What Illustrious XL v0.1 Actually Is
Illustrious XL v0.1 is an SDXL fine-tune released by OnomaAI in late 2024, trained on roughly 7.5 million Danbooru-tagged anime and illustration images — about three times the dataset size of Pony Diffusion XL, but exclusively in the anime domain. It uses plain Danbooru tag prompting with a short aesthetic prelude, no score_9 system, no safety alignment layer. By mid-2025 it had effectively replaced Pony as the default NSFW-permissive anime SDXL base, and most major anime fine-tune authors had migrated or dual-tracked their work onto it.
The model ships under the Fair AI Public License 1.0-SD, which is commercial-permissive in a way that the community generally finds easier to live with than Pony's licensing posture. The UNet is standard SDXL, so the entire SDXL ecosystem — LoRAs, ControlNets, IP-Adapters, regional prompters — works without special handling. You drop it into ComfyUI, Forge, or A1111 like any other SDXL checkpoint. There is no exotic loader, no custom node, no required preprocessor.
Why It Exists: The Post-Pony Migration
Through most of 2024, Pony Diffusion XL V6 was the dominant NSFW-capable SDXL base. The numbers are absurd in retrospect: roughly fifteen thousand LoRAs built on Pony, an enormous slice of all Civitai SDXL traffic, and a near-monoculture among anime and stylized fine-tune authors. Pony was the base. The score_9, score_8_up, score_7_up... prelude was the lingua franca. Authors trained against it because that was where the audience was.
Then Civitai changed how it ranked content. The mid-2024 algorithmic shifts deprioritized Pony-derived NSFW resources on the front page and in discovery surfaces. The exact mechanics were never fully public, but the practical effect was visible: a Pony V6 LoRA that would have ridden the trending shelf in early 2024 was getting buried by late 2024. Authors who depended on visibility for downloads, donations, and commission inflow started looking for alternatives. Plenty of anime-fine-tune people had already been quietly uncomfortable with the multi-domain Pony aesthetic anyway — the slight cartoon-leaning bleed, the anthro tag bias, the score-prelude ergonomics.
OnomaAI shipped Illustrious XL v0.1 at exactly the right month. Cleaner license. Anime focus instead of multi-domain mix. Plain Danbooru tag prompting with no score scaffolding. Standard SDXL UNet, so existing tooling worked unchanged. The migration didn't happen overnight — it moved in waves through early-to-mid 2025 as individual authors retrained their flagship LoRAs and announced Illustrious-base versions. Pony built the ecosystem, Civitai's algorithm changes drove it elsewhere, OnomaAI shipped at the right month. That's the whole story.
Inside the 7.5M-Image Anime Dataset
The dataset is the model. Illustrious's roughly 7.5 million training images are heavily Danbooru-derived, which matters in two specific ways that Pony's mix doesn't replicate.
First, Danbooru's tagging is character-dense. Volunteers tag named characters from anime, manga, games, and visual novels at a granularity Pony's broader scrape never matched. The downstream consequence is that Illustrious genuinely knows a much larger roster of anime IPs by name. Type a character tag from a popular series and you get a recognizable likeness without a LoRA. Type a character tag from a moderately obscure 2010s seinen series and you still get something usable. Pony covers some of this, but its multi-domain dataset diluted the per-character signal.
Second, the domain is clean. Pony was trained across anime, anthro, cartoon, semi-realistic, and a few stylized hybrids. The model averaged across all of that, which is why Pony's "pure anime" output has a faint anthro-cartoon undertone that the experienced eye picks up immediately. Illustrious doesn't have that bleed because there's nothing to bleed from. Its prior is anime, full stop. The aesthetic is more coherent, the style transfers cleaner, the explicit content sits in a stable anime register without prompt scaffolding to keep it there.
Training philosophy clean-domain-over-broad-domain has a cost — narrowness, which we'll get to — but for the central use case it pays off in every direction.
Plain Tags Win: No score_9 Prelude
The prompt syntax difference is small in characters and large in ergonomics.
Pony required:
``text score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime, 1girl, solo, ... ``
Illustrious takes:
``text masterpiece, best quality, very aesthetic, 1girl, solo, ... ``
The aesthetic prelude — masterpiece, best quality, very aesthetic — is convention rather than requirement. The model responds to it because Danbooru tags it on top-rated images, but you can drop the prelude entirely and the output stays usable. The score-tag system in Pony was load-bearing in a way Illustrious's prelude isn't: omit score_9, score_8_up... from a Pony prompt and quality drops visibly. Omit masterpiece, best quality from an Illustrious prompt and you get a mildly less polished result, not a structural collapse.
Standard Danbooru tags do the actual work. 1girl, solo, looking_at_viewer, white_shirt, sitting, indoors produces the expected composition with no further coaxing. Faster to learn for anyone migrating from base SDXL or Animagine. Less verbose. Less brittle to weight-tuning, since you're not balancing six score tags against your content tags. Negative prompts get simpler too — no score-tag negation needed, just the conventional lowres, bad anatomy, text, watermark and whatever style negatives you're rejecting.
For anyone who learned SD on Danbooru-style prompting in the SD1.5 era, Illustrious feels like coming home.
What Illustrious Does Better Than Pony
The strengths cluster in predictable places.
Pure anime aesthetics with no Pony-leaning bleed. The output looks like illustration the way an anime artist would draw it, not the way a multi-domain model averages toward illustration. Style LoRAs sit on top of this prior cleanly, without having to fight an underlying anthro-cartoon undertone.
Cleaner human anatomy at the base level. Hands are noticeably better out of the box. Proportions stay consistent at portrait and landscape resolutions without the mild Pony tendency toward stylized exaggeration. None of this is perfect — SDXL is SDXL, and hand-disasters happen — but the floor is higher.
Named-character knowledge. The character-dense Danbooru tagging gives Illustrious a substantially deeper roster than Pony in the anime/manga IP space. For named-character work on popular series, you often don't need a LoRA at all. For obscure characters, the base model gets you closer to a starting point that a small LoRA can finish.
Shorter learning curve for SDXL veterans. If you already know Danbooru-tag prompting from any prior anime SDXL or SD1.5 model, Illustrious requires zero new vocabulary. Pony's score-tag system was a real obstacle for newcomers and a real annoyance for everyone else.
Cleaner license and distribution dynamics. Fair AI Public License 1.0-SD is commercial-permissive, and Civitai's algorithmic posture toward Illustrious-base resources is friendlier than its current posture toward Pony-base ones.
Aesthetically-stable explicit content. Illustrious produces uncensored output without needing prompt scaffolding to hold the aesthetic in register. The native anime priors stay intact across SFW and NSFW prompts equally. Pony does this too, but the score-prelude-and-content-tag balance was always a little fiddlier.
What Pony Still Does Better
Pony hasn't been replaced everywhere. Honest about the gaps.
Anthro, feral, and furry. Illustrious essentially cannot do these without heavy LoRAs and significant prompt work. The dataset doesn't cover the domain in any meaningful quantity. Pony was trained partly on this material and produces strong anthro and feral output natively. If your work is in this space, Pony is still the base, full stop.
Western cartoon styles. Illustrious's prior is anime-Japanese illustration. Western cartoon — the show-bible style of US-produced animation — sits outside the training distribution. Pony covers it adequately at the base level.
Cross-domain hybrid styles. Anything that mixes anthro plus anime plus cartoon plus stylized realism in a single output is squarely Pony's home turf. The multi-domain dataset that limits Pony's pure-anime ceiling is exactly what enables this.
Larger total LoRA library. The V6 ecosystem is older and bigger. There are still characters, styles, concepts, and outfits that exist as Pony LoRAs and don't yet exist as Illustrious LoRAs. The gap is closing through 2025-2026 but it's real.
Realism via the dedicated Pony-realism fine-tune family. The fine-tunes that pushed Pony toward photoreal output — CyberRealistic Pony, Babes, RealMixPony — produce a category of output Illustrious has no equivalent for. Illustrious is anime-illustration. There is no Illustrious-realism branch of comparable maturity. If you want photoreal NSFW with the prompt ergonomics of an anime base, Pony's realism fine-tunes are the only viable answer right now.
Hardware: Same SDXL Floor
Nothing exotic. 8 GB VRAM is the realistic minimum for 1024x1024 generation, 12 GB is comfortable for hires-fix workflows and ControlNet stacks, 24 GB is excessive unless you're running heavy multi-pipeline setups or training. Generation speed at matched resolution is identical to any other SDXL fine-tune — Pony, Animagine, base SDXL — because the UNet is standard SDXL.
Compute is not a deciding factor between Pony and Illustrious. If your machine runs one, it runs the other at the same speed. ComfyUI, Forge, and A1111 all load Illustrious as a standard SDXL checkpoint with no special handling, no custom nodes, no preprocessor pipeline. Drop the file in your models directory, select it, generate.
The only mild gotcha: some early Illustrious-base fine-tunes shipped with quirky baked-in VAE behavior, and a few prefer the SDXL VAE explicitly loaded rather than relying on the embedded one. If output looks washed or color-shifted, swap to the standard SDXL VAE first before debugging anything else.
Recommended Sampler Settings
A reasonable default starting point:
``text Sampler: Euler a or DPM++ 2M Karras Steps: 25-30 CFG: 4.5-6.5 (lower than Pony — Illustrious is less cfg-tolerant) Resolution: 1024x1024 native, 832x1216 portrait, 1216x832 landscape Clip skip: 2 (standard for anime SDXL) Hires fix: latent upscaler 1.5x, 0.4 denoise ``
Note the CFG sits a touch lower than Pony's typical 7-9 range. Pushing CFG above 7 on Illustrious burns the output in ways it doesn't burn Pony — saturation goes wrong, faces get crispy, fine line detail breaks. The model is less CFG-tolerant by design. Stay in the 4.5 to 6.5 band, lean toward 5-5.5 for most work, and reach for higher CFG only when you genuinely need the prompt to dominate.
The 832x1216 portrait and 1216x832 landscape resolutions are the safe non-square sizes — staying inside the SDXL bucket-resolution training distribution avoids the proportion drift you get at off-bucket sizes. Hires fix at 1.5x with low denoise is usually enough; aggressive denoise on Illustrious tends to over-stylize.
The Illustrious Fine-Tune Family
The ecosystem built up through 2025 in waves. The big names mostly came in three batches: the early adopters who shipped Illustrious-base versions in Q1 2025, the mainstream migration through Q2-Q3, and the dual-trackers who kept Pony alive but made Illustrious their primary platform.
Hassaku is the canonical example of the migration. Originally a Pony-base anime fine-tune with a strong following, it shipped an Illustrious-base version mid-2025 and the Illustrious branch quickly became the primary line. The author's tagging conventions and aesthetic priorities translate cleanly to the new base.
Prefect is a clean refinement of the Illustrious base — less aggressive stylistic re-weighting than Hassaku, closer to "Illustrious but more polished." Good choice for users who want the base model's behavior with sharper output and tighter aesthetic priors but no strong stylistic personality imposed on top.
WAI's Illustrious-side line, parallel to their Pony work. Solid all-rounder anime fine-tune with strong character knowledge inherited from the base and the WAI team's usual attention to anatomy and prompt-following.
NoobAI XL is the "more permissive Illustrious cousin" — a related fine-tune line that pushed Illustrious's training further in specific directions, with a reputation for handling explicit content with more aesthetic stability than the base. It's not in our catalog, but anyone working seriously with Illustrious eventually encounters it.
JANKU and RouWei are adjacent — smaller-author Illustrious-base fine-tunes that have built loyal followings for specific aesthetic niches. Animagine XL (the older one) sits in adjacent territory as the previous-generation anime SDXL fine-tune that Illustrious effectively succeeded.
Where Illustrious Stands in April 2026
Illustrious 2.0 shipped early 2026, and the most important thing about it is what it didn't do: it didn't change base. Still SDXL. The same UNet architecture, the same ecosystem compatibility, the same LoRA format. v0.1 LoRAs are mostly forward-compatible to 2.0 with minor weight adjustments — some authors recommend a 0.7-0.9 weight scaling for v0.1 LoRAs run against 2.0, but the LoRAs work.
Compare Pony's V7 trajectory. Pony V7 jumped to AuraFlow as its base model. That was a defensible technical choice — AuraFlow is more capable than SDXL on paper — but it broke the existing LoRA library. The fifteen-thousand-LoRA V6 ecosystem doesn't run on V7. Authors who built their work on Pony V6 face a full retrain to migrate, and many haven't bothered. The V6 ecosystem is now a frozen archive that runs on the V6 base; the V7 ecosystem is a smaller, newer thing that's still building up.
Illustrious 2.0 preserved continuity. The fine-tune ecosystem is splitting along these lines: a smaller group of authors is retraining for Pony V7, a larger group is doubling down on Illustrious 2.0, and the ecosystem-stability tea-leaves favor the latter. If you're starting a serious image-gen project in 2026 and you need a base that will still have a living LoRA ecosystem in 2027, Illustrious 2.0 is the safer bet.
v0.1 itself remains useful — it's the version most existing fine-tunes still target, and the version most LoRA training documentation was written against. As a learning version and a stable production target it has at least another year of relevance.
When To Pick Illustrious vs Pony
Decision matrix that actually maps to use cases.
Pick Illustrious for: pure anime work where the aesthetic needs to stay clean, projects where human anatomy quality matters and you don't want to layer a realism fine-tune, named-character anime knowledge from the base model, shorter prompts and faster iteration, and any fresh 2026 project with no Pony legacy to maintain.
Pick Pony for: anthro, furry, and feral work — Illustrious genuinely cannot do this domain well. Cross-domain hybrid styles that mix anthro plus anime plus cartoon. Projects that depend on the deep V6 LoRA library for specific characters or concepts that don't yet have Illustrious equivalents. Photoreal NSFW via the dedicated Pony-realism fine-tune branch — CyberRealistic Pony, Babes, RealMixPony.
Pick both: most serious image-gen users keep both bases on disk and switch by job. The disk cost is six gigabytes per checkpoint. The mental overhead of switching is one dropdown selection. Treating this as a religious choice is a beginner mistake; treating it as two specialized tools in the same kit is how working users actually operate.
