⚙️ ComfyUI workflows included 🇯🇵 Danbooru tags supported
💬 New Discord, come play with us: https://discord.gg/rHCnjX9cW9
🌟 Starlight XL 星光 Animated
🔮 3.6 cfg 🪜 40 steps 🤖 DPM++ 3M SDE Karras
Starlight is a powerful 2.5D like SDXL model designed for creating stunning concept art and illustrations.
Key Features
Excels in anime, fantasy, and semi-realistic styles.
Designed for rich details and mesmerizing visuals.
Evolving based on community feedback.
How to Use
Starlight shines over a large range but performs best with a LOWER CFG setting (3 to 5). Lower CFG provides finer, richer details, aligning with the model's intended goal. Higher CFG offers more color but may have increased artifacts.
Example Prompt: 1girl, masterpiece, best quality, highly detailed, sharp focus, dynamic lighting, vivid colors, texture detail, particle effects, storytelling elements, narrative flair, 16k, UE5, HDR, subject-background isolation
Example Prompt: 8k portrait enchanting depiction of an iridescent dress worn by a beautiful woman, intricately detailed, digital masterpiece, mesmerizing play of colors and light, ultimate elegance, exquisitely detailed shimmering fabric, hdr
Sampler: DPM++ 3M SDE Karras
Recommended Steps: 25 to 65 (40)
CFG / Guidance: 3 to 5 (3.6)
Aspect Ratios:
2:3 — 832x1248
16:9 — 1360x768
4:5 - 912x1144
1:1 — 1024x1024
I really hope you enjoy this SDXL model — and if you do, please give Starlight 星光 a 5-star review!
☕️ Consider supporting Starlight's development by buying me a coffee!
🤝 Interested in partnering with me for innovation and marketing opportunities? Contact me at [email protected]
Description
This update introduces greater detail and knowledge, while preserving Starlight's distinctive style and composition
FAQ
Comments (54)
V3 is prolly the best model in SDXL now, I bet lots of those unskilled model merge people will mergeit and call it theirs xD
It's my favourite XL checkpoint so far. I consistently get good results with lots of different things I try, compared to other models.
Thanks for releasing v3! As v2 is my favorite model and has basically rendered all competition dull (yes, also the community favorites which I tested could not match composition and style of starlight v2) I am very much looking forward to testing the new release during the next few days!
I strongly agree with this, most of the community favorites are overrated, they are more of marketers than the real model makers. I hope real models like these will get the light on civitai.
You've got a knack for pushing a model to its limits! I enjoy your work and can't wait to see what you come up with next. 🤘
@chillpixel Thanks man, I'm on it!
Can you tell me a good settings for this Model with a refiner Please?
I don't use the refiner, but would keep the denoising value between 0.15 and 0.24
For refinement, including Img2Img techniques, I like to use Euler sampler with 35 steps 5 cfg
Here's a comment where I describe my current refinement workflow: https://www.reddit.com/r/StableDiffusion/comments/177id3w/comment/k4utbur/?utm_source=share&utm_medium=web2x&context=3
Refiner should only be used for photo realistic style images.
From SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis
We note that this step is optional, but improves sample quality for detailed backgrounds and human faces, as demonstrated in Fig. 6 and Fig. 13
I love this model. It can do anime, it can do realistic, it can even do some decent nsfw art.
Great job @chillpixel
I am guessing higher CFG is more ideal for anime like 4?
I usually use 40 steps with a 3.6 cfg for most art styles. If a style requires more vivid colors, I opt for 5.8 cfg. I've found these exact step and cfg values to work well with SDXL models.
On platforms like https://tensor.art/models/650104649833223064, which limit you to 30 steps on their free plan, I stick to 30 steps with a 4.4 cfg and 7 cfg.
@chillpixel Seems like you are saying that with > 30 steps, one can use a lower CFG, which seems to make sense.
But what is the advantage of using a lower CFG value? Most of the time people seems to recommend an CFG of around 7, and SDXL supposedly works well even with CFG > 10 (compared to SD1.5).
@NobodyButMeow i apologize for the late reply... I've noticed that using a lower CFG can reveal more details in certain types of artworks, especially realism. On the other hand, a higher CFG can make the colors pop, but it might sometimes make the image look 'overbaked' and add some weird stuff to it.
Generally, with a lower CFG, the AI has more room to rely on its intuition while denoising up the image.
I've primarily been using SD 1.5 models and have been quite satisfied with their performance. However, after rigorously testing ten of the most popular SDXL models, I was astounded by Starlight's responsiveness to prompts. The need for negative embeddings seems almost negligible. The capability of this checkpoint is truly remarkable!
good afternoon) Which VAE one did you use? 8400 spoils the image
no one... disable VAE when use sdxl checkpoint
@ponze2007 Thank you
Blessed V2 seemed to work for me quite well.
Could you please make a SD 1.5 version of this? As sadly i'v not gotten SDXL to work with tensorRT yet.
Hey!
I am not sure how to use this for animations.
Does animate diff work with this checkpoint?
Hey, when I say "animated" in the title, I'm talking about the 2.5D style of images/artworks that Starlight XL is designed to create, not that it's actually for animation. But the good news is that - once SDXL support is added for AnimateDiff, you'll be able to use this model and it should perform well.
The model seems great, but for some reason, I am using automatic1111, at the last tick, suddenly it goes from being beautiful, into being a yellow mustard mess. Anyone know what I am doing wrong? It has to be some setting issue.
if vae is set, try setting it to none... that was happening to me on another model.
I agree with @danielteeny956 If you notice "it's changing at the end" its almost certianly your VAE
23 Days late; Did you resolve your issue? Disable this setting: Apply color correction to img2img results to match original colors.
@yejokok902337 the correct fix is to disable any VAEs (I forgot to disable ClearVAE and i got a colorful mess as output, on both A1111 and ComfyUI)
Great job, thanks for the model !
You're welcome. Thanks for sharing your awesome art. It's really cool to see all the creative stuff you come up with!
One of the best SDXL models in my experience. Can do almost anything.
Also works REALLY well with LCM (Latent Consistency Model)
I've got all the other models, but I keep coming back to this one. It deserves way more downloads.
Thanks so much for the work you put into this model. Everything seems to come out looking great. Well done, it's now my go-to model.
Seeing as this is my favorite model, any chance of a turbo XL version? Thanks.
RevAnimated in SDXL era!
gorgeous model! do you plan to develop it further?
this is great <3
Can I use this model for commercial purposes?
This model is incredible and deserves more praise than it receives
Hey chillpixel, genius who made this model, is there a timeline for a new version? Not that I'm unhappy with the current version. Quite the opposite, it one of the BEST! Just curious.
any SDXL user who doesn't have Starlight on their drive doesn't know SDXL
perfect
The model shows up as 0.9 not 1.0 in metadata - I noticed that after realizing that model specific/SDXL enhancements weren't applied in StableSwarm when using this model.
Had to chime in also.. great model.. def a keeper.
it's a lot more than what it is advertised to be.
its still the best model its funny to me how model from december is still better than modern models of xl, i hope we could see the update for V4 sometime as SD3 wasnt good at all.
Beautiful colors, well done!
Really impressed. It makes wonderful dream like images. This is a keeper.
Hello! This checkpoint has been one of my all time favourites. Are you planning on port it to Flux? That would be awesome! I think it will benefit a lot from Flux new architecture
easy and forgiving model, few steps, old GPU, just awesome
Why this checkpoint is no longer available in on site generation ?
Can it come back ?
I love this model, when will it be back on Civita?
If you run SDXL locally and have a handful of checkpoints downloaded on your drive, Starlight is one of them. It has been, since basically forever.
If not, you're doin' it wrong.
Agreed .... I have 3 SDXL models, one being Starlight!
Anyone know what happened to "chillpixel"? It seems like he vanished 2 years ago. Any alternative suggestion for Starlight XL?
This model is a hidden gem, top on my personal list. To think this amazing model was released 2 years ago is.... damn.
Details
Files
starlightXLAnimated_v3.safetensors
Mirrors
starlightXLAnimated_v3.safetensors
starlightXLAnimated_v3.safetensors
starlightXLAnimated_v3.safetensors
starlightXLAnimated_v3.safetensors
starlightXLAnimated_v3.safetensors
starlightXLAnimated_v3.safetensors
starlightAnimated_v3.safetensors
starlightXLAnimated_v3.safetensors
starlightXLAnimated_v3.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.