For ASTRA use CFG of 1 !!
Astrallus (LCM Merge)
Astrallus is a custom-merged Latent Consistency Model (LCM) checkpoint designed for lightning-fast image generation with beautiful, vibrant results!
By combining distinct models rather than training from scratch, Astrallus is curated to deliver stunning characters, captivating digital art, and colorful scenes in a fraction of the time. Say goodbye to long render waits and hello to instant artistic gratification!
✨ Key Features:
Blazing Fast: Optimized for rapid iteration and creative exploration.
Vibrant & Colorful: Excels at generating images with rich, appealing color palettes—perfect for bringing your characters and digital illustrations to life.
Character & Digital Art Focused: Whether you're crafting dynamic heroes, fantastical creatures, or intricate digital paintings, Astrallus handles it with style.
LCM + SGM_Uniform: This merge leverages the power of LCM for speed combined with Karras noise scheduling for improved detail and quality.
⚙️ Recommended Settings (ComfyUI / A1111):
Sampler: LCM (or Euler A)
Steps: 5 - 15
CFG Scale: 1.0 - 1.5
⚠️ Important Note on Negative Prompts:
At CFG 1.0: Negative prompts are ignored by the sampler. The model relies entirely on your positive prompt.
At CFG 1.1 - 1.5: Negative prompts will have a minor effect.
Recommendation: Keep your positive prompts descriptive and clean. If you absolutely need a negative prompt to remove an element, try a CFG of 1.2-1.5, but 1.0 is the "sweet spot" for speed and stability.
Description
LCM 1CFG 15 steps and below
Ultra Fast
FAQ
Comments (3)
I really like your showcased pics style, nice job.
Two questions after reading your recommendations tho. You say that this checkpoint benefits from having been trained on LCM (sampler) and Karras (scheduler) but then in your recommended setting you say to use LCM (sampler) with either simple (scheduler) or SGM uniform (scheduler).
Also you recommend a CFG scale set at 1-1.5. On Forge (A1111 fork) setting the CFG scale below 1.5 prevents you from using a negative prompt.
I don't get it. 😜
I use comfyui for generations, I always keep the same negative prompts locked in but in LCM runs they appear on the images but are completely ignored in reality.
In simple terms
LCM models are optimized for low CFG (1.0 - 2.0). Mathematically, at CFG 1.0, the negative prompt is completely ignored by the sampler. To make negative prompts have any effect, you must raise the CFG above 1.0, but this can introduce burn/artifacts in LCMs. For best results, rely on a strong positive prompt.
It has to do with the formula used in diffusion (this ones are based off)
Gemini's explanation resumed to the important part: ↓
At CFG 1.0: The formula simplifies mathematically. The "Negative" parts cancel each other out completely, leaving only the Positive.
Scheduler-wise you can run various models, Karras, SGM_Uniform, ddim_uniform. You can also run Euler_a, you will get varied results. I've tested these myself, it relies on what you want basically.
I suggested Karras since it's the most common
Here's a quick explanation also from Gemini:
The "Gold Standard" (Use These)
sgm_uniform
Why: This is the exact scheduler used to train SDXL LCMs. It matches the math perfectly.
Result: Best stability, correct colors, no burn.
simple (ComfyUI specific)
Why: A basic, linear progression. It doesn't overthink the noise removal.
Result: Very reliable, slightly softer than SGM.
ddim_uniform
Why: Very similar to SGM Uniform.
Result: Sharp and predictable.
karras can be used for extra sharpness, but may require lowering CFG to avoid artifacts