🎨 Anime Illustrious Collection – Chroma · Z-Image · Qwen · Ernie
Transform your generations into authentic Illustrious-style anime art — now available as LoRAs for four different model architectures: Chroma HD, Z-Image, Qwen-Image, and the new Ernie-Image.
Perfect for detailed characters, vibrant colors, expressive scenes, and cinematic anime compositions. Actively developed and improved through community feedback and real-world testing.
🚀 Latest Releases
🆕 Ernie-Anime-V1 (April 2026) — First Illustrious LoRA for Baidu's Ernie-Image (works with Base and Turbo)
Z-Image-Anime-V1 (November 2025)
Qwen-Anime-V2 (November 2025) — Most refined Illustrious version yet
Qwen-Anime-V1 (October 2025)
Chroma-Anime-V3 (October 2025) — Recommended for Chroma HD
📌 Quick Start
Prompt starter: always begin with anime illustrious, ... Strength: 0.8 CFG Scale: 3–5 (Chroma / Qwen / Z-Image Base) · 4.0 (Ernie-Image Base) · 1 (all Turbo variants) Steps: 20–30 on base models · 8 on Turbo variants · 50 on Ernie-Image Base Sampler: Euler Clip Skip: 1
Use natural, descriptive language instead of comma-separated tag strings. The style is activated by the phrase anime illustrious at the start of the prompt.
⚡ Version Overview
🆕 Ernie-Anime-V1 — April 2026 (NEW)
The first Illustrious-style LoRA built for Baidu's Ernie-Image, an 8B single-stream DiT architecture known for its strong instruction following, text rendering, and structured composition. Adapted from the refined Chroma-Anime-V3 dataset and retrained for Ernie's generation engine.
One LoRA, both model variants — Ernie-Anime-V1 is fully compatible with:
Ernie-Image (Base) — the SFT 50-step model for maximum fidelity
Ernie-Image-Turbo — the DMD+RL distilled 8-step variant for fast iteration
🌟 Key Highlights:
✅ Single LoRA that runs on Base and Turbo
✅ Preserves Ernie's strong text rendering inside the anime aesthetic
✅ Clean Illustrious cel/painterly hybrid look
✅ Excellent for posters, multi-panel layouts, and layout-sensitive scenes
✅ SFW artistic focus with anatomical accuracy
✅ Stable character consistency at 1024px and portrait aspect ratios
🧩 Technical Details:
File Size: ~360 MB
Training Steps: 5100
Dataset Base: Z-Image-Anime (adapted for Ernie)
Resolution: 1024px
Focus: faces, hands, clothing, lighting, composition
Recommended settings per variant:
Ernie-Image Base → 50 steps, CFG 4.0, strength 0.8
Ernie-Image-Turbo → 8 steps, CFG 1, strength 0.8
🆕 Z-Image-Anime-V1 — November 2025
Illustrious-style LoRA adapted for Alibaba's Z-Image architecture (S3-DiT, 6B). Compact, stylistically consistent, and suitable for both standard and turbo workflows on Z-Image.
✅ Painterly cel-shaded Illustrious look
✅ Works with Z-Image Base and distilled (Turbo) variants
✅ Stable character rendering at 1024px
File Size: ~162 MB
Training Steps: 3000
Qwen-Anime-V2 — November 2025
The next evolution of the Illustrious look for Qwen-Image. Trained with over 40% more data and extended fine-tuning epochs, V2 offers richer variety, deeper detail, and more stylistic flexibility.
🌟 Key Highlights:
✅ Expanded dataset — +1,000 new curated anime frames for better diversity
✅ Sharper eyes, refined hands, improved textures
✅ Balanced color depth and glow (lighting & tone control)
✅ Wider expression range for emotional and dynamic scenes
✅ Cleaner multi-character compositions in group shots
✅ Compact & efficient — ~420 MB with ~3,200 training steps
🧩 Technical Details:
File Size: ~420 MB
Training Steps: ~3,200
Dataset Base: Chroma-Anime-V3 (expanded)
Resolution: 1024px
Focus: faces, hands, clothing, lighting, scene harmony
🔞 Includes artistic nudity for anatomical accuracy, but will not produce explicit NSFW content alone.
Qwen-Anime-V1 — October 2025
The first Qwen-optimized Illustrious LoRA — a painterly cel-shaded hybrid with SFW artistic focus. Built on the Chroma-Anime-V3 dataset and adapted for Qwen's generation engine.
✅ First Qwen-optimized Illustrious release
✅ Compact & stable (~415 MB, ~2,200 steps)
✅ Strong baseline before the V2 refinement
Chroma-Anime-V3 — October 2025
The most refined Chroma edition — smaller, faster, and more diverse than previous versions.
✅ 65% smaller file size vs. V2
✅ 154 diverse training images
✅ Multi-character stability
✅ Improved age balance
✅ Enhanced focus on eyes, hands, and poses
💎 Core Features (Across All Versions)
🎭 Authentic Illustrious aesthetic — vibrant, expressive anime style 🎨 Clean cel-shaded / painterly balance with sharp anatomy 👗 High character fidelity with textured clothing and detailed eyes 📸 Flexible composition — portraits, full-body, dynamic angles 🌅 Complete anime-style environments and backgrounds 🎯 Stable, consistent results with minimal artifacts
📖 How to Use
Prompt structure: write in natural descriptive sentences, not in tags. Open every prompt with anime illustrious to activate the style, then describe the scene as you would to a human illustrator.
Example prompt:
anime illustrious, a young woman with long flowing silver hair standing on a temple balcony at sunset, soft golden light falling across her face, cherry blossom petals drifting in the wind, detailed embroidered kimono, cinematic composition, shallow depth of field
Tips:
Always start with
anime illustriousDescribe in sentences — avoid long comma-separated tag strings
Lower CFG softens the style; higher CFG pushes it more strongly
25+ steps noticeably improve hands and poses on non-Turbo models
For Turbo variants (Ernie-Turbo, Z-Turbo), keep steps low (≈8) and CFG modest (≈2.5–3.0)
🎯 Quick Version Comparison
Chroma-Anime-V1 — Original release. Needs higher strength (1.2+).
Chroma-Anime-V2 — Better backgrounds, natural prompting, improved stability.
Chroma-Anime-V3 — Most balanced Chroma edition. Multi-character support, smaller size.
Qwen-Anime-V1 — First Qwen version. Painterly cel-shaded hybrid, SFW focus.
Qwen-Anime-V2 — Expanded data, longer training. Most refined Illustrious release so far.
Z-Image-Anime-V1 — First Illustrious look for Z-Image.
Ernie-Anime-V1 — First Illustrious LoRA for Ernie-Image. Works on Base and Turbo.
💬 Community & Support
❤️ Enjoying these LoRAs? Leave a review and share your art! 🐛 Found issues? Report them — feedback drives updates. 💡 Future updates (Chroma V4 / Qwen V3 / Ernie V2) are already in planning.
📢 Follow for updates — new features, better performance, and optional NSFW add-ons in development.
🕓 Version Timeline
Ernie-Anime-V1 (April 2026) — First Ernie-Image Illustrious LoRA (Base + Turbo)
Z-Image-Anime-V1 (Nov 2025) — First Illustrious LoRA for Z-Image
Qwen-Anime-V2 (Nov 2025) — Expanded data, improved details & diversity
Qwen-Anime-V1 (Oct 2025) — First Qwen-optimized version
Chroma-Anime-V3 (Oct 2025) — Multi-character refinement
Chroma-Anime-V2 — Background & strength fix
Chroma-Anime-V1 — Foundation of the Illustrious aesthetic
✨ Created with passion for the anime AI art community ❤️ Download your preferred version and experience the refined Illustrious look — on Chroma, Z-Image, Qwen, or Ernie!
Description
Still under construction
FAQ
Comments (15)
how did you get ai toolkit to train zimage, for me it sits forever on "starting job"
It worked for me right away, so I can’t really answer your question. All I can say is that I used the latest version and pointed it to a diffusion folder from my setup.
Let me find a YouTube link for you — maybe it’s easiest if you just follow the steps shown there.
Excellent work, as always, however I also have problems with glitter or particles, sometimes you add them and it is difficult to remove them, if you have any suggestions within the promp to do so I will thank you even more, good job!
@Zazalael Hey, thanks a lot for your question!
The “sparks” are actually intentional — that’s part of the training data.
Here’s an example image that shows how it’s meant to look:
https://civitai.com/images/107168859
I aimed to make the LoRA capable of creating very detailed anime-style images with strong colors and a lot of visual effects, but still generating them instead of just copying. I think it turned out quite well 🙂
I personally really like this type of anime illustration.
If you prefer a cleaner result, you can reduce the LoRA strength a bit — around 0.6 usually works fine.
You don’t really need an additional detail LoRA either.
Thanks, if the strength set to 1 it brings me back memories to SD1.5 days lol. Anyway I have also tried to train style Lora with ai-toolkit based on Ostris tutorial, but it turned out to be pretty bad even at 3000 steps. May I know did you use WD14 captioning to tag the dataset or something else?
I do a lot of this manually in WD14. The problem is that WD14 mainly uses tags, which works perfectly for SD1.5 / SDXL / ILL / Pony. But for models like Flux, Qwen, Chroma, Z-Image, and even wan2.x, they prefer natural language. You can either mix both approaches or use the correct style for the models I just mentioned. ^^
Also, even when using WD14, always double-check what’s in your .txt file. For example, if it says “blonde, long hair” but the character actually has short dark hair, that will cause problems. ^^
One more tip: less is often more. You’re not creating the image by writing every single detail—you’re just describing the main subject briefly. Example: “Son Goku from Dragonball, standing upright with both arms raised, black hair, blue eyes, wearing a blue tracksuit with yellow pants and dark shoes.” Even that is almost too much. I don’t need to describe the background if I don’t want to, or small details like folds in the clothing or half-open eyes.
I hope that makes sense! 😊👍
@soralz Still one of the best caption models for this: https://huggingface.co/spaces/fancyfeast/joy-caption-alpha-two
@SeeSeeLP Thanks!
Looks like author don't know what is prompt and how it works. All your prompts literally starts with "You are a masterful AI artist who..." lol, wtf
Hi there, and thanks for your question! ^^
Those lines you mentioned are just leftovers from my Lumina prompts. In Lumina you have to start the prompt with an introduction, something like:
“You are an assistant designed to generate anime images based on textual prompts. <Prompt Start> Tags: …”
For anyone who doesn’t know: Lumina’s text encoder is based on Google’s Gemma-2B model, and it requires this kind of structure. If you leave it out, it literally produces an empty image (meaning no image at all).
BUT:
You’re right that in Z-Image, which uses a Qwen text encoder, this isn’t necessary. It simply ignores that whole “You are a masterful AI artist who…” part, so it doesn’t affect the result in any way.
Could I have deleted those lines? Yes.
Did I bother? No, because the output would be the same either way. LoL.
I hope this explains why those sentences were still in my prompts.
Feel free to take a look at how the CLIP Text Encode (Prompt) node actually works before making assumptions — spoiler: it can handle more than plain text with dots and commas. ^^
Hope this helps clear things up!😅✌️
@magenta I've found this video for you again, it shows it very well: https://youtu.be/RXuTNuyM6GI?si=LGCpKmOogYVrNGZx&t=49
Best regards
@SeeSeeLP Thanks man, no pressure, your lora actually very good!
@magenta Thank you so much 😊
You described the advantages and development of different versions of the Qwen and Chroma models, but you didn't say anything about ZiT. Is it possible to find out how good she is when working with multiple characters, what about the problem of attributes flowing from one character to another, etc.?