRAs bind less precisely on v3, but NoobAI-trained LoRAs cover that gap.

If you want maximum forgiveness, WAI is easier; if you want the deepest vocabulary, NoobAI wins; and for the Pony lineage, AutismMix or Prefect Pony are the picks. But if your priority is fast, clean anime on a card that is not a 4090, Hassaku XL is the model we recommend.

Download it from the Civitai page, set Euler a, 24 steps, CFG 6, 832×1216, and lead your prompt with the character tags. And if you just want one image without touching a GPU, the on-site generator is one click away.

Frequently asked questions

What are the best settings for Hassaku XL?

Use Euler a as the sampler, 20 to 28 steps with 24 as a solid default, CFG 3 to 7 with 6 ideal when paired with Euler a, and a resolution of 832×1216, which the author lists as most recommended. Clip skip should be 2, and the VAE is inherited from Illustrious and baked in, so no external VAE is needed. For finishing, add hires fix at 1.5x with R-ESRGAN 4x+ Anime6B and denoise 0.35 to 0.45.

Why is Hassaku XL good for low-end GPUs?

Hassaku is unusually step-efficient. It produces a clean, usable image at lower step counts than most peers, and can make something decent in as few as 11 steps. On our RTX 3060 12GB, dropping to 20 steps kept quality high while cutting generation time to around 15 to 22 seconds. That efficiency makes iteration fast on mid-range cards, which is a big reason Hassaku is popular with budget-GPU users.

How should I order tags when prompting Hassaku XL?

The author recommends a specific order: number of characters first, then character tags, then all remaining tags. For example, 1girl, rem (re:zero), standing, masterpiece, upper body. Putting the character count and character tags before the quality block helps Hassaku bind characters more reliably. The author also advises avoiding Danbooru metadata tags and franchise tags, which can muddy results on this model.

Does Hassaku XL need a special VAE?

No. Hassaku XL inherits its VAE directly from Illustrious-v1.1, and it is baked into the checkpoint, so you do not need to load a separate VAE file. That inherited VAE is what gives Hassaku its consistent color palette, saturation, and tone within the Illustrious family. If your colors look off, the usual cause is CFG or Clip skip rather than the VAE, so check those first.

What quality tags does Hassaku XL use?

Hassaku is Illustrious-lineage, so it uses the Illustrious quality vocabulary, not Pony score tags. Use masterpiece, best quality, amazing quality, very aesthetic, absurdres in your positive prompt, placed after the character tags. Keep the negative lean, with worst quality, low quality, bad anatomy, bad hands, and censored. Add signature to the negative if stray logos or signatures appear, which is a documented quirk of this model.

Why do logos or signatures show up in my Hassaku images?

This is a known Hassaku quirk caused by signatures and logos present in some of the training data. The fix is simple: add signature to your negative prompt, and you can also include watermark and text. The author specifically recommends including signature as a negative when disruptive elements appear. After adding it, the stray marks usually disappear without affecting the rest of the image.

Do my Illustrious LoRAs work with Hassaku XL?

Mostly, but with a caveat. Because the v3 line adds significant extra training, some LoRAs trained on plain Illustrious bind less precisely than on the raw base. The author notes that NoobAI-trained LoRAs tend to work better than plain Illustrious LoRAs on this version. Start character LoRAs around weight 0.7 to 0.8 and adjust. Pony LoRAs do not transfer at all, since Hassaku is Illustrious-based.

Is Hassaku XL better than WAI for NSFW?

They have different strengths. Hassaku’s standout trait is step efficiency, making it faster on budget GPUs, with a clean, vibrant default look. WAI is the most forgiving overall and the easiest for total beginners, with a broad LoRA ecosystem. Both are Illustrious-lineage and share quality tags, so prompts carry over. We keep both installed and pick Hassaku when we want speed on a 3060 and WAI when we want maximum forgiveness.